betaFeedback

WorksMathematics

Joint universality of the Riemann zeta-function with arbitrary real shift parameters

Abstract

We prove joint universality of the Riemann zeta-function for shifts ζ(s+idjτ)ζ(s+id_jτ) with arbitrary nonzero real parameters djd_j whose absolute values are pairwise distinct, and show that these conditions are optimal. As a consequence, we establish generalized strong recurrence for every finite family of nonzero real parameters, meaning that the corresponding shifts of the Riemann zeta-function simultaneously approximate each other. The only cases for which generalized strong recurrence remains unproved are those involving both zero and nonzero shift parameters, for which it is equivalent to the Riemann hypothesis. Our proof combines Emsalem's generalization of the six exponentials theorem with a generalization of Pecherskii's rearrangement theorem that may be of independent interest.

Introduction

In 1981 Bagchi [1] discovered an interesting connection between the Riemann hypothesis and Voronin’s universality theorem [13] for the Riemann zeta-function ζ(s)\zeta(s). Namely, he showed that the Riemann hypothesis is true if and only if, for every compact set K⊂D:={s∈C:12<Re⁡s<1}K \subset D := \{s \in\mathbb{C} : \frac{1}{2} < \operatorname{Re} s < 1\} with connected complement and every ε>0\varepsilon> 0, we have

lim inf⁡T→∞1Tmeas⁡{τ∈[0,T]:sup⁡s∈K∣ζ(s+iτ)−ζ(s)∣<ε}>0.(1)\liminf_{T\to\infty}\frac{1}{T}\operatorname{meas}\left\{\tau\in[0,T] : \sup_{s\in K}|\zeta(s+i\tau)-\zeta(s)|<\varepsilon\right\}>0. \tag*{(1)}

In the language of topological dynamics, this is called the strong recurrence property of the Riemann zeta-function (see also [2]).

An interesting generalization of (1) was suggested by Nakamura [9], who considered the problem of finding all real numbers d1,…,dnd_1,\ldots,d_n for which

lim inf⁡T→∞1Tmeas⁡{τ∈[0,T]:max⁡1≤j<k≤nsup⁡s∈K∣ζ(s+idjτ)−ζ(s+idkτ)∣<ε}>0(2)\liminf_{T\to\infty}\frac{1}{T}\operatorname{meas}\left\{\tau\in[0,T] : \max_{1\leq j<k\leq n}\sup_{s\in K}|\zeta(s+id_j\tau)-\zeta(s+id_k\tau)|<\varepsilon\right\}>0 \tag*{(2)}

for every such compact set KK and every ε>0\varepsilon> 0. This property is usually called generalized strong recurrence or self-approximation. A natural approach to proving it is to show that the shifts ζ(s+idjτ)\zeta(s+id_j\tau) simultaneously approximate arbitrarily given non-vanishing analytic functions. Nakamura proved such a joint universality phenomenon when d1,…,dnd_1,\ldots,d_n are algebraic real numbers linearly independent over Q\mathbb{Q}.

For n=2n=2, the problem of joint universality for ζ(s+id1τ),ζ(s+id2τ)\zeta(s+id_1\tau),\zeta(s+id_2\tau) was completely solved in [10, 11]. In [10], the case of irrational d2/d1d_2/d_1 was treated using the six exponentials theorem, which shows that possible linear relations among the numbers djlog⁡pd_j\log p, j=1,2j=1,2, p∈Pp\in\mathbb{P}, involve at most finitely many exceptional primes. The remaining case of nonzero rational d2/d1≠±1d_2/d_1\ne\pm1 was established in [11] by modifying the classical denseness argument based on Pecherskii’s rearrangement theorem. Consequently, for two nonzero real numbers d1,d2d_1,d_2 with d1≠±d2d_1\ne\pm d_2, the shifts ζ(s+id1τ)\zeta(s+id_1\tau) and ζ(s+id2τ)\zeta(s+id_2\tau) are jointly universal. This implies that the only case for which (2) with n=2n=2 remains unresolved is d1=1,d2=0d_1=1,d_2=0, which is equivalent to the Riemann hypothesis (see [11 Theorem 1.4]).

The situation for n≥3n\geq3 is much more difficult. In [12], the author proved joint universality of ζ(s+id1τ),…,ζ(s+idnτ)\zeta(s+id_1\tau),\ldots,\zeta(s+id_n\tau) when KK is a singleton and d1,…,dnd_1,\ldots,d_n are distinct positive rational numbers, namely for any fixed s=σ+its=\sigma+it with 1/2<σ≤11/2<\sigma\leq1 and t>0t>0, the values log⁡ζ(s+idjτ)\log\zeta(s+id_j\tau) can simultaneously approximate arbitrary complex numbers. Very recently, Mishou [7] obtained joint universality for compact sets and positive integer parameters d1<⋯<dnd_1 < \cdots< d_n satisfying certain additional separation conditions. Moreover, at a conference in Vilnius in September 2026, he announced that the separation conditions can be removed, so that joint universality holds for arbitrary distinct positive integers d1,…,dnd_1,\ldots,d_n. Nevertheless, the problem of characterizing the set of possible d1,…,dnd_1,\ldots,d_n satisfying joint universality of ζ(s+id1τ),…,ζ(s+idnτ)\zeta(s+id_1\tau),\ldots,\zeta(s+id_n\tau), or at least (2), remained open.

The main difficulty is related to the linear dependence over Q\mathbb{Q} of the numbers djlog⁡pd_j\log p, where pp runs over the primes. In the case considered by Nakamura, these numbers are linearly independent, and the classical Kronecker–Weyl theorem can be applied directly. For general real djd_j, relations can occur both among the numbers d1,…,dnd_1,\ldots,d_n and between djlog⁡pd_j\log p corresponding to different primes. The former can be taken into account by choosing a Z\mathbb{Z}-basis β1,…,βm\beta_1,\ldots,\beta_m of the additive group generated by d1,…,dnd_1,\ldots,d_n. To deal with the latter, we use a result due to Emsalem concerning linear relations among logarithms of algebraic numbers, which generalizes the six exponentials theorem (see [14, 15]). As in the case of two shifts considered in [10], this allows us to isolate a finite set SS of exceptional primes such that the numbers βℓlog⁡p\beta_\ell\log p, 1≤ℓ≤m1\leq\ell\leq m, p∉Sp\notin S, are linearly independent over Q\mathbb{Q}.

Another crucial ingredient of our proof is a generalization of Pecherskii’s rearrangement theorem. Let HH be a complex Hilbert space. In Voronin’s original approach [13], one first fixes a sequence (vp)(v_p) in HH and requires that, for every 0≠g∈H0\ne g\in H,

∑pmax⁡(0,Re⁡⟨vp,g⟩)=+∞,∑pmin⁡(0,Re⁡⟨vp,g⟩)=−∞.\sum_p \max(0,\operatorname{Re}\langle v_p,g\rangle)=+\infty,\qquad \sum_p \min(0,\operatorname{Re}\langle v_p,g\rangle)=-\infty.

Together with ∑p∥vp∥2<∞\sum_p\lVert v_p\rVert^2<\infty, these conditions allow one to approximate any element of HH by rearranging the series ∑pvp\sum_p v_p.

On the other hand, in Bagchi’s approach (see [1] or [6 Chapter 6, §6.1, Theorem 1.16]), one considers vectors zpvpz_pv_p, where ∣zp∣=1|z_p|=1 and vp∈Hv_p\in H is fixed for each pp. The corresponding condition is

∑psup⁡z∈TRe⁡⟨zvp,g⟩=∑p∣⟨vp,g⟩∣=∞(0≠g∈H).\sum_p\sup_{z\in\mathbb{T}}\operatorname{Re}\langle zv_p,g\rangle =\sum_p|\langle v_p,g\rangle|=\infty\qquad(0\ne g\in H).

For example, in the proof of joint universality for Dirichlet LL-functions, one takes vp=(χ1(p)p−s,…,χn(p)p−s)v_p=(\chi_1(p)p^{-s},\ldots,\chi_n(p)p^{-s}), where χ1,…,χn\chi_1,\ldots,\chi_n are non-equivalent Dirichlet characters.

We generalize this criterion by replacing the sets {zvp:z∈T}\{zv_p:z\in\mathbb{T}\} with arbitrary nonempty bounded sets Ap⊂HA_p\subset H. More precisely, assuming

∑psup⁡a∈Ap∥a∥2<∞,\sum_p\sup_{a\in A_p}\lVert a\rVert^2<\infty,

we prove that the condition

∑psup⁡a∈ApRe⁡⟨a,g⟩=∞(0≠g∈H)\sum_p\sup_{a\in A_p}\operatorname{Re}\langle a,g\rangle=\infty\qquad(0\ne g\in H)

implies that the finite sums ∑p∈Map\sum_{p\in M}a_p, where ap∈Apa_p\in A_p, are dense in HH.

Our main result shows that joint universality of ζ(s+id1τ),…,ζ(s+idnτ)\zeta(s+id_1\tau),\ldots,\zeta(s+id_n\tau) holds under the optimal conditions dj≠0d_j\ne0 and ∣dj∣≠∣dk∣|d_j|\ne|d_k| for j≠kj\ne k. In particular, no linear independence assumptions on d1,…,dnd_1,\ldots,d_n are needed.

Theorem 1.1. Let n≥1n\geq1 and let d1,…,dnd_1,\ldots,d_n be nonzero real numbers satisfying

∣dj∣≠∣dk∣(1≤j<k≤n).(3)|d_j|\ne|d_k|\qquad(1\leq j<k\leq n). \tag*{(3)}

For each 1≤j≤n1\leq j\leq n, let Kj⊂DK_j\subset D be compact with connected complement, and let fjf_j be continuous and non-vanishing on KjK_j and holomorphic in its interior. Then for every ε>0\varepsilon>0,

lim inf⁡T→∞1Tmeas⁡{τ∈[T,2T]:max⁡1≤j≤nsup⁡s∈Kj∣ζ(s+idjτ)−fj(s)∣<ε}>0.(4)\liminf_{T\to\infty}\frac{1}{T}\operatorname{meas}\left\{\tau\in[T,2T]:\max_{1\leq j\leq n}\sup_{s\in K_j}|\zeta(s+id_j\tau)-f_j(s)|<\varepsilon\right\}>0. \tag*{(4)}

The conditions on d1,…,dnd_1,\ldots,d_n in Theorem 1.1 are optimal for joint universality. Indeed, if dj=0d_j=0, the corresponding function does not depend on τ\tau. If dj=dkd_j=d_k, the two corresponding shifts are identical. Finally, if dj=−dkd_j=-d_k, then for every real σ∈(1/2,1)\sigma\in(1/2,1),

ζ(σ+idjτ)=ζ(σ+idkτ)‾(τ∈R),\zeta(\sigma+id_j\tau)=\overline{\zeta(\sigma+id_k\tau)}\qquad(\tau\in\mathbb{R}),

so the values of ζ(s+idjτ)\zeta(s+id_j\tau) and ζ(s+idkτ)\zeta(s+id_k\tau) cannot approximate two independent functions.

As an immediate consequence of Theorem 1.1, we obtain generalized strong recurrence for every nonzero real parameters d1,…,dnd_1,\ldots,d_n.

Corollary 1.2. Let n≥2n\ge2 and let d1,…,dnd_1,\ldots,d_n be arbitrary nonzero real numbers. Then for every compact set K⊂DK\subset D with connected complement and every ε>0\varepsilon>0, we have

lim inf⁡T→∞1Tmeas⁡{τ∈[T,2T]:max⁡1≤j<k≤nsup⁡s∈K∣ζ(s+idjτ)−ζ(s+idkτ)∣<ε}>0.(5)\liminf_{T\to\infty}\frac{1}{T}\operatorname{meas}\left\{\tau\in[T,2T]:\max_{1\le j<k\le n}\sup_{s\in K}\left|\zeta(s+id_j\tau)-\zeta(s+id_k\tau)\right|<\varepsilon\right\}>0. \tag*{(5)}

Proof. Let a1,…,ama_1,\ldots,a_m be the distinct numbers among ∣d1∣,…,∣dn∣|d_1|,\ldots,|d_n|, and put K∗={s‾:s∈K}K^*=\{\overline{s}:s\in K\}. Choose a closed rectangle K0⊂DK_0\subset D containing K∪K∗K\cup K^*. By Theorem 1.1, applied to the positive parameters a1,…,ama_1,\ldots,a_m with all approximated functions equal to 11 on K0K_0, there is a set of τ\tau of positive lower density such that

max⁡1≤h≤msup⁡s∈K0∣ζ(s+iahτ)−1∣<ε2.\max_{1\le h\le m}\sup_{s\in K_0}\left|\zeta(s+ia_h\tau)-1\right|<\frac{\varepsilon}{2}.

Since ζ(s)‾=ζ(s‾)\overline{\zeta(s)}=\zeta(\overline{s}), the same estimate holds on KK for every shift ζ(s+idjτ)\zeta(s+id_j\tau), including those with dj<0d_j<0. This implies (5). □

Corollary 1.2 states that Nakamura’s generalized strong recurrence problem is settled for any finite family of nonzero real parameters. Let us emphasize that if at least one parameter djd_j is zero and at least one is nonzero, (2) is equivalent to the Riemann hypothesis by Bagchi’s theorem and Theorem 1.1.

Linear independence and Emsalem’s theorem

As can be seen from the proof of the main result of [9], the proof of Theorem 1.1 is quite standard if the set of numbers {djlog⁡p:j=1,2,…,n, p∈P}\{d_j\log p:j=1,2,\ldots,n,\ p\in\mathbb{P}\} is linearly independent over Q\mathbb{Q}. In such a case one can use the Kronecker-Weyl theorem to show that the quantities

djτlog⁡p2π(mod1),1≤j≤n,p∈Q,\frac{d_j\tau\log p}{2\pi}\pmod1,\qquad1\le j\le n,\quad p\in Q,

can simultaneously approximate arbitrary given real numbers (mod 1), for any finite set of primes QQ. This is a crucial step in many known proofs of universality and, simultaneously, the main obstacle in proving Theorem 1.1 for d1,…,dnd_1,\ldots,d_n when the numbers djlog⁡pd_j\log p are linearly dependent over Q\mathbb{Q}.

The case when ∑j=1ncjdj=0\sum_{j=1}^n c_jd_j=0 for some numbers cj∈Qc_j\in\mathbb{Q}, not all zero, which is equivalent to ∑j=1ncjdjlog⁡p=0\sum_{j=1}^n c_jd_j\log p=0 for every prime pp, we shall treat by choosing a Z\mathbb{Z}-basis β1,…,βm\beta_1,\ldots,\beta_m of the additive group generated by d1,…,dnd_1,\ldots,d_n, as in (6) below. Then, for each fixed prime pp, the numbers β1log⁡p,…,βmlog⁡p\beta_1\log p,\ldots,\beta_m\log p are linearly independent over Q\mathbb{Q}, and every djlog⁡pd_j\log p is a linear combination of them with integer coefficients.

However, linear relations involving different primes may still occur. For example, if d1=log⁡2d_1=\log2 and d2=log⁡3d_2=\log3, then d1log⁡3−d2log⁡2=0d_1\log3-d_2\log2=0. Such relations for n=2n=2 were investigated in [10] using the six exponentials theorem.

Our first main observation is that, by applying a generalization of the six exponentials theorem due to Emsalem (see Theorem 2.1), one can show that only finitely many primes can occur in nontrivial linear relations among the numbers

βℓlog⁡p,1≤ℓ≤m,p∈P.\beta_\ell\log p,\qquad1\le\ell\le m,\quad p\in\mathbb{P}.

More precisely, we prove that there exists a finite set of primes SS such that if

∑p∑ℓ=1mcℓ,pβℓlog⁡p=0,cℓ,p∈Q,\sum_p\sum_{\ell=1}^{m}c_{\ell,p}\beta_\ell\log p=0,\qquad c_{\ell,p}\in\mathbb{Q},

and only finitely many cℓ,pc_{\ell,p}’s are non-zero, then cℓ,p=0c_{\ell,p}=0 for p∉Sp\notin S, 1≤ℓ≤m1\le\ell\le m. Thus, the numbers βℓlog⁡p\beta_\ell\log p with p∉Sp\notin S are linearly independent over Q\mathbb{Q} and are also independent of all the numbers βℓlog⁡q\beta_\ell\log q with q∈Sq\in S.

The set {βjlog⁡q:q∈S, j=1,2,…,m}\{\beta_j\log q:q\in S,\ j=1,2,\ldots,m\} may still contain linear dependencies. We therefore choose a Z\mathbb{Z}-basis λ1,…,λr\lambda_1,\ldots,\lambda_r of the additive group it generates. Consequently, the numbers

λ1,…,λr,βℓlog⁡p(p∉S, 1≤ℓ≤m)\lambda_1,\ldots,\lambda_r,\qquad\beta_\ell\log p\quad(p\notin S,\ 1\le\ell\le m)

are linearly independent over Q\mathbb{Q}, which enables us to apply the Kronecker-Weyl theorem.

Exceptional set and Emsalem’s theorem

L0=span⁡Q{log⁡p:p prime}⊂{log⁡a:a∈Q>0}=:L.L_0=\operatorname{span}_{\mathbb{Q}}\{\log p:p\text{ prime}\}\subset\{\log a:a\in\mathbb{Q}_{>0}\}=:L.

Theorem 2.1. Let m≥1m\ge1 and let VV be a complex hyperplane in Cm\mathbb{C}^m such that V∩Qm={0}V\cap\mathbb{Q}^m=\{0\}. Then

dim⁡Q(V∩Lm)≤m(m−1).\dim_{\mathbb{Q}}(V\cap L^m)\le m(m-1).

In particular, V∩L0mV\cap L_0^m is of finite dimension over Q\mathbb{Q}.

This follows from Emsalem’s theorem in the form stated in [14 Chapter 13, p. 13-2] (see also [15 Corollary 11.10]). The cited result concerns the larger Q\mathbb{Q}-vector space {z∈C:ez∈Q‾×}\{z\in\mathbb{C}:e^z\in\overline{\mathbb{Q}}^{\times}\} of complex logarithms of nonzero algebraic numbers.

Before we apply Theorem 2.1 to our case, let us introduce some notations. Assume that d1,…,dn∈R∖{0}d_1,\ldots,d_n\in\mathbb{R}\setminus\{0\} and put

G=∑j=1nZdj⊂R.G=\sum_{j=1}^{n}\mathbb{Z}d_j\subset\mathbb{R}.

As a finitely generated torsion-free abelian group, GG has a Z\mathbb{Z}-basis β1,…,βm\beta_1,\ldots,\beta_m for some 1≤m≤n1\le m\le n. In particular, β1,…,βm\beta_1,\ldots,\beta_m are linearly independent over Q\mathbb{Q} and

dj=∑ℓ=1majℓβℓ,for suitable ajℓ∈Z.(6)d_j=\sum_{\ell=1}^{m}a_{j\ell}\beta_\ell,\qquad\text{for suitable }a_{j\ell}\in\mathbb{Z}. \tag*{(6)}

Define

Vβ={z∈Cm:∑ℓ=1mβℓzℓ=0}.V_\beta=\left\{z\in\mathbb{C}^m:\sum_{\ell=1}^{m}\beta_\ell z_\ell=0\right\}.

It is a complex hyperplane, with Vβ∩Qm={0}V_\beta\cap\mathbb{Q}^m=\{0\}.

Theorem 2.2. There exists a finite set SS of prime numbers such that for any finitely supported family bp∈Gb_p\in G,

∑pbplog⁡p=0⟹bp=0(p∉S).(7)\sum_p b_p\log p=0\quad\Longrightarrow\quad b_p=0\quad(p\notin S). \tag*{(7)}

Proof. Applying Theorem 2.1 to VβV_\beta gives that the vector space

Vβ∩L0m⊆Vβ∩LmV_\beta\cap L_0^m\subseteq V_\beta\cap L^m

has finite dimension over Q\mathbb{Q}, since Vβ∩Qm={0}V_\beta\cap\mathbb{Q}^m=\{0\} and

dim⁡Q(Vβ∩L0m)≤dim⁡Q(Vβ∩Lm)≤m(m−1).\dim_{\mathbb{Q}}(V_\beta\cap L_0^m)\le\dim_{\mathbb{Q}}(V_\beta\cap L^m)\le m(m-1).

Choose a finite basis v(1),…,v(r)v^{(1)},\ldots,v^{(r)} of Vβ∩L0mV_\beta\cap L_0^m. Each coordinate of every v(a)v^{(a)} is a finite linear combination over Q\mathbb{Q} of logarithms of primes. Since the numbers log⁡p\log p, with pp running over all primes, are linearly independent over Q\mathbb{Q}, such a representation is unique. Thus, writing

vj(a)=∑p∈Pcj,p(a)log⁡p,cj,p(a)∈Q,v_j^{(a)}=\sum_{p\in\mathbb{P}}c_{j,p}^{(a)}\log p,\qquad c_{j,p}^{(a)}\in\mathbb{Q},

where only finitely many coefficients are nonzero, we define

S=⋃a,j{p∈P:cj,p(a)≠0}.S=\bigcup_{a,j}\{p\in\mathbb{P}:c_{j,p}^{(a)}\ne0\}.

Clearly, SS is finite. Moreover, every vector of Vβ∩L0mV_{\beta}\cap L_{0}^{m} has all its coordinates in span⁡Q{log⁡q:q∈S}\operatorname{span}_{\mathbb{Q}}\{\log q:q\in S\}.

Suppose now that bp∈Gb_p\in G and ∑pbplog⁡p=0\sum_p b_p\log p=0. Write bp=∑ℓcℓ,pβℓb_p=\sum_{\ell}c_{\ell,p}\beta_{\ell} with cℓ,p∈Zc_{\ell,p}\in\mathbb{Z}. We obtain

∑ℓ=1mβℓ(∑pcℓ,plog⁡p)=0.\sum_{\ell=1}^{m}\beta_{\ell}\left(\sum_p c_{\ell,p}\log p\right)=0.

Therefore the vector (∑pc1,plog⁡p,…,∑pcm,plog⁡p)∈Vβ∩L0m\left(\sum_p c_{1,p}\log p,\ldots,\sum_p c_{m,p}\log p\right)\in V_{\beta}\cap L_{0}^{m}. By the construction of SS, every coordinate has a representation involving only log⁡q\log q with q∈Sq\in S. Since the logarithms of distinct primes are linearly independent over Q\mathbb{Q}, we have cℓ,p=0c_{\ell,p}=0 for every p∉Sp\notin S and every ℓ\ell. This proves (7). □

Parametrization of the factors p−idjτp^{-id_j\tau}

Theorem 2.2 allows us to choose real numbers that are linearly independent over Q\mathbb{Q} and hence to apply the classical Kronecker–Weyl theorem. However, the proof of Theorem 1.1 involves the factors

p−idjτ,1≤j≤n,p∈Q,p^{-id_j\tau},\qquad1\leq j\leq n,\quad p\in Q,

where QQ is a finite set of primes, and the numbers djlog⁡pd_j\log p need not be linearly independent over Q\mathbb{Q}. We therefore describe how to parametrize these factors using the independent numbers obtained from Theorem 2.2.

Recall from (6) that

G=⨁ℓ=1mZβℓ,dj=∑ℓ=1majℓβℓ,ajℓ∈Z.G=\bigoplus_{\ell=1}^{m}\mathbb{Z}\beta_{\ell},\qquad d_j=\sum_{\ell=1}^{m}a_{j\ell}\beta_{\ell},\qquad a_{j\ell}\in\mathbb{Z}.

The group ∑q∈SGlog⁡q\sum_{q\in S}G\log q is finitely generated and torsion-free, and hence admits a Z\mathbb{Z}-basis λ1,…,λr\lambda_1,\ldots,\lambda_r. We allow r=0r=0 when S=∅S=\varnothing. Although the numbers βℓlog⁡q\beta_{\ell}\log q, 1≤ℓ≤m1\leq\ell\leq m, q∈Sq\in S, generate ∑q∈SGlog⁡q\sum_{q\in S}G\log q, they may satisfy nontrivial linear relations over Q\mathbb{Q} (see Remark 2.3). For every q∈Sq\in S and 1≤j≤n1\leq j\leq n, write

djlog⁡q=∑h=1rbj,q,hλh,bj,q,h∈Z.(8)d_j\log q=\sum_{h=1}^{r}b_{j,q,h}\lambda_h,\qquad b_{j,q,h}\in\mathbb{Z}. \tag*{(8)}

Let Q⊇SQ\supseteq S be a finite set of primes. By (7), the numbers

λ1,…,λr,βℓlog⁡p(p∈Q∖S, 1≤ℓ≤m)(9)\lambda_1,\ldots,\lambda_r,\qquad\beta_{\ell}\log p\quad(p\in Q\setminus S,\ 1\leq\ell\leq m) \tag*{(9)}

are linearly independent over Q\mathbb{Q}. Indeed, consider a linear combination over Q\mathbb{Q} among the numbers in (9). Multiplying by a common denominator and expressing the terms involving λh\lambda_h as integer combinations of djlog⁡qd_j\log q with q∈Sq\in S show that, by (7), the coefficient in GG at each p∈Q∖Sp\in Q\setminus S must vanish. Since β1,…,βm\beta_1,\ldots,\beta_m are linearly independent over Q\mathbb{Q}, all the coefficients at such primes are zero. The remaining coefficients are zero as well, because λ1,…,λr\lambda_1,\ldots,\lambda_r are linearly independent over Q\mathbb{Q}.

Thus the Kronecker–Weyl theorem applies to the numbers in (9). To recover the factors p−idjτp^{-id_j\tau} from (9), we introduce the following parametrization.

Put NQ=r+m∣Q∖S∣N_Q=r+m|Q\setminus S| and identify

[0,1]NQ=[0,1]r×∏p∈Q∖S[0,1]m.[0,1]^{N_Q}=[0,1]^r\times\prod_{p\in Q\setminus S}[0,1]^m.

For u=(u1,…,ur)∈[0,1]r\boldsymbol{u}=(u_1,\ldots,u_r)\in[0,1]^r and vp=(vp,1,…,vp,m)∈[0,1]m\boldsymbol{v}_p=(v_{p,1},\ldots,v_{p,m})\in[0,1]^m, p∈Q∖Sp\in Q\setminus S, define

ΨQ:[0,1]NQ⟶(Tn)∣Q∣\Psi_Q:[0,1]^{N_Q}\longrightarrow(\mathbb{T}^{n})^{|Q|}

by

ΨQ(u,(vp)p∈Q∖S)=(ψS(u),(ψp(vp))p∈Q∖S),\Psi_Q\left(\boldsymbol{u},(\boldsymbol{v}_p)_{p\in Q\setminus S}\right) =\left(\psi_S(\boldsymbol{u}),(\psi_p(\boldsymbol{v}_p))_{p\in Q\setminus S}\right),

where

ψS(u)=((e(∑h=1rbj,q,huh))j=1n)q∈S\psi_S(\boldsymbol{u})=\left(\left(e\left(\sum_{h=1}^{r}b_{j,q,h}u_h\right)\right)_{j=1}^{n}\right)_{q\in S}

and

ψp(vp)=(e(∑ℓ=1majℓvp,ℓ))j=1n(p∈Q∖S).\psi_p(\mathbf{v}_p)=\left(e\left(\sum_{\ell=1}^{m}a_{j\ell}v_{p,\ell}\right)\right)_{j=1}^{n}\qquad(p\in Q\setminus S).

Define HS=ψS([0,1]r)H_S=\psi_S([0,1]^r) and

Hd={(e(∑ℓ=1majℓvℓ))j=1n:v∈[0,1]m}.H_d=\left\{\left(e\left(\sum_{\ell=1}^{m}a_{j\ell}v_\ell\right)\right)_{j=1}^{n}: \mathbf{v}\in[0,1]^m\right\}.

The image of ψp\psi_p is HdH_d for every p∈Q∖Sp\in Q\setminus S. Consequently,

HQ:=ΨQ([0,1]NQ)=HS×∏p∈Q∖SHd.(10)H_Q:=\Psi_Q([0,1]^{N_Q})=H_S\times\prod_{p\in Q\setminus S}H_d. \tag*{(10)}

Now let us put

αQ=−12π(λ1,…,λr,(βℓlog⁡p)p∈Q∖S, 1≤ℓ≤m),\alpha_Q=-\frac{1}{2\pi}\left(\lambda_1,\ldots,\lambda_r,(\beta_\ell\log p)_{p\in Q\setminus S,\ 1\leq\ell\leq m}\right),

and, for τ∈R\tau\in\mathbb{R}, define

γQ(τ)=((p−idjτ)j=1n)p∈Q.\gamma_Q(\tau)=\left((p^{-id_j\tau})_{j=1}^{n}\right)_{p\in Q}.

Since the coefficients ajℓa_{j\ell} and bj,q,hb_{j,q,h} are integers, (6) and (8) give

γQ(τ)=ΨQ({ταQ})∈HQ,\gamma_Q(\tau)=\Psi_Q(\{\tau\alpha_Q\})\in H_Q,

where the fractional parts are taken coordinatewise. By (9), the coordinates of αQ\alpha_Q are linearly independent over Q\mathbb{Q}. The Kronecker–Weyl theorem therefore implies that {ταQ}\{\tau\alpha_Q\} is uniformly distributed modulo 1. In particular,

HQ={γQ(τ):τ∈R}.H_Q=\{\gamma_Q(\tau):\tau\in\mathbb{R}\}.

In the proof of Theorem 1.1, we shall apply [4 Theorems A.8.1 and A.8.3] directly to {ταQ}\{\tau\alpha_Q\}.

Remark 2.3. The following example shows that HSH_S does not have to be a product over primes in SS.

Take n=m=2n=m=2 and

d1=β1=log⁡2,d2=β2=log⁡3.d_1=\beta_1=\log2,\qquad d_2=\beta_2=\log3.

Since d2log⁡2=d1log⁡3d_2\log2=d_1\log3 and d1,d2≠0d_1,d_2\ne0, we have {2,3}⊆S\{2,3\}\subseteq S. Moreover,

2−id2τ=3−id1τ(τ∈R).2^{-id_2\tau}=3^{-id_1\tau}\qquad(\tau\in\mathbb{R}).

Since

HS={γS(τ):τ∈R}‾,H_S=\overline{\{\gamma_S(\tau):\tau\in\mathbb{R}\}},

every element ((zq,1,zq,2))q∈S((z_{q,1},z_{q,2}))_{q\in S} of HSH_S satisfies z2,2=z3,1z_{2,2}=z_{3,1}.

On the other hand, since d1,d2d_1,d_2 are linearly independent over Q\mathbb{Q}, the Kronecker–Weyl theorem implies that, for every fixed q∈Sq\in S, the projection of HSH_S onto the coordinates corresponding to qq is the whole T2\mathbb{T}^2. Consequently, HSH_S cannot be the product of these projections.

Denseness lemma

For a bounded smooth Jordan domain U⊂U‾⊂DU\subset\overline{U}\subset D, let B2(U)B^2(U) denote the Bergman space of analytic functions g:U→Cg:U\to\mathbb{C} satisfying

∥g∥B2(U)2:=∬U∣g(s)∣2 dσdt<∞.\lVert g\rVert_{B^2(U)}^2:=\iint_U|g(s)|^2\,d\sigma dt<\infty.

We equip B2(U)B^2(U) with the inner product

⟨f,g⟩=∬Uf(s)g(s)‾ dσdt.\langle f,g\rangle=\iint_U f(s)\overline{g(s)}\,d\sigma dt.

It is well known (see [4 Theorem A.2.7]) that, for a compact set K⊂UK\subset U, we have

sup⁡s∈K∣g(s)∣≤CK,U∥g∥B2(U).\sup_{s\in K}|g(s)|\leq C_{K,U}\lVert g\rVert_{B^2(U)}.

We will work with the Hilbert space H=B2(U1)×⋯×B2(Un)H = B^2(U_1) \times\cdots\times B^2(U_n), where Uj⊂Uj‾⊂DU_j \subset\overline{U_j} \subset D are bounded smooth Jordan domains and

⟨u,g⟩H=∑j=1n∬Ujuj(s)gj(s)‾ dσdt.\langle u,g\rangle_H=\sum_{j=1}^{n}\iint_{U_j}u_j(s)\overline{g_j(s)}\,d\sigma dt.

Our aim is to show that, for every v=(v1,…,vn)∈Hv=(v_1,\ldots,v_n)\in H and every ε>0\varepsilon>0, there exist a finite set M⊂P∖SM\subset\mathbb{P}\setminus S and real numbers ϑp\vartheta_p, p∈Mp\in M, such that

∥v−(∑p∈M∑k=1∞e(kdjϑp)kpks)j=1n∥H<ε,(11)\left\|v-\left(\sum_{p\in M}\sum_{k=1}^{\infty}\frac{e(kd_j\vartheta_p)}{kp^{ks}}\right)_{j=1}^{n}\right\|_H<\varepsilon, \tag*{(11)}

where SS denotes a finite exceptional set from Theorem 2.2. Choose 1/2<σ1<σ2<11/2<\sigma_1<\sigma_2<1 such that σ1<Re⁡s<σ2\sigma_1<\operatorname{Re}s<\sigma_2 on all Uj‾\overline{U_j}. For p∉Sp\notin S and x∈Rx\in\mathbb{R}, define

up(x)=(−log⁡(1−e(djx)p−s))j=1n,up,j(x;s)=∑k≥1e(kdjx)kpks.(12)u_p(x)=\left(-\log(1-e(d_jx)p^{-s})\right)_{j=1}^{n},\qquad u_{p,j}(x;s)=\sum_{k\geq1}\frac{e(kd_jx)}{kp^{ks}}. \tag*{(12)}

A standard approach to proving (11) uses Pecherskii’s rearrangement theorem (see [4 Theorem A.6.1]). To apply this theorem directly, one would first need to choose a sequence of real numbers (ϑp)p∉S(\vartheta_p)_{p\notin S}, independently of g∈Hg\in H, such that the fixed sequence (up(ϑp))p∉S(u_p(\vartheta_p))_{p\notin S} satisfies the assumptions of Pecherskii’s theorem. Note that the vectors up(x)u_p(x) cannot, in general, be written as zvpzv_p, where z∈Tz\in\mathbb{T} and vp∈Hv_p\in H is independent of xx. Thus, the version of Pecherskii’s theorem with unimodular coefficients mentioned in Introduction does not apply directly here.

More precisely, since HH may be considered as a real Hilbert space with the inner product Re⁡⟨u,g⟩H\operatorname{Re}\langle u,g\rangle_H and

∑p∉Ssup⁡x∈R∥up(x)∥H2≪∑pp−2σ1<∞,(13)\sum_{p\notin S}\sup_{x\in\mathbb{R}}\|u_p(x)\|_H^2\ll\sum_p p^{-2\sigma_1}<\infty, \tag*{(13)}

in order to use Pecherskii’s theorem, it suffices to check (see [4 p. 243]) that, for every 0≠g∈H0\ne g\in H,

∑p∉Smax⁡(0,Re⁡⟨up(ϑp),g⟩H)=+∞\sum_{p\notin S}\max(0,\operatorname{Re}\langle u_p(\vartheta_p),g\rangle_H)=+\infty

and

∑p∉Smin⁡(0,Re⁡⟨up(ϑp),g⟩H)=−∞.\sum_{p\notin S}\min(0,\operatorname{Re}\langle u_p(\vartheta_p),g\rangle_H)=-\infty.

To examine this condition, define

Δj(ω)=∬Uje−sωgj(s)‾ dσdt,g=(g1,…,gn)∈H.(14)\Delta_j(\omega)=\iint_{U_j}e^{-s\omega}\overline{g_j(s)}\,d\sigma dt,\qquad g=(g_1,\ldots,g_n)\in H. \tag*{(14)}

By (12), uniformly for x∈Rx\in\mathbb{R}, we have

Re⁡⟨up(x),g⟩H=Re⁡∑j=1ne(djx)Δj(log⁡p)+Og(p−2σ1).(15)\operatorname{Re}\langle u_p(x),g\rangle_H=\operatorname{Re}\sum_{j=1}^{n}e(d_jx)\Delta_j(\log p)+O_g(p^{-2\sigma_1}). \tag*{(15)}

Thus, a direct application of Pecherskii’s theorem would require a single sequence (ϑp)(\vartheta_p) for which the real series

∑p∉SRe⁡∑j=1ne(djϑp)Δj(log⁡p)\sum_{p\notin S}\operatorname{Re}\sum_{j=1}^{n}e(d_j\vartheta_p)\Delta_j(\log p)

has both divergent positive and negative parts for every 0≠g∈H0\ne g\in H. One possible approach, related to those used in [7, 10, 11], would be to choose ϑp\vartheta_p so that, for suitable primes pp,

Re⁡(e(djϑp)Δj(log⁡p))>0for all j with Δj(log⁡p)≠0.\operatorname{Re}(e(d_j\vartheta_p)\Delta_j(\log p))>0\qquad\text{for all }j\text{ with }\Delta_j(\log p)\ne0.

Such a requirement is unnecessarily strong and may impose additional restrictions on d1,…,dnd_1,\ldots,d_n. Instead, we shall estimate the entire sum

Re⁡∑j=1ne(djϑp)Δj(log⁡p)\operatorname{Re}\sum_{j=1}^{n}e(d_j\vartheta_p)\Delta_j(\log p)

directly, without imposing separate sign conditions on its individual terms.

Since the values e(d1x),…,e(dnx)e(d_1x),\ldots,e(d_nx) need not vary independently as xx ranges over R\mathbb{R}, finding a single sequence (ϑp)(\vartheta_p), independent of gg, satisfying the required divergence conditions for every 0≠g∈H0 \ne g \in H is difficult.

We therefore use a more flexible variant of Pecherskii’s theorem (Lemma 3.2 below) with Ap={up(x):x∈R}A_p=\{u_p(x):x\in\mathbb{R}\}. In view of (13), to obtain (11) it suffices to prove that

∑p∉Ssup⁡x∈RRe⁡⟨up(x),g⟩H=∞(0≠g∈H).\sum_{p\notin S}\sup_{x\in\mathbb{R}}\operatorname{Re}\langle u_p(x),g\rangle_H=\infty\qquad(0\ne g\in H).

By (15), this follows if

∑p∉Ssup⁡x∈RRe⁡∑j=1ne(djx)Δj(log⁡p)=∞(0≠g∈H),\sum_{p\notin S}\sup_{x\in\mathbb{R}}\operatorname{Re}\sum_{j=1}^{n}e(d_jx)\Delta_j(\log p)=\infty\qquad(0\ne g\in H),

since ∑pp−2σ1<∞\sum_p p^{-2\sigma_1}<\infty. To this end, we first prove the following lemma.

Lemma 3.1. Assume that dj≠0d_j\ne0 for all jj and ∣dj∣≠∣dk∣|d_j|\ne|d_k| for 1≤j<k≤n1\le j<k\le n. There exists a constant cd>0c_d>0 such that, for every A=(A1,…,An)∈CnA=(A_1,\ldots,A_n)\in\mathbb{C}^n,

sup⁡x∈RRe⁡∑j=1nAje(djx)≥cd∥A∥2,inf⁡x∈RRe⁡∑j=1nAje(djx)≤−cd∥A∥2,(16)\sup_{x\in\mathbb{R}}\operatorname{Re}\sum_{j=1}^{n}A_je(d_jx)\ge c_d\lVert A\rVert_2,\qquad\inf_{x\in\mathbb{R}}\operatorname{Re}\sum_{j=1}^{n}A_je(d_jx)\le-c_d\lVert A\rVert_2, \tag*{(16)}

where ∥A∥2=(∑j=1n∣Aj∣2)1/2\lVert A\rVert_2=\left(\sum_{j=1}^{n}|A_j|^2\right)^{1/2}.

Proof. Let hA(x)=Re⁡∑jAje(djx)h_A(x)=\operatorname{Re}\sum_j A_je(d_jx). Since the numbers ±dj\pm d_j are pairwise distinct and nonzero, direct integration of exponential functions gives

lim⁡X→∞1X∫0XhA(x) dx=0,lim⁡X→∞1X∫0XhA(x)2 dx=12∥A∥22.(17)\lim_{X\to\infty}\frac{1}{X}\int_0^X h_A(x)\,dx=0,\qquad \lim_{X\to\infty}\frac{1}{X}\int_0^X h_A(x)^2\,dx=\frac{1}{2}\lVert A\rVert_2^2. \tag*{(17)}

Indeed, since

(Re⁡z)2=12(∣z∣2+Re⁡z2),(\operatorname{Re}z)^2=\frac{1}{2}\left(|z|^2+\operatorname{Re}z^2\right),

and since dj−dk≠0d_j-d_k\ne0 for j≠kj\ne k and dj+dk≠0d_j+d_k\ne0 for all j,kj,k, we have

lim⁡X→∞1X∫0XhA(x)2 dx=12lim⁡X→∞1X∫0X∣∑j=1nAje(djx)∣2 dx+12Re⁡lim⁡X→∞1X∫0X(∑j=1nAje(djx))2 dx=12∑j=1n∣Aj∣2.\begin{aligned} \lim_{X\to\infty}\frac{1}{X}\int_0^X h_A(x)^2\,dx &=\frac{1}{2}\lim_{X\to\infty}\frac{1}{X}\int_0^X\left|\sum_{j=1}^{n}A_je(d_jx)\right|^2\,dx\\ &\quad+\frac{1}{2}\operatorname{Re}\lim_{X\to\infty}\frac{1}{X}\int_0^X\left(\sum_{j=1}^{n}A_je(d_jx)\right)^2\,dx\\ &=\frac{1}{2}\sum_{j=1}^{n}|A_j|^2. \end{aligned}

Let MA=sup⁡x∈RhA(x)M_A=\sup_{x\in\mathbb{R}}h_A(x) and mA=−inf⁡x∈RhA(x)m_A=-\inf_{x\in\mathbb{R}}h_A(x). By (17), mA,MA>0m_A,M_A>0 if AA is nonzero. On the other hand, by the Cauchy–Schwarz inequality, we have MA,mA≤∑j∣Aj∣≤n∥A∥2M_A,m_A\le\sum_j|A_j|\le\sqrt{n}\lVert A\rVert_2. Moreover, since (hA(x)+mA)(MA−hA(x))≥0(h_A(x)+m_A)(M_A-h_A(x))\ge0 for every x∈Rx\in\mathbb{R}, we have hA(x)2≤(MA−mA)hA(x)+MAmAh_A(x)^2\le(M_A-m_A)h_A(x)+M_Am_A. Hence, using (17) gives MAmA≥∥A∥22/2M_Am_A\ge\lVert A\rVert_2^2/2 and, in consequence,

MA≥∥A∥22n,mA≥∥A∥22n.M_A\ge\frac{\lVert A\rVert_2}{2\sqrt{n}},\qquad m_A\ge\frac{\lVert A\rVert_2}{2\sqrt{n}}.

Hence, both inequalities in (16) hold with cd=1/(2n)c_d=1/(2\sqrt{n}) when A≠0A\ne0. The case A=0A=0 is trivial, and the proof is complete. □

Now, we prove the following variant of Pecherskii’s theorem.

Lemma 3.2. Let HH be a complex Hilbert space and let Ak⊂HA_k\subset H be nonempty bounded sets. Put bk=sup⁡a∈Ak∥a∥b_k=\sup_{a\in A_k}\lVert a\rVert. Assume that

∑kbk2<∞,∑ksup⁡a∈AkRe⁡⟨a,g⟩=∞(0≠g∈H).(18)\sum_k b_k^2<\infty,\qquad\sum_k\sup_{a\in A_k}\operatorname{Re}\langle a,g\rangle=\infty\qquad(0\ne g\in H). \tag*{(18)}

Then, for every N≥0N \ge0, the set

{∑k∈Mak:M⊂{k:k>N} is finite, ak∈Ak}\left\{ \sum_{k \in M} a_k : M \subset\{k : k > N\}\text{ is finite},\ a_k \in A_k \right\}

is dense in HH.

The proof uses a Hahn–Banach separation argument similar to those in [3 Theorem 4(ii)] and [6 Chapter 6, § 6.1, Theorem 1.16]. To pass from sums with terms in conv⁡(Ak)\operatorname{conv}(A_k) to sums with terms in AkA_k, we use the convergence of ∑ksup⁡a∈Ak∥a∥2\sum_k \sup_{a\in A_k}\lVert a\rVert^2. A similar argument appears in [6 Chapter 6, § 6.1, Lemma 1.15], where sums with coefficients in the unit disc are approximated by sums with coefficients on the unit circle.

Proof. By replacing AkA_k with Ak∪{0}A_k \cup\{0\}, we may assume that 0∈Ak0 \in A_k. Indeed, the first condition in (18) remains unchanged, while the second is preserved since the corresponding supremum can only increase. Moreover, the set of finite sums in the conclusion is unchanged.

For each kk, put bk=sup⁡u∈Ak∥u∥b_k = \sup_{u\in A_k}\lVert u\rVert. For integers M>LM > L, let

BL,M={∑k=L+1Mmk:mk∈conv⁡(Ak)}.B_{L,M} = \left\{ \sum_{k=L+1}^{M} m_k : m_k \in\operatorname{conv}(A_k) \right\}.

where conv⁡(Ak)\operatorname{conv}(A_k) denotes the real convex hull of AkA_k. Obviously, the sets BL,MB_{L,M} are convex. Moreover, the fact that 0∈Ak0 \in A_k implies BL,M⊆BL,M+1B_{L,M} \subseteq B_{L,M+1}. Hence

BL=⋃M>LBL,MB_L = \bigcup_{M>L} B_{L,M}

is closed and convex.

We claim that BL=HB_L = H. Otherwise, by the real Hahn–Banach separation theorem and the Riesz representation theorem, there exist g∈Hg \in H, g≠0g \ne0, and α∈R\alpha\in\mathbb{R} such that

Re⁡⟨u,g⟩≤α<Re⁡⟨u0,g⟩(u∈BL)\operatorname{Re}\langle u,g\rangle\le\alpha< \operatorname{Re}\langle u_0,g\rangle\qquad(u \in B_L)

for some fixed u0∈H∖BLu_0 \in H \setminus B_L. On the other hand,

sup⁡u∈BLRe⁡⟨u,g⟩=sup⁡M>L∑k=L+1Msup⁡a∈AkRe⁡⟨a,g⟩=∑k>Lsup⁡a∈AkRe⁡⟨a,g⟩=∞,\sup_{u\in B_L} \operatorname{Re}\langle u,g\rangle = \sup_{M>L} \sum_{k=L+1}^{M} \sup_{a\in A_k}\operatorname{Re}\langle a,g\rangle = \sum_{k>L} \sup_{a\in A_k}\operatorname{Re}\langle a,g\rangle = \infty,

which gives a contradiction. Thus BL=HB_L = H for every LL.

Now fix v∈Hv \in H, ε>0\varepsilon> 0 and N≥0N \ge0. Choose L≥NL \ge N so large that

∑k>Lbk2<ε24.\sum_{k>L} b_k^2 < \frac{\varepsilon^2}{4}.

Since BL=HB_L = H, there exist M>LM > L and mk∈conv⁡(Ak)m_k \in\operatorname{conv}(A_k), L<k≤ML < k \le M, such that

∥v−∑k=L+1Mmk∥<ε2.\left\lVert v-\sum_{k=L+1}^{M}m_k\right\rVert< \frac{\varepsilon}{2}.

By the definition of the convex hull, for each L<k≤ML < k \le M we can write

mk=∑r=0Rktk,ruk,r,uk,r∈Ak,tk,r≥0,∑r=0Rktk,r=1,m_k = \sum_{r=0}^{R_k} t_{k,r}u_{k,r}, \qquad u_{k,r} \in A_k,\quad t_{k,r} \ge0,\qquad\sum_{r=0}^{R_k} t_{k,r}=1,

where RkR_k is finite.

We now choose indices rkr_k with L<k≤ML < k \le M, successively, so that

∥∑k=L+1M(uk,rk−mk)∥2≤∑k=L+1Mbk2.\left\lVert\sum_{k=L+1}^{M}(u_{k,r_k}-m_k)\right\rVert^2 \le\sum_{k=L+1}^{M} b_k^2.

Indeed, suppose that rL+1,…,rk−1r_{L+1},\ldots,r_{k-1} have already been chosen, and put

Ek−1=∑j=L+1k−1(uj,rj−mj),EL=0.E_{k-1}=\sum_{j=L+1}^{k-1}(u_{j,r_j}-m_j),\qquad E_L=0.

Since ∑r=0Rktk,r(uk,r−mk)=0\sum_{r=0}^{R_k}t_{k,r}(u_{k,r}-m_k)=0, we have

∑r=0Rktk,r∥Ek−1+uk,r−mk∥2=∥Ek−1∥2+∑r=0Rktk,r∥uk,r∥2−∥mk∥2≤∥Ek−1∥2+bk2.\sum_{r=0}^{R_k}t_{k,r}\lVert E_{k-1}+u_{k,r}-m_k\rVert^2=\lVert E_{k-1}\rVert^2+\sum_{r=0}^{R_k}t_{k,r}\lVert u_{k,r}\rVert^2-\lVert m_k\rVert^2\leq\lVert E_{k-1}\rVert^2+b_k^2.

As the coefficients tk,rt_{k,r} are nonnegative and sum to 1, there exists an index rkr_k such that

∥Ek−1+uk,rk−mk∥2≤∥Ek−1∥2+bk2.\lVert E_{k-1}+u_{k,r_k}-m_k\rVert^2\leq\lVert E_{k-1}\rVert^2+b_k^2.

Writing Ek=Ek−1+uk,rk−mkE_k=E_{k-1}+u_{k,r_k}-m_k and iterating, we obtain

∥∑k=L+1M(uk,rk−mk)∥2=∥EM∥2≤∑k=L+1Mbk2<ε24.\left\lVert\sum_{k=L+1}^{M}(u_{k,r_k}-m_k)\right\rVert^2=\lVert E_M\rVert^2\leq\sum_{k=L+1}^{M}b_k^2<\frac{\varepsilon^2}{4}.

Consequently,

∥v−∑k=L+1Muk,rk∥≤∥v−∑k=L+1Mmk∥+∥∑k=L+1M(uk,rk−mk)∥<ε2+ε2=ε.\left\lVert v-\sum_{k=L+1}^{M}u_{k,r_k}\right\rVert\leq\left\lVert v-\sum_{k=L+1}^{M}m_k\right\rVert+\left\lVert\sum_{k=L+1}^{M}(u_{k,r_k}-m_k)\right\rVert<\frac{\varepsilon}{2}+\frac{\varepsilon}{2}=\varepsilon.

By omitting the indices for which uk,rk=0u_{k,r_k}=0, we obtain a finite sum of elements of the original sets AkA_k, with all indices greater than NN. This completes the proof. □

Thus, we are ready to formulate the following denseness lemma.

Theorem 3.3. Assume that dj≠0d_j\ne0 for all jj and ∣dj∣≠∣dk∣\lvert d_j\rvert\ne\lvert d_k\rvert whenever 1≤j<k≤n1\leq j<k\leq n. For every v∈Hv\in H, ε>0\varepsilon>0 and y>0y>0, and for any real ϑp\vartheta_p at the primes p≤yp\leq y outside SS, there exist a finite set M⊂P∖SM\subset P\setminus S containing all primes p≤yp\leq y outside SS, and real numbers ϑp\vartheta_p for all p∈Mp\in M, such that

∥∑p∈Mup(ϑp)−v∥H<ε.(19)\left\lVert\sum_{p\in M}u_p(\vartheta_p)-v\right\rVert_H<\varepsilon. \tag*{(19)}

Proof of Theorem 3.3. With σ1\sigma_1 fixed above, put

Ap={up(x):x∈R}(p∉S).A_p=\{u_p(x):x\in\mathbb{R}\}\qquad(p\notin S).

Uniformly in x∈Rx\in\mathbb{R}, (12) gives

∥up(x)∥H≪p−σ1,∥up(x)−(e(djx)p−s)j=1n∥H≪p−2σ1.\lVert u_p(x)\rVert_H\ll p^{-\sigma_1},\qquad\left\lVert u_p(x)-\left(e(d_jx)p^{-s}\right)_{j=1}^{n}\right\rVert_H\ll p^{-2\sigma_1}.

In particular,

∑p∉Ssup⁡a∈Ap∥a∥H2<∞.\sum_{p\notin S}\sup_{a\in A_p}\lVert a\rVert_H^2<\infty.

For 0≠g=(g1,…,gn)∈H0\ne g=(g_1,\ldots,g_n)\in H, by (14) and Lemma 3.1, we get

sup⁡a∈ApRe⁡⟨a,g⟩H≥sup⁡x∈RRe⁡∑j=1ne(djx)Δj(log⁡p)+Og(p−2σ1)≥cd(∑j=1n∣Δj(log⁡p)∣2)1/2+Og(p−2σ1).\begin{aligned} \sup_{a\in A_p}\operatorname{Re}\langle a,g\rangle_H\geq\sup_{x\in\mathbb{R}}\operatorname{Re}\sum_{j=1}^{n}e(d_jx)\Delta_j(\log p)+O_g(p^{-2\sigma_1}) \\ &\geq c_d\left(\sum_{j=1}^{n}\lvert\Delta_j(\log p)\rvert^2\right)^{1/2}+O_g(p^{-2\sigma_1}). \end{aligned}

Choose j0j_0 with gj0≠0g_{j_0}\ne0. We claim that ∑p∉S∣Δj0(log⁡p)∣=∞\sum_{p\notin S}\lvert\Delta_{j_0}(\log p)\rvert=\infty. By [5 Lemma 3,4, and p. 229], there are xν→∞x_\nu\to\infty, constants B,c>0B,c>0, and intervals Jν⊂[xν,xν+1]J_\nu\subset[x_\nu,x_\nu+1] such that

∣Jν∣≥Bxν−2,∣Δj0(t)∣≥ce−σ2xν(t∈Jν).\lvert J_\nu\rvert\geq Bx_\nu^{-2},\qquad\lvert\Delta_{j_0}(t)\rvert\geq ce^{-\sigma_2x_\nu}\quad(t\in J_\nu).

The prime number theorem gives ∣{p:log⁡p∈Jν}∣≫exνxν3|\{p:\log p\in J_\nu\}|\gg\frac{e^{x_\nu}}{x_\nu^3}. Hence, since σ2<1\sigma_2<1,

∑log⁡p∈Jν∣Δj0(log⁡p)∣≫e(1−σ2)xνxν3⟶∞,\sum_{\log p\in J_\nu}|\Delta_{j_0}(\log p)|\gg\frac{e^{(1-\sigma_2)x_\nu}}{x_\nu^3}\longrightarrow\infty,

which proves our claim.

Thus, we have

∑p∉Ssup⁡a∈ApRe⁡⟨a,g⟩H=∞,\sum_{p\notin S}\sup_{a\in A_p}\operatorname{Re}\langle a,g\rangle_H=\infty,

and we are ready to apply Lemma 3.2. Let us fix y>0y>0 and choose y′≥yy'\ge y such that S⊂{p∈P:p≤y′}S\subset\{p\in\mathbb{P}:p\le y'\}. Write Py′′={p∈P∖S:p≤y′}P_{y'}'=\{p\in\mathbb{P}\setminus S:p\le y'\}. For each p∈Py′′p\in P_{y'}', retain already fixed ϑp\vartheta_p for p≤yp\le y, and choose ϑp∈R\vartheta_p\in\mathbb{R} arbitrarily for y<p≤y′y<p\le y'. Put vy′′=v−∑p∈Py′′up(ϑp)v_{y'}'=v-\sum_{p\in P_{y'}'}u_p(\vartheta_p). Then, by Lemma 3.2, one can find a finite set of primes M′⊂{p∈P:p>y′}M'\subset\{p\in\mathbb{P}:p>y'\} and real numbers ϑp\vartheta_p for each p∈M′p\in M' such that

∥∑p∈M′up(ϑp)−vy′′∥H<ε.\left\|\sum_{p\in M'}u_p(\vartheta_p)-v_{y'}'\right\|_H<\varepsilon.

Therefore, taking M=Py′′∪M′M=P_{y'}'\cup M' yields (19). □

Proof of Theorem 1.1

We now complete the proof by following the standard approach originating with Voronin. The only point requiring special attention is the finite exceptional set SS.

Proof of Theorem 1.1. Fix compact sets KjK_j, functions fjf_j, and ε>0\varepsilon>0 as in the statement. By Mergelyan’s theorem, since fjf_j has no zeros on KjK_j, one can find polynomials PjP_j such that

max⁡1≤j≤nsup⁡s∈Kj∣ePj(s)−fj(s)∣<ε/3.(20)\max_{1\le j\le n}\sup_{s\in K_j}\left|e^{P_j(s)}-f_j(s)\right|<\varepsilon/3. \tag*{(20)}

Choose bounded smooth Jordan domains UjU_j with Kj⊂Uj⊂Uj‾⊂DK_j\subset U_j\subset\overline{U_j}\subset D, and fix real numbers 1/2<σ1<σ2<11/2<\sigma_1<\sigma_2<1 such that σ1<Re⁡s<σ2\sigma_1<\operatorname{Re}s<\sigma_2 on every Uj‾\overline{U_j}. Write H=B2(U1)×⋯×B2(Un)H=B^2(U_1)\times\cdots\times B^2(U_n) and choose CB≥1C_B\ge1 so that

max⁡jsup⁡s∈Kj∣gj(s)∣≤CB∥g∥H(g=(g1,…,gn)∈H).(21)\max_j\sup_{s\in K_j}|g_j(s)|\le C_B\|g\|_H\qquad(g=(g_1,\ldots,g_n)\in H). \tag*{(21)}

Put B=max⁡jsup⁡s∈KjeRe⁡Pj(s)B=\max_j\sup_{s\in K_j}e^{\operatorname{Re}P_j(s)}, and fix 0<η<1/80<\eta<1/8 sufficiently small that

B((1+η)e5η/4−1)<ε/3.B\bigl((1+\eta)e^{5\eta/4}-1\bigr)<\varepsilon/3.

Moreover, there is a constant C0C_0 depending only on UjU_j and σ1\sigma_1 such that

∑j=1n∬Uj∑p>y∑k≥11k2p2kσ dσ dt≤C0y1−2σ1<η264CB2\sum_{j=1}^{n}\iint_{U_j}\sum_{p>y}\sum_{k\ge1}\frac{1}{k^2p^{2k\sigma}}\,d\sigma\,dt\le C_0y^{1-2\sigma_1}<\frac{\eta^2}{64C_B^2}

for sufficiently large y≥2y\ge2.

Let SS be the finite exceptional set from Theorem 2.2, and define

LS(s)=∑q∈Slog⁡(1−q−s),σ>0.L_S(s)=\sum_{q\in S}\log(1-q^{-s}),\qquad\sigma>0.

By Theorem 3.3 and (21), we obtain a finite set M⊂P∖SM\subset\mathbb{P}\setminus S containing every prime p≤yp\le y outside SS, and real numbers ϑp\vartheta_p, p∈Mp\in M, such that

max⁡jsup⁡s∈Kj∣∑p∈M∑k≥1e(kdjϑp)kpks−Pj(s)−LS(s)∣<η8.(22)\max_j\sup_{s\in K_j}\left|\sum_{p\in M}\sum_{k\ge1}\frac{e(kd_j\vartheta_p)}{kp^{ks}}-P_j(s)-L_S(s)\right|<\frac{\eta}{8}. \tag*{(22)}

Set ϑp=0\vartheta_p=0 for p∈Sp\in S and C=S∪MC=S\cup M. In particular, CC contains all primes p≤yp\le y. For each p∈Mp\in M, put

vp,ℓ≡βℓϑp(mod1),1≤ℓ≤m.v_{p,\ell}\equiv\beta_\ell\vartheta_p\pmod1,\qquad1\le\ell\le m.

Since ajℓ∈Za_{j\ell} \in\mathbb{Z} and dj=∑ℓ=1majℓβℓd_j = \sum_{\ell=1}^{m} a_{j\ell}\beta_\ell, we have

e(djϑp)=e(ϑp∑ℓ=1majℓβℓ)=e(∑ℓ=1majℓvp,ℓ).e(d_j\vartheta_p) = e\left(\vartheta_p\sum_{\ell=1}^{m}a_{j\ell}\beta_\ell\right) = e\left(\sum_{\ell=1}^{m}a_{j\ell}v_{p,\ell}\right).

for every j=1,…,nj=1,\ldots,n. Hence the vector

ωp∗=(e(d1ϑp),…,e(dnϑp))\omega_p^*=\left(e(d_1\vartheta_p),\ldots,e(d_n\vartheta_p)\right)

belongs to HdH_d. For the exceptional block choose all its basis parameters to be zero. Hence, by (10), we have

ω∗=(((1,…,1))q∈S′,(e(d1ϑp),…,e(dnϑp))p∈M)∈HC,\omega^*=\left(((1,\ldots,1))_{q\in S'},(e(d_1\vartheta_p),\ldots,e(d_n\vartheta_p))_{p\in M}\right)\in H_C,

where the first vector has nn times the entry 11.

For ω=(ωp,j)p∈C, 1≤j≤n∈HC\omega=(\omega_{p,j})_{p\in C,\,1\leq j\leq n}\in H_C, define

Fj(s,ω)=∑p∈C∑k=1∞ωp,jkkpks,1≤j≤n.F_j(s,\omega)=\sum_{p\in C}\sum_{k=1}^{\infty}\frac{\omega_{p,j}^{k}}{kp^{ks}},\qquad1\leq j\leq n.

Since CC is finite and ∣ωp,j∣=1|\omega_{p,j}|=1, these series converge absolutely and uniformly for s∈Uj‾s\in\overline{U_j}. Moreover, let us recall (see Section 2.2) that for the map γC(τ)=((e(−djτlog⁡p2π))j=1n)p∈C\gamma_C(\tau)=\left(\left(e\left(-\frac{d_j\tau\log p}{2\pi}\right)\right)_{j=1}^{n}\right)_{p\in C}, we have γC(τ)=ΨC({ταC})∈HC\gamma_C(\tau)=\Psi_C(\{\tau\alpha_C\})\in H_C, where

αC=−12π(λ1,…,λr,(βℓlog⁡p)p∈M, 1≤ℓ≤m).\alpha_C=-\frac{1}{2\pi}\left(\lambda_1,\ldots,\lambda_r,(\beta_\ell\log p)_{p\in M,\,1\leq\ell\leq m}\right).

Hence,

Fj(s,γC(τ))=∑p∈C∑k≥1e(−kdjτlog⁡p2π)kpks.(23)F_j(s,\gamma_C(\tau))=\sum_{p\in C}\sum_{k\geq1}\frac{e\left(-\frac{kd_j\tau\log p}{2\pi}\right)}{kp^{ks}}. \tag*{(23)}

On the other hand, by the definition of ω∗∈HC\omega^*\in H_C, we have

Fj(s,ω∗)=∑q∈S∑k=1∞1kqks+∑p∈M∑k=1∞e(kdjϑp)kpks=−LS(s)+∑p∈M∑k=1∞e(kdjϑp)kpks.F_j(s,\omega^*)=\sum_{q\in S}\sum_{k=1}^{\infty}\frac{1}{kq^{ks}}+\sum_{p\in M}\sum_{k=1}^{\infty}\frac{e(kd_j\vartheta_p)}{kp^{ks}}=-L_S(s)+\sum_{p\in M}\sum_{k=1}^{\infty}\frac{e(kd_j\vartheta_p)}{kp^{ks}}.

Therefore, by (22),

max⁡1≤j≤nsup⁡s∈Kj∣Fj(s,ω∗)−Pj(s)∣<η8.\max_{1\leq j\leq n}\sup_{s\in K_j}|F_j(s,\omega^*)-P_j(s)|<\frac{\eta}{8}.

We next verify that this estimate remains valid for vectors sufficiently close to ω∗\omega^*. Recall that Re⁡s≥σ1>1/2\operatorname{Re}s\geq\sigma_1>1/2 on each KjK_j. For any z,w∈Tz,w\in\mathbb{T} and k≥1k\geq1, the factorization

zk−wk=(z−w)∑ℓ=0k−1zk−1−ℓwℓz^k-w^k=(z-w)\sum_{\ell=0}^{k-1}z^{k-1-\ell}w^\ell

implies

∣zk−wk∣≤k∣z−w∣.|z^k-w^k|\leq k|z-w|.

Therefore, one can easily notice that for sufficiently small ρ>0\rho>0 and for every

ω∈V:={ω∈HC:max⁡p∈C1≤j≤n∣ωp,j−ωp,j∗∣<ρ},\omega\in\mathcal{V}:=\left\{\omega\in H_C:\max_{\substack{p\in C\\1\leq j\leq n}}|\omega_{p,j}-\omega_{p,j}^*|<\rho\right\},

we have

max⁡1≤j≤nsup⁡s∈Kj∣Fj(s,ω)−Pj(s)∣<η4.\max_{1\leq j\leq n}\sup_{s\in K_j}|F_j(s,\omega)-P_j(s)|<\frac{\eta}{4}.

Finally, using (23), we conclude that

max⁡1≤j≤nsup⁡s∈Kj∣Fj(s,γC(τ))−Pj(s)∣<η4.(24)\max_{1\leq j\leq n}\sup_{s\in K_j}|F_j(s,\gamma_C(\tau))-P_j(s)|<\frac{\eta}{4}. \tag*{(24)}

provided γC(τ)∈V\gamma_C(\tau)\in\mathcal{V}.

Recall that ω∗∈HC\omega^* \in H_C, so there is t∗∈[0,1)NCt^* \in[0,1)^{N_C} such that ΨC(t∗)=ω∗\Psi_C(t^*)=\omega^*. More precisely, we may take

t∗=(0,…,0⏟r,({βℓϑp})p∈M, 1≤ℓ≤m).t^*=\left(\underbrace{0,\ldots,0}_{r},\left(\{\beta_\ell\vartheta_p\}\right)_{p\in M,\ 1\leq\ell\leq m}\right).

Since ΨC\Psi_C is continuous and V\mathcal{V} is an open neighborhood of ω∗\omega^* in HCH_C, there exists 0<δ0<1/20<\delta_0<1/2 such that

ΨC(Bδ0)⊂V,\Psi_C(B_{\delta_0})\subset\mathcal{V},

where

Bδ0={t∈[0,1]NC:max⁡1≤ν≤NC∥tν−tν∗∥≤δ0}.B_{\delta_0}=\left\{t\in[0,1]^{N_C}:\max_{1\leq\nu\leq N_C}\|t_\nu-t_\nu^*\|\leq\delta_0\right\}.

Here ∥⋅∥\|\cdot\| denotes the distance to the nearest integer. In particular, the Lebesgue measure of Bδ0B_{\delta_0} is

meas⁡(Bδ0)=(2δ0)NC>0.\operatorname{meas}(B_{\delta_0})=(2\delta_0)^{N_C}>0.

For T>0T>0, define

AT={τ∈[T,2T]:{ταC}∈Bδ0}.A_T=\{\tau\in[T,2T]:\{\tau\alpha_C\}\in B_{\delta_0}\}.

Then, by (9) and the Kronecker–Weyl theorem (see [4 Theorem A.8.1]), we have

lim⁡T→∞meas⁡(AT)T=(2δ0)NC>0.(25)\lim_{T\to\infty}\frac{\operatorname{meas}(A_T)}{T}=(2\delta_0)^{N_C}>0. \tag*{(25)}

Moreover, for every τ∈AT\tau\in A_T, we have γC(τ)=ΨC({ταC})∈V\gamma_C(\tau)=\Psi_C(\{\tau\alpha_C\})\in\mathcal{V}, so (24) holds for every τ∈AT\tau\in A_T.

Let z≥2z\geq2 be fixed, with C⊆Qz:={p:p≤z}C\subseteq Q_z:=\{p:p\leq z\}. Recalling (10) gives

t=(u,(vp)p∈Qz∖S)∈[0,1]NQz,t=(u,(v_p)_{p\in Q_z\setminus S})\in[0,1]^{N_{Q_z}},

where u=(u1,…,ur)∈[0,1]ru=(u_1,\ldots,u_r)\in[0,1]^r parametrizes the exceptional block HSH_S, while vp=(vp,1,…,vp,m)∈[0,1]mv_p=(v_{p,1},\ldots,v_{p,m})\in[0,1]^m parametrizes HdH_d with p∉Sp\notin S.

Let us put

t=(tC,(vp)p∈Qz∖C)withtC=(u,(vp)p∈M)∈[0,1]NC,t=(t_C,(v_p)_{p\in Q_z\setminus C})\quad\text{with}\quad t_C=(u,(v_p)_{p\in M})\in[0,1]^{N_C},

and

Rj,z(s;t)=∑p∈Qz∖C∑k≥1e(k∑ℓ=1majℓvp,ℓ)kpks,R_{j,z}(s;t)=\sum_{p\in Q_z\setminus C}\sum_{k\geq1}\frac{e\left(k\sum_{\ell=1}^{m}a_{j\ell}v_{p,\ell}\right)}{k p^{ks}},

where (aj1,…,ajm)∈Zm∖{0}(a_{j1},\ldots,a_{jm})\in\mathbb{Z}^m\setminus\{0\} were defined in Section 2.2.

By the definition of αQz\alpha_{Q_z},

Rj,z(s;{ταQz})=∑p∈Qz∖C∑k≥11kpk(s+idjτ).(26)R_{j,z}(s;\{\tau\alpha_{Q_z}\})=\sum_{p\in Q_z\setminus C}\sum_{k\geq1}\frac{1}{k p^{k(s+id_j\tau)}}. \tag*{(26)}

Put Dz=Bδ0×∏p∈Qz∖C[0,1]mD_z=B_{\delta_0}\times\prod_{p\in Q_z\setminus C}[0,1]^m and

Fz(t)=∑j=1n∥Rj,z(⋅;t)∥B2(Uj)2=∑j=1n∬Uj∣Rj,z(s;t)∣2 dσ dt.F_z(t)=\sum_{j=1}^{n}\|R_{j,z}(\cdot;t)\|_{B^2(U_j)}^2=\sum_{j=1}^{n}\iint_{U_j}|R_{j,z}(s;t)|^2\,d\sigma\,dt.

Since the function FzF_z is continuous on [0,1]NQz[0,1]^{N_{Q_z}}, we get by [4 Theorem A.8.3] that

lim⁡T→∞1T∫AT∑j=1n∥Rj,z(s;{ταQz})∥B2(Uj)2 dτ=∫DzFz(t) dt=(2δ0)NC∑j=1n∬Uj∑p∈Qz∖Ck≥1∑p′∈Qz∖Ck′≥11k2p2kRe⁡s dσ dt.\begin{aligned} \lim_{T\to\infty}\frac{1}{T}\int_{A_T}\sum_{j=1}^{n}\|R_{j,z}(s;\{\tau\alpha_{Q_z}\})\|_{B^2(U_j)}^2\,d\tau =\int_{D_z}F_z(t)\,dt \\ &=(2\delta_0)^{N_C}\sum_{j=1}^{n}\iint_{U_j}\sum_{\substack{p\in Q_z\setminus C\\ k\geq1}}\sum_{\substack{p'\in Q_z\setminus C\\ k'\geq1}}\frac{1}{k^2p^{2k\operatorname{Re}s}}\,d\sigma\,dt. \end{aligned}

The last equality follows from meas⁡(Bδ0)=(2δ0)NC\operatorname{meas}(B_{\delta_0})=(2\delta_0)^{N_C} and the orthogonality relation

∫∏v∈Qz∖C[0,1]me(k1∑ℓ=1majℓvp1,ℓ−k2∑ℓ=1majℓvp2,ℓ)∏v∈Qz∖Cdvv={1,p1=p2, k1=k2,0,otherwise,\int_{\prod_{v\in Q_z\setminus C}[0,1]^m}e\left(k_1\sum_{\ell=1}^{m}a_{j\ell}v_{p_1,\ell}-k_2\sum_{\ell=1}^{m}a_{j\ell}v_{p_2,\ell}\right)\prod_{v\in Q_z\setminus C}dv_v = \begin{cases} 1, & p_1=p_2,\ k_1=k_2,\\ 0, & \text{otherwise}, \end{cases}

which holds since aj≠0a_j\ne0.

Since CC contains every prime p≤yp \le y, we have

∑j=1n∬Uj∑p≤zk≥1p∉C∑k≥11k2p2kRe⁡s dσdt≪∑p>yp−2σ1≪y1−2σ1.\sum_{j=1}^{n}\iint_{U_j}\sum_{\substack{p\le z\\ k\ge1\\ p\notin C}}\sum_{k\ge1}\frac{1}{k^2p^{2k\operatorname{Re}s}}\,d\sigma dt\ll\sum_{p>y}p^{-2\sigma_1}\ll y^{1-2\sigma_1}.

Thus, with a constant C0C_0 independent of MM, zz, and δ0\delta_0, we obtain

lim sup⁡T→∞1T∫AT∑j=1n∥Rj,z(s;{ταQz})∥B22(Uj)2 dτ≤(2δ0)NCC0y1−2σ1<(2δ0)NCη264CB2,\limsup_{T\to\infty}\frac{1}{T}\int_{A_T}\sum_{j=1}^{n}\left\|R_{j,z}(s;\{\tau\alpha_{Q_z}\})\right\|_{B_2^2(U_j)}^2\,d\tau\le(2\delta_0)^{N_C}C_0y^{1-2\sigma_1}<\frac{(2\delta_0)^{N_C}\eta^2}{64C_B^2},

where the last inequality follows from our earlier choice of yy.

By (21), if

max⁡jsup⁡s∈Kj∣Rj,z(s;{ταQz})∣>η,\max_j\sup_{s\in K_j}\left|R_{j,z}(s;\{\tau\alpha_{Q_z}\})\right|>\eta,

then

∑j=1n∥Rj,z(s;{ταQz})∥B22(Uj)2>η2CB2.\sum_{j=1}^{n}\left\|R_{j,z}(s;\{\tau\alpha_{Q_z}\})\right\|_{B_2^2(U_j)}^2>\frac{\eta^2}{C_B^2}.

Consequently, Markov’s inequality gives

lim sup⁡T→∞1Tmeas⁡{τ∈AT:max⁡jsup⁡s∈Kj∣Rj,z(s;{ταQz})∣>η}≤(2δ0)NC64.\limsup_{T\to\infty}\frac{1}{T}\operatorname{meas}\left\{\tau\in A_T:\max_j\sup_{s\in K_j}\left|R_{j,z}(s;\{\tau\alpha_{Q_z}\})\right|>\eta\right\}\le\frac{(2\delta_0)^{N_C}}{64}.

Thus, it follows from (24), (26) and (27) that

max⁡jsup⁡s∈Kj∣∑p≤z∑k≥11kpk(s+idjτ)−Pj(s)∣<5η4\max_j\sup_{s\in K_j}\left|\sum_{p\le z}\sum_{k\ge1}\frac{1}{kp^{k(s+id_j\tau)}}-P_j(s)\right|<\frac{5\eta}{4}

for all τ∈AT\tau\in A_T outside an exceptional set of upper density at most (2δ0)NC/64(2\delta_0)^{N_C}/64.

lim inf⁡T→∞1Tmeas⁡{τ∈[T,2T]:max⁡jsup⁡s∈Kj∣∏p≤z(1−p−s−idjτ)−1−ePj(s)∣<ε3}≥6364(2δ0)NC>0.(27)\liminf_{T\to\infty}\frac{1}{T}\operatorname{meas}\left\{\tau\in[T,2T]:\max_j\sup_{s\in K_j}\left|\prod_{p\le z}(1-p^{-s-id_j\tau})^{-1}-e^{P_j(s)}\right|<\frac{\varepsilon}{3}\right\}\ge\frac{63}{64}(2\delta_0)^{N_C}>0. \tag*{(27)}

so it remains to show that the Riemann zeta-function can be approximated by these finite Euler products.

We apply Carlson’s theorem [4 Theorem A.2.10] to the entire function

Gz(s)=(1−21−s)(ζ(s)∏p≤z(1−p−s)−1).G_z(s)=(1-2^{1-s})\left(\zeta(s)\prod_{p\le z}(1-p^{-s})-1\right).

For Re⁡s>1\operatorname{Re}s>1, it has the Dirichlet series representation Gz(s)=∑m>zbz(m)msG_z(s)=\sum_{m>z}\frac{b_z(m)}{m^s} with ∣bz(m)∣≤2\left|b_z(m)\right|\le2. Hence Carlson’s theorem gives

lim⁡T→∞1T∫0T∣Gz(σ+it)∣2 dt=∑m>z∣bz(m)∣2m2σ≪∑m>zm−2σ≪z1−2σ1,\lim_{T\to\infty}\frac{1}{T}\int_0^T\left|G_z(\sigma+it)\right|^2\,dt=\sum_{m>z}\frac{|b_z(m)|^2}{m^{2\sigma}}\ll\sum_{m>z}m^{-2\sigma}\ll z^{1-2\sigma_1},

uniformly for σ1≤σ≤σ2\sigma_1\le\sigma\le\sigma_2.

For every dj∈R∖{0}d_j\in\mathbb{R}\setminus\{0\} and s=σ+iu∈Ujs=\sigma+iu\in U_j, the substitution t=u+djτt=u+d_j\tau gives

lim⁡T→∞1T∫T2T∣Gz(s+idjτ)∣2 dτ=lim⁡T→∞1∣dj∣T∫min⁡(u+djT,u+2djT)max⁡(u+djT,u+2djT)∣Gz(σ+it)∣2 dt≪z1−2σ1.\lim_{T\to\infty}\frac{1}{T}\int_T^{2T}|G_z(s+id_j\tau)|^2\,d\tau=\lim_{T\to\infty}\frac{1}{|d_j|T}\int_{\min(u+d_jT,u+2d_jT)}^{\max(u+d_jT,u+2d_jT)}|G_z(\sigma+it)|^2\,dt\ll z^{1-2\sigma_1}.

For dj<0d_j<0, we also use ∣Gz(σ−it)∣=∣Gz(σ+it)∣|G_z(\sigma-it)|=|G_z(\sigma+it)|, since the Dirichlet coefficients bz(m)b_z(m) are real. Thus,

lim sup⁡T→∞1T∫T2T∑j=1n∬Uj∣ζ(s+idjτ)∏p≤z(1−p−s−idjτ)−1∣2 dσdt dτ≤C1z1−2σ1.(28)\limsup_{T\to\infty}\frac{1}{T}\int_T^{2T}\sum_{j=1}^{n}\iint_{U_j}\left|\zeta(s+id_j\tau)\prod_{p\le z}(1-p^{-s-id_j\tau})-1\right|^2\,d\sigma dt\,d\tau\le C_1z^{1-2\sigma_1}. \tag*{(28)}

where C1C_1 is independent of zz. Hence, we may choose zz sufficiently large such that C⊆QzC \subseteq Q_z and

C1z1−2σ1<(2δ0)NCη264CB2.C_1 z^{1-2\sigma_1}<\frac{(2\delta_0)N_C\eta^2}{64C_B^2}.

By (21), (28), and Markov’s inequality,

lim sup⁡T→∞1Tmeas⁡{τ∈AT:max⁡jsup⁡s∈Kj∣ζ(s+idjτ)∏p≤z(1−p−s−idjτ)−1∣>η}≤(2δ0)NC64.(29)\limsup_{T\to\infty}\frac{1}{T}\operatorname{meas}\left\{\tau\in A_T:\max_j\sup_{s\in K_j}\left|\zeta(s+id_j\tau)\prod_{p\leq z}(1-p^{-s-id_j\tau})-1\right|>\eta\right\}\leq\frac{(2\delta_0)^{N_C}}{64}. \tag*{(29)}

Let GTG_T be the set of τ∈AT\tau\in A_T for which

max⁡jsup⁡s∈Kj∣∑p∈Qz∖C∑k≥11kpk(s+idjτ)∣≤η,max⁡jsup⁡s∈Kj∣ζ(s+idjτ)∏p≤z(1−p−s−idjτ)−1∣≤η.\max_j\sup_{s\in K_j}\left|\sum_{p\in Q_z\setminus C}\sum_{k\geq1}\frac{1}{kp^{k(s+id_j\tau)}}\right|\leq\eta,\qquad\max_j\sup_{s\in K_j}\left|\zeta(s+id_j\tau)\prod_{p\leq z}(1-p^{-s-id_j\tau})-1\right|\leq\eta.

By (25), (27) and (29), we have

lim inf⁡T→∞meas⁡(GT)T≥3132(2δ0)NC>0.\liminf_{T\to\infty}\frac{\operatorname{meas}(G_T)}{T}\geq\frac{31}{32}(2\delta_0)^{N_C}>0.

For τ∈GT\tau\in G_T, by (4), we have

max⁡jsup⁡s∈Kj∣ζ(s+idjτ)−ePj(s)∣≤max⁡jsup⁡s∈KjeRe⁡Pj(s)((1+η)e5η/4−1)<ε3,\max_j\sup_{s\in K_j}\left|\zeta(s+id_j\tau)-e^{P_j(s)}\right|\leq\max_j\sup_{s\in K_j}e^{\operatorname{Re}P_j(s)}\left((1+\eta)e^{5\eta/4}-1\right)<\frac{\varepsilon}{3},

by our choice of η\eta. Together with (20), this gives

max⁡jsup⁡s∈Kj∣ζ(s+idjτ)−fj(s)∣<ε(τ∈GT).\max_j\sup_{s\in K_j}\left|\zeta(s+id_j\tau)-f_j(s)\right|<\varepsilon\qquad(\tau\in G_T).

This completes the proof. □\square

References

Faculty of Mathematics and Computer Science, Adam Mickiewicz University, Uniwersytetu Poznańskiego 4, 61-614 Poznań, Poland.

Email address: [email protected]

References

  1. [1]B. Bagchi, The statistical behavior and universality properties of the Riemann zeta-function and other allied Dirichlet series, Ph.D. thesis, Indian Statistical Institute, Calcutta, 1981.
  2. [2]B. Bagchi, Recurrence in topological dynamics and the Riemann hypothesis, Acta Math. Hungar. 50 (1987), 227–240.DOI
  3. [3]A. Good, On the distribution of the values of Riemann’s zeta function, Acta Arith. 38 (1980/81), no. 4, 347–388.DOI
  4. [4]A. A. Karatsuba and S. M. Voronin, The Riemann Zeta-Function, Walter de Gruyter, Berlin–New York, 1992.
  5. [5]J. Kaczorowski and M. Kulas, On the non-trivial zeros off the critical line for L-functions from the extended Selberg class, Monatsh. Math. 150 (2007), 217–232.DOI
  6. [6]A. Laurinčikas, Limit Theorems for the Riemann Zeta-Function, Mathematics and Its Applications, vol. 352, Kluwer Academic Publishers, Dordrecht, 1996.DOI
  7. [7]H. Mishou, Joint universality for shifts of the Riemann zeta function with integer multiplicities, preprint, 2026.
  8. [9]T. Nakamura, The joint universality and the generalized strong recurrence for Dirichlet L-functions, Acta Arith. 138 (2009), 357–362.DOI
  9. [10]Ł. Pańkowski, Some remarks on the generalized strong recurrence for L-functions, in: New Directions in Value Distribution Theory of Zeta and L-Functions, Ber. Math., Shaker Verlag, Aachen, 2009, pp. 305–315.
  10. [11]Ł. Pańkowski, Joint universality and generalized strong recurrence with rational parameter, J. Number Theory 163 (2016), 61–74.
  11. [12]Ł. Pańkowski, Joint value-distribution of shifts of the Riemann zeta-function, Results Math. 77 (2022), article 76.arxiv.org/abs/2010.08332
  12. [13]S. M. Voronin, Theorem on the universality of the Riemann zeta-function, Math. USSR-Izv. 9 (1975), 443–453.DOI
  13. [14]M. Waldschmidt, Linear Independence of Logarithms of Algebraic Numbers, The Institute of Mathematical Sciences, Madras, IMSc Report 116 (1992), Chapter 13, p. 13-2, available at https://webusers.imj-prg.fr/~michel.waldschmidt/articles/pdf/LIL.pdf.
  14. [15]M. Waldschmidt, Diophantine Approximation on Linear Algebraic Groups: Transcendence Properties of the Exponential Function in Several Variables, Grundlehren der mathematischen Wissenschaften, vol. 326, Springer, Berlin, 2000.

Paper details

vlnc:W_9qve7m5h
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents