betaFeedback

WorksMathematics

Prime supports and vanishing in restricted coprime representations

Abstract

We study the prime supports that eliminate every representation of an even integer NN by two positive integers coprime to 6P6P. Let J(k)J(k) be the largest vanishing instance obtainable with at most kk factors of PP not dividing N/2N/2. We prove k(log⁡k)2≪J(k)≪εk4+εk(\log k)^2\ll J(k)\ll_\varepsilon k^{4+\varepsilon} for every ε>0\varepsilon>0. The upper bound follows by comparing a classical family of shifted-prime representations with a two-dimensional upper sieve for the loss at each additional factor. Retaining the same local density factor in both estimates yields a weighted survival criterion and positivity whenever the number of noncolliding factors above NθN^θ is sufficiently small compared with NθN^θ, for fixed 0<θ<1/40<θ<1/4. An exact reduction separates a minimum prime cover into forced prime-pair contributions and a residual problem, and gives the lower bound for JJ along powers of two. Residue occupancies characterize the extremal concentration at a first zero. Explicit discrepancy estimates show that supports with at most eleven factors cannot eliminate every representation when N>2780460N>2780460, nor supports with at most seven noncolliding factors when N≥2266292N\ge2266292. Positivity for the complete support of primes up to N\sqrt N remains open.

Introduction

Let N=2n≥26N = 2n \ge26 be even, and let PP be a product of distinct primes at least 5. We consider

gP(N)=#{1≤h≤n:gcd⁡(h(N−h),6P)=1}.(1)g_P(N) = \#\{1 \le h \le n : \gcd(h(N-h), 6P) = 1\}. \tag*{(1)}

The endpoint 1+(N−1)1 + (N - 1) is included when admissible. The question is how many prime factors, and what arrangement of their excluded classes, are needed to make this count vanish.

The one-prime case was studied in [8]. Exact formulas for squarefree supports, together with the parametrization h=6z+βh = 6z + \beta, discrepancy estimates, and positivity thresholds, were developed in [9]. Here the primes of an instance remain fixed, and the excluded residues are prescribed by that same NN. Removing a factor reveals the preceding survivor set. Imposing the factors in increasing order identifies a first vanishing step.

Write δq(x)\delta_q(x) for the least nonnegative remainder of the integer xx upon division by qq. For an odd prime q∣Pq \mid P, the forbidden residues are 00 and δq(N)\delta_q(N). They coincide precisely when q∣Nq \mid N, or equivalently q∣nq \mid n. We call these factors colliding and write

k(N,P)=ω(P/gcd⁡(P,n)).k(N,P)=\omega(P/\gcd(P,n)).

Here ω\omega counts distinct prime factors. For each integer k≥0k \ge0, our principal size parameter is

J(k)=sup⁡({0}∪{N≥26:N even, gP(N)=0 for some P with k(N,P)≤k}).(2)J(k)=\sup\left(\{0\}\cup\{N\ge26:N\ \text{even},\ g_P(N)=0\ \text{for some }P\text{ with }k(N,P)\le k\}\right). \tag*{(2)}

It measures the largest vanishing instance allowed by at most kk noncolliding factors, with no bound on their size or on the number of colliding factors. The supremum compares different fixed instances (N,P)(N,P).

Theorem 1.1 (Size of a vanishing instance). For every ε>0\varepsilon>0,

k(log⁡k)2≪J(k)≪εk4+ε(k⟶∞).(3)k(\log k)^2\ll J(k)\ll_{\varepsilon} k^{4+\varepsilon}\qquad(k\longrightarrow\infty). \tag*{(3)}

In particular, J(k)J(k) is finite for every fixed kk.

The upper bound follows from Corollary 5.5. The lower bound is proved in Proposition 6.4 by constructing prime covers along the sequence N=2jN=2^j. This choice avoids a logarithmic loss from the local density factor.

The estimates arise from two complementary descriptions of the same covering problem. First, for any fixed 0<θ<1/40<\theta<1/4, the classical linear sieve supplies

Tθ(N)={2≤h<N/2:N−h prime, P−(h)>Nθ},\mathcal{T}_{\theta}(N)=\{2\le h<N/2:N-h\ \text{prime},\ P^{-}(h)>N^{\theta}\},

with cardinality ≫θS(N)N/log⁡2N\gg_{\theta}\mathfrak{S}(N)N/\log^2N, where S(N)=∏ℓ∣N, ℓ>2(ℓ−1)/(ℓ−2)\mathfrak{S}(N)=\prod_{\ell\mid N,\ \ell>2}(\ell-1)/(\ell-2). An upper sieve applied to m(N−qm)m(N-qm) bounds the loss caused by a factor Nθ<q≤NN^{\theta}<q\le\sqrt{N} by ≪θS(N)N/(qlog⁡2N)\ll_{\theta}\mathfrak{S}(N)N/(q\log^2N). The common local factor cancels. For sufficiently large even NN, Theorem 5.4 gives a weighted positivity criterion and, in particular,

#{q∣P:Nθ<q<N, q∤N}≤ηθNθ⟹gP(N)≫θS(N)Nlog⁡2N.(4)\#\{q\mid P:N^{\theta}<q<N,\ q\nmid N\}\le\eta_{\theta}N^{\theta} \quad\Longrightarrow\quad g_P(N)\gg_{\theta}\mathfrak{S}(N)\frac{N}{\log^2N}. \tag*{(4)}

The shifted-prime input is classical [2, 6]. The result here compares its supply with the losses of a fixed support.

Second, let τ(N)\tau(N) be the minimum total number of prime factors in a support with gP(N)=0g_P(N)=0. Prime pairs force disjoint covering requirements, each of which can be met by choosing its smaller summand. Theorem 6.1 separates them exactly:

τ(N)=H(N)+τ0(N),(5)\tau(N)=H(N)+\tau_0(N), \tag*{(5)}

where H(N)H(N) counts the prime pairs with both summands at least 55, together with the admissible endpoint, and τ0(N)\tau_0(N) is a residual covering number. Fractional covering and disjoint-product packing admit the same reduction, although their optima need not equal the integer covering number. The same prime-pair requirements force k(N,P)≥H(N)−1k(N,P) \ge H(N)-1. Under the conjectural bound H(N)≫N/log⁡2NH(N) \gg N/\log^{2} N, Remark 6.5 therefore gives J(k)≍k(log⁡k)2J(k) \asymp k(\log k)^{2}.

To describe the first vanishing step, write P=QpP=Qp and put

SQ(N)={1≤h<N:gcd⁡(h(N−h),6Q)=1}.S_Q(N)=\{1\le h<N:\gcd(h(N-h),6Q)=1\}.

Its occupancies satisfy cr=cδp(N−r)c_r=c_{\delta_p(N-r)}. At a first zero, this nonempty set lies entirely in the one or two classes removed by pp. Theorem 3.3 characterizes this concentration through a sharp second-moment bound with prescribed central mass. The arithmetic estimates below compare survivor counts with class capacities to exclude such a step.

The distinction between total and noncolliding factors is also useful in explicit ranges. Discrepancy estimates show that supports with at most eleven factors cannot vanish for N>2780460N>2780460, and supports with at most seven noncolliding factors cannot vanish for N≥2266292N\ge2266292, with no restriction on the size of the factors. These effective conclusions complement (1.3), whose constants have not been made explicit.

The parameter J(k)J(k) is related in purpose to Jacobsthal-type covering problems. Iwaniec [5] proved a bound of order (klog⁡k)2(k\log k)^2 for the maximal run covered by kk primes in the ordinary one-class problem. Ziller and Morack [10] study paired progressions and their phases. Nguyen [7] studies finite-window CRT intersections and shift correlations for symmetric pairs on primorial wheels. Here the classes are tied to the fixed sum NN, and colliding factors are not charged. The parameter J(k)J(k) measures the extremal size under this noncolliding budget.

We call pmax⁡2<Np_{\max}^{2}<N the natural cutoff. With Π(N)=∏5≤q≤Nq\Pi(N)=\prod_{5\le q\le\sqrt{N}}q, factorization gives

gΠ(N)(N)=1N−1 prime+#{N<λ≤ϱ:λ,ϱ prime, λ+ϱ=N}.(6)g_{\Pi(N)}(N)=1_{N-1\ \mathrm{prime}}+\#\{\sqrt{N}<\lambda\le\varrho:\lambda,\varrho\ \mathrm{prime},\ \lambda+\varrho=N\}. \tag*{(6)}

Every composite summand below NN has a prime divisor below N\sqrt{N}, so the only admissible summands are 1 and primes exceeding N\sqrt{N}. Since NN is even, N\sqrt{N} cannot be an odd prime. Positivity at the complete cutoff therefore implies that every even N≥26N\ge26 with N−1N-1 composite is a sum of two primes exceeding N\sqrt{N}. This conclusion remains open here.

Sections 2–3 describe the count and the structure of a first zero. Section 4 gives explicit exclusions. Section 5 proves the weighted survival bound and the upper bound for JJ. Section 6 proves the cover reduction and the lower bound for JJ. Section 7 states the remaining quantitative gap.

Parametrization and discrepancy

Throughout, N=2n≥26N=2n\ge26. All prime products are squarefree and supported on primes at least 5. The empty product Q=1Q=1 is allowed. For P≠1P\ne1, let pmax⁡p_{\max} be its largest factor. For integers q≥1q\ge1 and x∈Zx\in\mathbb{Z}, we write the canonical remainder operator as in [8, 9]:

δq(x)=x−q⌊x/q⌋,Rq={0,1,…,q−1}.(7)\delta_q(x)=x-q\lfloor x/q\rfloor,\qquad\mathcal{R}_q=\{0,1,\ldots,q-1\}. \tag*{(7)}

Thus δq:Z→Rq\delta_q:\mathbb{Z}\to\mathcal{R}_q, including for negative xx, and δq(x+y)=δq(δq(x)+δq(y))\delta_q(x+y)=\delta_q(\delta_q(x)+\delta_q(y)). Divisibility is equivalently expressed by q∣xq\mid x or δq(x)=0\delta_q(x)=0. We reserve δ\delta for this operator and use ε>0\varepsilon>0 in exponent bounds.

Alongside (1.1), consider the ordered count

sQ(N)=#SQ(N),DQ(N)=1gcd⁡(n,6Q)=1.s_Q(N)=\#S_Q(N),\qquad D_Q(N)=\mathbf{1}_{\gcd(n,6Q)=1}.

The reflection h↦N−hh\mapsto N-h has only one possible fixed point, namely h=nh=n. Hence

sQ(N)=2gQ(N)−DQ(N),sQ(N)>0⟺gQ(N)>0.(8)s_Q(N)=2g_Q(N)-D_Q(N),\qquad s_Q(N)>0\Longleftrightarrow g_Q(N)>0. \tag*{(8)}

Put

ℓ={n/3,δ3(n)=0,(n+2)/3,δ3(n)=1,(n−2)/3,δ3(n)=2,β={1,δ3(n)≠2,5,δ3(n)=2,b6=1+13∣N.(9)\ell= \begin{cases} n/3, & \delta_3(n)=0,\\ (n+2)/3, & \delta_3(n)=1,\\ (n-2)/3, & \delta_3(n)=2, \end{cases} \qquad \beta= \begin{cases} 1, & \delta_3(n)\ne2,\\ 5, & \delta_3(n)=2, \end{cases} \qquad b_6=1+1_{3\mid N}. \tag*{(9)}

The parameter interval is 0≤z<ℓ0\le z<\ell, with h=6z+βh=6z+\beta. If 3∣n3\mid n, this family represents one of the two admissible values of δ6(h)\delta_6(h), and its reflection represents the other. If 3∤n3\nmid n, it represents all permitted summands. Consequently, setting

CQ(N)=#{0≤z<ℓ:gcd⁡((6z+β)(N−6z−β),Q)=1},(10)C_Q(N)=\#\{0\le z<\ell:\gcd((6z+\beta)(N-6z-\beta),Q)=1\}, \tag*{(10)}

we have

sQ(N)=b6CQ(N),ℓ≥N−46.(11)s_Q(N)=b_6C_Q(N),\qquad\ell\ge\frac{N-4}{6}. \tag*{(11)}

This is the parametrization of [9], written for the ordered count.

For each q∣Qq\mid Q, let ιq∈{1,…,q−1}\iota_q\in\{1,\ldots,q-1\} be the unique integer satisfying δq(6ιq)=1\delta_q(6\iota_q)=1. The excluded parameter residues are

δq(−ιqβ),δq(ιq(N−β)).\delta_q(-\iota_q\beta),\qquad\delta_q(\iota_q(N-\beta)).

Their number is νq=2−1q∣N\nu_q=2-1_{q\mid N}. Define

ρQ=∏q∣Q(1−νqq),DQ=∏q∣Q(1+νq)−∏q∣Q(1+νqq).(12)\rho_Q=\prod_{q\mid Q}\left(1-\frac{\nu_q}{q}\right),\qquad D_Q=\prod_{q\mid Q}(1+\nu_q)-\prod_{q\mid Q}\left(1+\frac{\nu_q}{q}\right). \tag*{(12)}

The dependence on NN is suppressed in ρQ\rho_Q and DQD_Q. The discrepancy bound DQD_Q is distinct from the diagonal indicator DQ(N)D_Q(N). For Q=1Q=1, we have ρQ=1\rho_Q=1 and DQ=0D_Q=0.

Lemma 2.1 (Parameter discrepancy). We have

∣CQ(N)−ℓρQ∣≤DQ.(13)|C_Q(N)-\ell\rho_Q|\le D_Q. \tag*{(13)}

If Q≠1Q\ne1, then DQ<∏q∣Q(1+νq)−1D_Q<\prod_{q\mid Q}(1+\nu_q)-1. In particular,

sQ(N)≥b6(ℓρQ−DQ).(14)s_Q(N)\ge b_6(\ell\rho_Q-D_Q). \tag*{(14)}

Proof. We recall the proof of [9 Proposition 5.2]. For each d∣Qd\mid Q, choosing one excluded class at every prime dividing dd gives ∏q∣dνq\prod_{q\mid d}\nu_q distinct prescribed values of δd(z)\delta_d(z) by the Chinese remainder theorem. The count of each prescribed value in an interval of length ℓ\ell differs from ℓ/d\ell/d by at most 1−1/d1-1/d. Inclusion–exclusion gives the main term

ℓ∑d∣Qμ(d)∏q∣dνqd=ℓρQ.\ell\sum_{d\mid Q}\mu(d)\frac{\prod_{q\mid d}\nu_q}{d}=\ell\rho_Q.

The sum of the remainder bounds is

∑d∣Q(∏q∣dνq)(1−1d)=DQ.\sum_{d\mid Q}\left(\prod_{q\mid d}\nu_q\right)\left(1-\frac{1}{d}\right)=D_Q.

The second product in (2.6) exceeds 1 if Q≠1Q\ne1. Finally, apply (2.5).

The structure of a first zero

If the factors are imposed in increasing order, the first vanishing step must concentrate the preceding set in the classes removed by the last prime. We first analyze that concentration without an ordering assumption. Let P=QpP=Qp, where p∤6Qp\nmid6Q. We do not yet require pp to be the largest factor of PP. Abbreviate

S=SQ(N),s=∣S∣,cr=#{h∈S:δp(h)=r},r∈Rp.(15)S=S_Q(N),\qquad s=|S|,\qquad c_r=\#\{h\in S:\delta_p(h)=r\},\qquad r\in\mathcal{R}_p. \tag*{(15)}

Reflection of the summands gives

cr=cδp(N−r).(16)c_r=c_{\delta_p(N-r)}. \tag*{(16)}

All occupancy indices belong to Rp\mathcal{R}_p. For each phase a∈Rpa\in\mathcal{R}_p, define

Y(a)={s−cδp(−a)−cδp(N−a),p∤N,s−cδp(−a),p∣N.(17)Y(a)= \begin{cases} s-c_{\delta_p(-a)}-c_{\delta_p(N-a)}, & p\nmid N,\\ s-c_{\delta_p(-a)}, & p\mid N. \end{cases} \tag*{(17)}

These phases arise by translating hh by multiples of 6Q6Q. Every condition at QQ remains unchanged, while δp(6Qj)\delta_p(6Qj) runs through Rp\mathcal{R}_p as jj does. Translated values need not lie in the original interval. The phase counting the representations of NN is a=0a=0.

Put

ν=2−1p∣N,μp=s(1−νp),Vp=1p∑a∈Rp(Y(a)−μp)2.(18)\nu=2-\mathbf{1}_{p\mid N},\qquad\mu_p=s\left(1-\frac{\nu}{p}\right),\qquad V_p=\frac{1}{p}\sum_{a\in\mathcal{R}_p}(Y(a)-\mu_p)^2. \tag*{(18)}

and

A=∑r∈Rpcr2,B=∑r∈Rpcrcδp(r+N).(19)A=\sum_{r\in\mathcal{R}_p}c_r^2,\qquad B=\sum_{r\in\mathcal{R}_p}c_rc_{\delta_p(r+N)}. \tag*{(19)}

The mean of YY is μp\mu_p, since each occupancy occurs ν\nu times in the sum of the losses.

Proposition 3.1 (Variance identity and pointwise bound). Let t=Y(0)=sP(N)=2gP(N)−DP(N)t=Y(0)=s_P(N)=2g_P(N)-D_P(N). Then

Vp={2p(A+B)−4s2p2,p∤N,Ap−s2p2,p∣N.(20)V_p= \begin{cases} \frac{2}{p}(A+B)-\frac{4s^2}{p^2}, & p\nmid N,\\ \frac{A}{p}-\frac{s^2}{p^2}, & p\mid N. \end{cases} \tag*{(20)}

and

t≥{s−2p3Vp+8s23p,p∤N,s−pVp+s2p,p∣N.(21)t\ge \begin{cases} s-\sqrt{\frac{2p}{3}V_p+\frac{8s^2}{3p}}, & p\nmid N,\\ s-\sqrt{pV_p+\frac{s^2}{p}}, & p\mid N. \end{cases} \tag*{(21)}

Proof. Expanding the square of the losses in (17) gives (20). If p∤Np\nmid N, symmetry yields c0=cδp(N)=(s−t)/2c_0=c_{\delta_p(N)}=(s-t)/2. Nonnegativity gives

A≥(s−t)22,B≥(s−t)24.A\ge\frac{(s-t)^2}{2},\qquad B\ge\frac{(s-t)^2}{4}.

Substitution into (20) yields Vp+4s2/p2≥3(s−t)2/(2p)V_p+4s^2/p^2\ge3(s-t)^2/(2p), proving the first bound. If p∣Np\mid N, then c0=s−tc_0=s-t and A≥(s−t)2A\ge(s-t)^2, which proves the second.

For s>0s > 0, write U(N,Q,p)\mathrm{U}(N,Q,p) for the inequality

Vp<{3p−82p2s2,p∤N,p−1p2s2,p∣N.(22)V_p < \begin{cases} \dfrac{3p-8}{2p^2}s^2, & p \nmid N,\\ \dfrac{p-1}{p^2}s^2, & p \mid N. \end{cases} \tag*{(22)}

By the preceding proposition, U(N,Q,p)\mathrm{U}(N,Q,p) implies gP(N)>0g_P(N) > 0. Using (20), its two branches become

A+B<34s2(p∤N),A<s2(p∣N).(23)A+B < \frac{3}{4}s^2 \quad(p \nmid N), \qquad A < s^2 \quad(p \mid N). \tag*{(23)}

Equality structure

When the two excluded classes coincide, the criterion has an immediate interpretation.

Proposition 3.2 (Collision). If p∣Np \mid N and s>0s > 0, then A≤s2A \le s^2 and

U(N,Q,p)⟺gP(N)>0.\mathrm{U}(N,Q,p) \Longleftrightarrow g_P(N) > 0.

Proof. The identity s2−A=2∑r<r′crcr′s^2-A=2\sum_{r<r'}c_rc_{r'} shows that equality holds precisely when a single class carries mass. By (16), this class must satisfy δp(2r)=0\delta_p(2r)=0. Since pp is odd and r∈Rpr\in R_p, it is 00. This is equivalent to t=0t=0.

For the other branch, write b=δp(N)≠0b=\delta_p(N)\ne0 and

u=cδp(n),v=s−u.(24)u=c_{\delta_p(n)}, \qquad v=s-u. \tag*{(24)}

The residue δp(n)\delta_p(n) is the unique fixed point of r↦δp(N−r)r\mapsto\delta_p(N-r). The quantity uu counts an entire class, possibly containing several integers n+jpn+jp. It need not equal the diagonal indicator DQ(N)D_Q(N).

Theorem 3.3 (Second moment with prescribed central mass). Let p≥5p\ge5 be prime, let b∈Rp∖{0}b\in R_p\setminus\{0\}, and let the nonnegative real numbers crc_r satisfy cr=cδp(b−r)c_r=c_{\delta_p(b-r)}. Put

r∗=δp((p+1)b2),s=∑r∈Rpcr,u=cr∗,v=s−u.r_*=\delta_p\left(\frac{(p+1)b}{2}\right), \qquad s=\sum_{r\in R_p}c_r, \qquad u=c_{r_*}, \qquad v=s-u.

Then

E(c):=∑rcr2+∑rcrcδp(r+b)≤max⁡{u2+34v2, u2+uv+12v2}.(25)\mathcal{E}(c):=\sum_r c_r^2+\sum_r c_rc_{\delta_p(r+b)} \le\max\left\{u^2+\frac{3}{4}v^2,\ u^2+uv+\frac{1}{2}v^2\right\}. \tag*{(25)}

The bound is attained for every pair of real numbers u,v≥0u,v\ge0. If u=0u=0 and s>0s>0, then

E(c)≤34s2,(26)\mathcal{E}(c)\le\frac{3}{4}s^2, \tag*{(26)}

and equality holds precisely when c0=cb=s/2c_0=c_b=s/2 and all other occupancies vanish.

Proof. The quadratic form can be written as

E(c)=12∑r∈Rp(cr+cδp(r+b))2.(27)\mathcal{E}(c)=\frac{1}{2}\sum_{r\in R_p}\left(c_r+c_{\delta_p(r+b)}\right)^2. \tag*{(27)}

It is positive definite. Indeed, a vector in its kernel would satisfy cδp(r+b)=−crc_{\delta_p(r+b)}=-c_r for every rr. Addition by bb traverses a cycle of odd length pp, so cr=−crc_r=-c_r and the vector vanishes. Thus E\mathcal{E} is strictly convex.

The reflection r↦δp(b−r)r\mapsto\delta_p(b-r) has one fixed point r∗r_\ast and (p−1)/2(p-1)/2 two-point orbits. Once uu is fixed, the occupancies are described by distributing the mass vv among these orbits, with equal mass at the two points of each orbit. The set of distributions is a simplex. Convexity allows E\mathcal{E} to be maximized at a vertex, with all the mass vv on a single orbit.

For the orbit {0,b}\{0,b\}, the two points are adjacent under the step bb, and neither is adjacent to the fixed point. The value is u2+3v2/4u^2+3v^2/4. For the orbit {δp(−r∗),δp(3r∗)}\{\delta_p(-r_\ast),\delta_p(3r_\ast)\}, both points are adjacent to the fixed point but not to each other, and the value is u2+uv+v2/2u^2+uv+v^2/2. Every other orbit gives u2+v2/2u^2+v^2/2. These classifications use p≥5p\geq5. The two special orbits exist and are distinct even for p=5p=5. This proves the bound and its sharpness.

If u=0u=0, the only vertex attaining 3s2/43s^2/4 is the orbit {0,b}\{0,b\}. Strict convexity rules out equality at a nontrivial combination of vertices, proving the equality characterization. □

Corollary 3.4 (Central survival or strict variance). Suppose p∤Np\nmid N and s>0s>0. Then

gQp(N)>0⟺u>0 or U(N,Q,p),u=cδp(n).g_{Qp}(N)>0 \quad\Longleftrightarrow\quad u>0\ \text{or}\ U(N,Q,p),\qquad u=c_{\delta_p(n)}.

Proof. The central residue is distinct from 00 and δp(N)\delta_p(N), so u>0u>0 supplies a survivor. If u=0u=0, Theorem 3.3 shows that failure of the strict criterion is equivalent to concentration in the excluded classes. The converse implications follow from these same observations and Proposition 3.1. □

For example, N=26N=26, Q=35Q=35, and p=11p=11 give SQ(N)={13}S_Q(N)=\{13\}. Here s=u=1s=u=1, so A+B=1>3/4A+B=1>3/4, although g385(26)=1g_{385}(26)=1. The variance criterion is therefore sufficient in general and exact in the empty-central-class branch. The positivity estimates below are obtained independently from the complete survivor set.

Class capacity and explicit positivity

Each value of δp(h)\delta_p(h) is taken by a limited number of integers hh in the interval with gcd⁡(h(N−h),6)=1\gcd(h(N-h),6)=1. This allows us to compare a lower bound for the preceding set with an upper bound for the loss.

Lemma 4.1 (Residue-class capacity). With the notation of (3.1), let

MN,p=b6⌈N−16p⌉.(28)M_{N,p}=b_6\left\lceil\frac{N-1}{6p}\right\rceil. \tag*{(28)}

Then

max⁡rcr≤MN,p,A≤MN,ps,B≤MN,ps.(29)\max_r c_r\leq M_{N,p},\qquad A\leq M_{N,p}s,\qquad B\leq M_{N,p}s. \tag*{(29)}

Proof. The condition gcd⁡(h(N−h),6)=1\gcd(h(N-h),6)=1 permits b6b_6 values of δ6(h)\delta_6(h). Together with a fixed value of δp(h)\delta_p(h), each determines one value of δ6p(h)\delta_{6p}(h). An interval of length N−1N-1 contains at most ⌈(N−1)/(6p)⌉\lceil(N-1)/(6p)\rceil integers with a prescribed value of δ6p(h)\delta_{6p}(h). The restrictions at QQ can only decrease the count. The bounds for AA and BB follow by bounding one factor of each summand by MN,pM_{N,p}. □

Proposition 4.2 (Two capacity criteria). If s>0s > 0, the conditions

s>83MN,p(p∤N),s>MN,p(p∣N)s > \frac{8}{3}M_{N,p}\quad(p \nmid N),\qquad s > M_{N,p}\quad(p \mid N)

are sufficient for U(N,Q,p)\mathrm{U}(N,\mathbb{Q},p). To obtain gQp(N)>0g_{\mathbb{Q}p}(N) > 0, it is enough that

s>νMN,p,ν=2−1p∣N.(30)s > \nu M_{N,p},\qquad\nu= 2 - 1_{p\mid N}. \tag*{(30)}

Proof. The first assertions follow from (4.2) and (3.9). For the last one, symmetry gives

t=s−νc0≥s−νMN,p.(31)t=s-\nu c_0\ge s-\nu M_{N,p}. \tag*{(31)}

Thus an estimate based on the maximum occupancy can prove positivity under a weaker hypothesis than that needed for the variance criterion. The independent bound (2.8) allows both comparisons.

Complete blocks of the preceding modulus

Proposition 4.3. Suppose Q>1Q > 1 and pp exceeds every prime factor of QQ. Let AQ=QρQ=∏q∣Q(q−νq)A_Q=Q\rho_Q=\prod_{q\mid Q}(q-\nu_q). Then

CQp(N)≥⌊ℓQ⌋⌊AQ−ν⌈Qp⌉⌋,ν=2−1p∣N.(32)C_{Qp}(N)\ge\left\lfloor\frac{\ell}{Q}\right\rfloor\left\lfloor A_Q-\nu\left\lceil\frac{Q}{p}\right\rceil\right\rfloor,\qquad\nu=2-1_{p\mid N}. \tag*{(32)}

The bracket is at least 1. In particular, ℓ≥Q\ell\ge Q implies gQp(N)>0g_{Qp}(N)>0, and the same holds if N≥6QN\ge6Q.

Proof. Each block of QQ consecutive parameters contains exactly AQA_Q survivors of the restrictions at QQ. Each value of δp(z)\delta_p(z) occurs at most ⌈Q/p⌉\lceil Q/p\rceil times in the block. Summing the complete blocks proves (4.5).

Let A0=∏q∣Q(q−2)A_0=\prod_{q\mid Q}(q-2). This is an odd integer at least 3, and AQ≥A0A_Q\ge A_0. The telescoping product over odd integers from 5 to p−2p-2 gives

A0Q≥3p−2,Qp<A03≤A0−12.\frac{A_0}{Q}\ge\frac{3}{p-2},\qquad\frac{Q}{p}<\frac{A_0}{3}\le\frac{A_0-1}{2}.

Since the last quantity is an integer, 2⌈Q/p⌉≤A0−12\lceil Q/p\rceil\le A_0-1. The bracket in (4.5) is therefore at least 1. Finally, the three branches of (2.3) show that N≥6QN\ge6Q implies ℓ≥Q\ell\ge Q.

Thresholds for a given support

Assume now that every prime factor of QQ is less than pp. Put

ρ0=∏q∣Q(1−2q),D0=3ω(Q)−∏q∣Q(1+2q).(33)\rho_0=\prod_{q\mid Q}\left(1-\frac{2}{q}\right),\qquad D_0=3^{\omega(Q)}-\prod_{q\mid Q}\left(1+\frac{2}{q}\right). \tag*{(33)}

We have ρQ≥ρ0\rho_Q\ge\rho_0 and DQ≤D0D_Q\le D_0. For the second comparison, the expression for DQD_Q in (2.6) increases with each νq\nu_q. Also,

ρ0≥∏5≤j≤p−2j oddj−2j=3p−2.(34)\rho_0\ge\prod_{\substack{5\le j\le p-2\\j\ \mathrm{odd}}}\frac{j-2}{j}=\frac{3}{p-2}. \tag*{(34)}

The product on the right includes additional factors smaller than one.

Theorem 4.4 (Explicit thresholds). Let QQ be a product of primes in [5,p)[5,p), with p≥5p \ge5 prime, and let ρ0,D0\rho_0,D_0 be as in (33). Then

N>4ρ0+6D0+12−2/pρ0−2/p⟹gQp(N)>0.(35)N > \frac{4\rho_0+6D_0+12-2/p}{\rho_0-2/p} \Longrightarrow g_{Qp}(N)>0. \tag*{(35)}

Furthermore,

N>4ρ0+6D0+16−8/(3p)ρ0−8/(3p)⟹s>0 and U(N,Q,p).(36)N > \frac{4\rho_0+6D_0+16-8/(3p)}{\rho_0-8/(3p)} \Longrightarrow s>0\ \text{and}\ U(N,Q,p). \tag*{(36)}

Both denominators are positive.

Proof. By Lemma 2.1, s≥b6((N−4)ρ0/6−D0)s \ge b_6((N-4)\rho_0/6-D_0). Also, MN,p≤b6((N−1)/(6p)+1)M_{N,p} \le b_6((N-1)/(6p)+1). Comparing the first expression with 2MN,p2M_{N,p} gives (35), and comparing it with 8MN,p/38M_{N,p}/3 gives (36). The latter comparison also covers the branch p∣Np \mid N. Finally, (34) implies ρ0>8/(3p)>2/p\rho_0>8/(3p)>2/p.

The term D0D_0 retains exponential growth in the number of prime factors of QQ. The theorem uses the fact that the last prime need not be included in the inclusion–exclusion remainder. This improvement of the finite bounds does not by itself cover all supports at the natural cutoff.

Supports with at most eleven prime factors

Let ℓ1=5,ℓ2=7,…\ell_1=5,\ell_2=7,\ldots be the primes starting at 5. For r≥1r\ge1, define

ρr=∏i<r(1−2ℓi),Er=3r−1−1,Tr=4ρr+6Er+12−2/ℓrρr−2/ℓr.(37)\rho_r=\prod_{i<r}\left(1-\frac{2}{\ell_i}\right),\qquad E_r=3^{r-1}-1,\qquad T_r=\frac{4\rho_r+6E_r+12-2/\ell_r}{\rho_r-2/\ell_r}. \tag*{(37)}

Lemma 4.5. If ω(P)=r\omega(P)=r and N>TrN>T_r, then gP(N)>0g_P(N)>0.

Proof. Remove the largest prime pp. The prime factors of Q=P/pQ=P/p are at least ℓ1,…,ℓr−1\ell_1,\ldots,\ell_{r-1}, and p≥ℓrp\ge\ell_r. Hence ρQ≥ρr\rho_Q\ge\rho_r and DQ≤ErD_Q\le E_r. Using (14) and (31),

t≥b66[N(ρr−2ℓr)−4ρr−6Er−12+2ℓr].t\ge\frac{b_6}{6}\left[N\left(\rho_r-\frac{2}{\ell_r}\right)-4\rho_r-6E_r-12+\frac{2}{\ell_r}\right].

The substitution p≥ℓrp\ge\ell_r is valid in the term −2(N−1)/p-2(N-1)/p. The right-hand side is positive for N>TrN>T_r.

Table 1 lists the ceilings of the first eleven bounds. All comparisons use exact fractions. In particular,

max⁡1≤r≤11Tr=T11=9465510038258634042961=2780460.265562…(38)\max_{1\le r\le11}T_r=T_{11}=\frac{94655100382586}{34042961}=2780460.265562\ldots \tag*{(38)}
rr1234567891011
ℓr\ell_r57111317192329313741
⌈Tr⌉\lceil T_r\rceil26842508602754935929745906032925068932412780461

Table 1. Thresholds by cardinality. The sufficient condition is N>TrN>T_r.

Theorem 4.6 (Eleven prime factors). If 1≤ω(P)≤111 \le\omega(P) \le11 and N>2780460N > 2780460, then gP(N)>0g_P(N)>0. No restriction on the size of the factors is needed.

Proof. By (4.11), every even N>2780460N > 2780460 satisfies N>TrN > T_r for 1≤r≤111 \le r \le11. Apply Lemma 4.5.

Prime factors dividing half the integer

Discrepancy also allows us to retain information about collisions. Write

C=gcd⁡(P,n),R=P/C,c=ω(C),k=ω(R).(39)C=\gcd(P,n),\qquad R=P/C,\qquad c=\omega(C),\qquad k=\omega(R). \tag*{(39)}

The prime factors of CC and RR form disjoint sets, C∣nC\mid n, and gcd⁡(R,n)=1\gcd(R,n)=1. For this support, Lemma 2.1 implies

ℓ∏q∣C(1−1q)∏q∣R(1−2q)≥2c3k−1⟹gP(N)>0.(40)\ell\prod_{q\mid C}\left(1-\frac{1}{q}\right)\prod_{q\mid R}\left(1-\frac{2}{q}\right)\ge2^{c}3^{k}-1 \quad\Longrightarrow\quad g_P(N)>0. \tag*{(40)}

The non-strict hypothesis is valid because the discrepancy bound by 2c3k−12^{c}3^{k}-1 is strict. For c≥1c\ge1, let Cc=∏i=1cℓiC_c=\prod_{i=1}^{c}\ell_i, and define

Bc,k=φ(Cc)3(1−2Cc)∏i=c+1c+k(1−2ℓi).(41)B_{c,k}=\frac{\varphi(C_c)}{3}\left(1-\frac{2}{C_c}\right)\prod_{i=c+1}^{c+k}\left(1-\frac{2}{\ell_i}\right). \tag*{(41)}

Lemma 4.7 (Extrema with disjoint groups). If c≥1c\ge1, the left-hand side of (4.13) is at least Bc,kB_{c,k}. Moreover, for all c,k≥0c,k\ge0 with c+k≥1c+k\ge1, a sufficient condition for positivity is

n≥Θc,k:=3(2c3k−1)∏i=1kℓiℓi−2∏i=k+1k+cℓiℓi−1+2.(42)n\ge\Theta_{c,k}:=3(2^{c}3^{k}-1)\prod_{i=1}^{k}\frac{\ell_i}{\ell_i-2}\prod_{i=k+1}^{k+c}\frac{\ell_i}{\ell_i-1}+2. \tag*{(42)}

Proof. Since n≥Cn\ge C and ℓ≥(n−2)/3\ell\ge(n-2)/3, the left-hand side of (4.13) is at least

φ(C)3(1−2C)∏q∣R(1−2q).(43)\frac{\varphi(C)}{3}\left(1-\frac{2}{C}\right)\prod_{q\mid R}\left(1-\frac{2}{q}\right). \tag*{(43)}

Replacing a prime by a smaller unused prime decreases this expression, regardless of its group. Its minimum therefore uses the first c+kc+k primes. If a<ba<b, assigning aa to CC and bb to RR decreases the product of factors, since

(a−1)b−2b<(b−1)a−2a.(a-1)\frac{b-2}{b}<(b-1)\frac{a-2}{a}.

It also decreases CC, and hence the factor 1−2/C1-2/C. The minimizing assignment puts the cc smallest primes in CC, proving the first assertion. Without using n≥Cn \ge C, criterion (40) gives the threshold

3(2c3k−1)∏q∣Cqq−1∏q∣Rqq−2+2.3(2c3^k-1)\prod_{q\mid C}\frac{q}{q-1}\prod_{q\mid R}\frac{q}{q-2}+2.

Its factors decrease as the primes increase. If a<ba<b, the ratio of the assignment a∈R,b∈Ca\in R,b\in C to the opposite assignment is

(a−1)(b−2)(a−2)(b−1)>1.\frac{(a-1)(b-2)}{(a-2)(b-1)}>1.

The maximum assigns the kk smallest primes to RR and gives (42).

Theorem 4.8 (Seven noncolliding factors). If ω(P/gcd⁡(P,n))≤7\omega(P/\gcd(P,n))\le7 and n≥1133146n\ge1133146, then gP(N)>0g_P(N)>0, with no restriction on ω(P)\omega(P) or on the size of its factors. If at least six factors of PP divide nn, the restriction on nn is unnecessary.

Proof. If P=1P=1, positivity follows from (11). Suppose P>1P>1. For c=6c=6, k=7k=7, we have exactly

B6,7=34212642163660800189431751553>139967=2637−1.B_{6,7}=\frac{34212642163660800}{189431751553}>139967=2^6 3^7-1.

Decreasing kk increases Bc,kB_{c,k} and decreases 2c3k−12^c3^k-1. As cc increases, the ratio Bc+1,k/Bc,kB_{c+1,k}/B_{c,k} exceeds ℓc+1−1≥4\ell_{c+1}-1\ge4, whereas 2c3k−12^c3^k-1 grows by a factor at most 3 for c≥1c\ge1. Thus Lemma 4.7 covers all cases c≥6c\ge6, k≤7k\le7.

If c≤5c\le5 and k≤7k\le7, the maximum Θc,k\Theta_{c,k} does not exceed Θ5,7\Theta_{5,7}. Monotonicity follows by adding factors to the maximization problem in the lemma. The exact value is

Θ5,7=2331480779877798720575296000<1133146.\Theta_{5,7}=\frac{23314807798777987}{20575296000}<1133146.

Consequently, all n≥1133146n\ge1133146 are covered.

Theorem 4.9 (Restrictions on any zero). If n≥13n\ge13 and gP(2n)=0g_P(2n)=0, then k≥1k\ge1 and

(c+1)!<5(2k+3)33k,(44)(c+1)!<\frac{5(2k+3)}{3}3^k, \tag*{(44)}
n<(2(c+k)+3)(2c3k−1)+2.(45)n<(2(c+k)+3)(2^c3^k-1)+2. \tag*{(45)}

These conditions do not require the natural cutoff.

Proof. If k=0k=0, every prime factor of PP divides nn. The displacement d=212∤n313∤n≤6<nd=2^{1_{2\nmid n}}3^{1_{3\nmid n}}\le6<n produces the pair n−d,n+dn-d,n+d, coprime to 6P6P. This contradicts vanishing.

For c≥1c\ge1, the iith prime factor of CC is at least 2i+32i+3, so

φ(C)≥2c(c+1)!,1−2/C≥3/5.\varphi(C)\ge2^c(c+1)!,\qquad1-2/C\ge3/5.

Furthermore,

∏q∣R(1−2q)≥∏i=1k2i+12i+3=32k+3.\prod_{q\mid R}\left(1-\frac{2}{q}\right)\ge\prod_{i=1}^{k}\frac{2i+1}{2i+3}=\frac{3}{2k+3}.

Expression (43) is therefore at least 3 2c(c+1)!/(5(2k+3))3\,2^c(c+1)!/(5(2k+3)). If (44) failed, the threshold in (40) would be reached. For c=0c=0, (44) is immediate because k≥1k\ge1. Put r=c+kr=c+k and ρ=∏q∣C(1−1/q)∏q∣R(1−2/q)\rho=\prod_{q\mid C}(1-1/q)\prod_{q\mid R}(1-2/q). Vanishing requires (n−2)ρ/3<233k−1(n-2)\rho/3<2^{3}3^{k}-1. Moreover,

ρ−1≤∏q∣Rqq−2≤∏i=1r2i+32i+1=2r+33.\rho^{-1}\le\prod_{q\mid R}\frac{q}{q-2}\le\prod_{i=1}^{r}\frac{2i+3}{2i+1}=\frac{2r+3}{3}.

Combining these bounds proves (45).

The elementary restrictions above are effective and retain information about both groups of primes. Equation (57) below gives a stronger asymptotic restriction by using the distribution of primes.

Sparse supports and the size of a vanishing instance

We now estimate the number of prime factors needed to eliminate every representation. The argument constructs a family with a prime summand and compares its size with the loss at each factor. Keeping the same local density factor in both estimates is essential for uniformity.

A family supplied by the linear sieve

For h>1h>1, let P−(h)P^{-}(h) denote its least prime factor. Fix 0<θ<1/40<\theta<1/4 and put

Tθ(N)={2≤h<N/2:N−h prime, P−(h)>Nθ}.(46)\mathcal{T}_{\theta}(N)=\{2\le h<N/2:N-h\text{ prime},\ P^{-}(h)>N^{\theta}\}. \tag*{(46)}

Write

S(N)=∏ℓ∣Nℓ>2ℓ−1ℓ−2,Sw(N)=∏2<ℓ<wℓ∤Nℓ−1ℓ−2,(47)\mathfrak{S}(N)=\prod_{\substack{\ell\mid N\\\ell>2}}\frac{\ell-1}{\ell-2},\qquad \mathfrak{S}_{w}(N)=\prod_{\substack{2<\ell<w\\\ell\nmid N}}\frac{\ell-1}{\ell-2}, \tag*{(47)}

where the products run over primes. We have S(N)≥1\mathfrak{S}(N)\ge1.

Lemma 5.1 (Shifted-prime sieve). For every 0<θ<1/40<\theta<1/4, there are cθ>0c_{\theta}>0 and N0(θ)N_{0}(\theta) such that

∣Tθ(N)∣≥cθS(N)Nlog⁡2N(48)|\mathcal{T}_{\theta}(N)|\ge c_{\theta}\mathfrak{S}(N)\frac{N}{\log^{2}N} \tag*{(48)}

for all even N≥N0(θ)N\ge N_{0}(\theta). Both summands of every representation in this family are coprime to 6N6N.

Proof. Consider AN={N−r:N/2<r<N, r prime}\mathcal{A}_{N}=\{N-r:N/2<r<N,\ r\text{ prime}\} and XN=li⁡(N)−li⁡(N/2)≍N/log⁡NX_{N}=\operatorname{li}(N)-\operatorname{li}(N/2)\asymp N/\log N. A prime r>N/2r>N/2 cannot divide the even integer NN, so every element is coprime to NN. For squarefree dd, the divisibility density is

gN(d)={1/φ(d),(d,N)=1,0,(d,N)>1.g_{N}(d)= \begin{cases} 1/\varphi(d), & (d,N)=1,\\ 0, & (d,N)>1. \end{cases}

Thus ∣(AN)d∣=XNgN(d)+Rd|(\mathcal{A}_{N})_{d}|=X_{N}g_{N}(d)+R_{d}. Bombieri–Vinogradov gives, for any A>0A>0 and a suitable BB,

∑d≤Dμ2(d)∣Rd∣≪ANlog⁡−AN,D=N1/2/log⁡BN.(49)\sum_{d\le D}\mu^{2}(d)|R_{d}|\ll_{A}N\log^{-A}N,\qquad D=N^{1/2}/\log^{B}N. \tag*{(49)}

For d≤Dd \le D with (d,N)>1(d,N)>1, the count and its density are both zero. For the other dd, the estimate is uniform in NN, since it uses the maximum over reduced classes at both endpoints of the prime interval.

The lower linear sieve applies uniformly after deleting the odd primes dividing NN. Its density product is

VN(z)=∏3≤ℓ<zℓ∤N(1−1ℓ−1)≍Sz(N)log⁡z.V_N(z)=\prod_{\substack{3\le\ell<z\\ \ell\nmid N}}\left(1-\frac{1}{\ell-1}\right)\asymp\frac{\mathcal{S}_z(N)}{\log z}.

For z=2Nθz=2N^\theta,

0≤log⁡S(N)Sz(N)≪log⁡Nzlog⁡z=oθ(1).0\le\log\frac{\mathcal{S}(N)}{\mathcal{S}_z(N)}\ll\frac{\log N}{z\log z}=o_\theta(1).

Indeed, at most log⁡N/log⁡z\log N/\log z prime factors of NN are at least zz. The sieve parameter satisfies log⁡D/log⁡z→1/(2θ)>2\log D/\log z\to1/(2\theta)>2, so its lower function is positive. Consequently,

S(AN,z)≫θS(N)Nlog⁡2N.S(\mathcal{A}_N,z)\gg_\theta\mathcal{S}(N)\frac{N}{\log^2 N}.

Removing the possible element 11 proves (5.3). For Nθ>3N^\theta>3, the surviving summands are also coprime to 66.

This is the classical Goldbach-sequence argument in [6 Theorem 8 and Sections 4.2--4.3], with a variable power cutoff. An explicit general form of the linear sieve appears in [2 Theorem 6]. The restriction θ<1/4\theta<1/4 comes from its lower-sieve parameter and is not a new exponent for that classical construction.

The loss at an additional prime

The roughness of the complementary factor improves the loss estimate by one logarithm. The local factor S(N)\mathcal{S}(N) must be retained.

Lemma 5.2 (Two linear forms). Fix 0<θ<1/40<\theta<1/4 and let σ=min⁡(θ/2,1/100)\sigma=\min(\theta/2,1/100). For q=1q=1, or for any prime Nθ<q≤NN^\theta<q\le\sqrt{N} with q∤Nq\nmid N, put X=N/(2q)X=N/(2q) and w=Nσw=N^\sigma. Then, uniformly in qq and for even NN sufficiently large,

#{1≤m<X:gcd⁡(m(N−qm),∏ℓ<wℓ)=1}≪θS(N)Nqlog⁡2N.(50)\#\left\{1\le m<X:\gcd\left(m(N-qm),\prod_{\ell<w}\ell\right)=1\right\}\ll_\theta\mathcal{S}(N)\frac{N}{q\log^2 N}. \tag*{(50)}

Proof. No sieving prime divides qq. The polynomial m(N−qm)m(N-qm) has one root at ℓ=2\ell=2, one root at each odd ℓ∣N\ell\mid N, and two roots at all other primes ℓ<w\ell<w. Denote this number by ωN(ℓ)\omega_N(\ell) and extend it multiplicatively to squarefree integers. The count for a squarefree dd supported below ww is

XωN(d)d+Rd,∣Rd∣≤ωN(d)≤2ω(d).X\frac{\omega_N(d)}{d}+R_d,\qquad|R_d|\le\omega_N(d)\le2^{\omega(d)}.

The upper-bound sieve of dimension two applies with a bounded upper sieve function, uniformly in these local densities. Indeed, its inverse density products are bounded by

those with two roots at every odd prime. Take D=X/log⁡20ND = X/\log^{20} N. For sufficiently large NN, one has 40≤log⁡D/log⁡w≤1/σ40 \le\log D/\log w \le1/\sigma, and

∑d<Dμ2(d)4ω(d)∣Rd∣≤∑d<D9ω(d)≪D(1+log⁡D)7≪Xlog⁡−13N.\sum_{d<D}\mu^2(d)4^{\omega(d)}|R_d| \le\sum_{d<D} 9^{\omega(d)} \ll D(1+\log D)^7 \ll X\log^{-13}N.

The standard upper sieve therefore bounds the left side of (50) by

≪θX∏ℓ<w(1−ωN(ℓ)ℓ)+O(Xlog⁡−13N)≪θXSw(N)log⁡2w.\ll_\theta X\prod_{\ell<w}\left(1-\frac{\omega_N(\ell)}{\ell}\right)+O(X\log^{-13}N)\ll_\theta X\frac{\mathfrak{S}_w(N)}{\log^2w}.

Here the root at 22 contributes 1/21/2, and the correction at every odd divisor of NN is (ℓ−1)/(ℓ−2)(\ell-1)/(\ell-2). Since Sw(N)≤S(N)\mathfrak{S}_w(N)\le\mathfrak{S}(N) and log⁡w=σlog⁡N\log w=\sigma\log N, the claim follows. The upper sieve used here is the standard dimension-two specialization of [3 Theorem 9.1], with the divisibility remainders verified above.

Let Lq(N)=#{h∈Tθ(N):q∣h(N−h)}L_q(N)=\#\{h\in\mathcal{T}_\theta(N):q\mid h(N-h)\}. For a support PP, define the normalized loss bound

Wθ(N,P)=∑q∣PNθ<q≤Nq∤N1q+log⁡2NS(N)N∑q∣PN<q<Nq∤N(N2q+1).(51)W_\theta(N,P)= \sum_{\substack{q\mid P\\N^\theta<q\le\sqrt{N}\\q\nmid N}}\frac{1}{q} +\frac{\log^2N}{\mathfrak{S}(N)N} \sum_{\substack{q\mid P\\\sqrt{N}<q<N\\q\nmid N}}\left(\frac{N}{2q}+1\right). \tag*{(51)}

All sums over factors of a support run over primes. Also put

κθ(N,P)=#{q∣P:Nθ<q<N, q∤N}.(52)\kappa_\theta(N,P)=\#\{q\mid P:N^\theta<q<N,\ q\nmid N\}. \tag*{(52)}

Lemma 5.3 (Loss at one factor). For even NN sufficiently large,

Lq(N)≤{CθS(N)N/(qlog⁡2N),Nθ<q≤N, q∤N,N/(2q)+1,N<q<N, q∤N,0,otherwise.L_q(N)\le \begin{cases} C_\theta\mathfrak{S}(N)N/(q\log^2N), & N^\theta<q\le\sqrt{N},\ q\nmid N,\\ N/(2q)+1, & \sqrt{N}<q<N,\ q\nmid N,\\ 0, & \text{otherwise}. \end{cases}

In particular, max⁡qLq(N)≪θS(N)N1−θ/log⁡2N\max_q L_q(N)\ll_\theta\mathfrak{S}(N)N^{1-\theta}/\log^2N.

Proof. The zero-loss cases follow from Lemma 5.1 and the defining roughness. If Nθ<q≤NN^\theta<q\le\sqrt{N}, the prime summand exceeds N/2N/2 and cannot equal qq. Write h=qmh=qm. Every counted mm avoids all prime divisors below ww in Lemma 5.2, so that lemma applies. For q>Nq>\sqrt{N}, multiples of qq among h<N/2h<N/2 contribute at most N/(2q)N/(2q), and N−h=qN-h=q contributes at most one. Finally, N/2+1=oθ(N1−θ/log⁡2N)\sqrt{N}/2+1=o_\theta(N^{1-\theta}/\log^2N). □\square

Theorem 5.4 (Weighted survival and sparse supports). Fix 0<θ<1/40<\theta<1/4 and write x+=max⁡{x,0}x_+=\max\{x,0\}. There are positive constants cθ,Cθ,ηθc_\theta,C_\theta,\eta_\theta such that, for all sufficiently large even NN and every support PP,

gP(N)≥S(N)Nlog⁡2N(cθ−CθWθ(N,P))+.(53)g_P(N)\geq\mathcal{S}(N)\frac{N}{\log^2 N}\left(c_\theta-C_\theta W_\theta(N,P)\right)_+. \tag*{(53)}

In particular,

κθ(N,P)≤ηθNθ⟹gP(N)≫θS(N)Nlog⁡2N.(54)\kappa_\theta(N,P)\leq\eta_\theta N^\theta\quad\Longrightarrow\quad g_P(N)\gg_\theta\mathcal{S}(N)\frac{N}{\log^2 N}. \tag*{(54)}

There is no upper bound on the prime factors of PP.

Proof. At least ∣Tθ(N)∣−∑q∣PLq(N)|\mathcal{T}_\theta(N)|-\sum_{q\mid P}L_q(N) members survive. Lemmas 5.1 and 5.3, enlarging CθC_\theta to be at least 1, give (5.8). Each nonzero weight in (5.6) is at most N−θN^{-\theta} once NN is sufficiently large. For the second sum this follows from

log⁡2NS(N)N(N2+1)=oθ(N−θ).\frac{\log^2 N}{\mathcal{S}(N)N}\left(\frac{\sqrt{N}}{2}+1\right)=o_\theta(N^{-\theta}).

Hence Wθ(N,P)≤κθ(N,P)N−θW_\theta(N,P)\leq\kappa_\theta(N,P)N^{-\theta}. Choose ηθ=cθ/(2Cθ)\eta_\theta=c_\theta/(2C_\theta).

The same factor S(N)\mathcal{S}(N) multiplies the supply and the upper bound for each small-prime loss. Its cancellation explains the absence of a logarithm in (5.9). The weighted statement also permits many more factors when they lie close to N\sqrt{N}.

Corollary 5.5 (Cost and size of a zero). For every 0<θ<1/40<\theta<1/4, a vanishing instance with sufficiently large NN satisfies

κθ(N,P)≫θNθ.(55)\kappa_\theta(N,P)\gg_\theta N^\theta. \tag*{(55)}

Consequently, for each ε>0\varepsilon>0 and k=ω(P/gcd⁡(P,N/2))≥1k=\omega(P/\gcd(P,N/2))\geq1,

gP(N)=0⟹N≪εk4+ε.(56)g_P(N)=0\quad\Longrightarrow\quad N\ll_\varepsilon k^{4+\varepsilon}. \tag*{(56)}

Thus J(k)≪εk4+εJ(k)\ll_\varepsilon k^{4+\varepsilon}.

Proof. Vanishing requires every member of Tθ(N)\mathcal{T}_\theta(N) to be removed. Compare its lower bound with the maximum loss in Lemma 5.3 multiplied by κθ\kappa_\theta. The common factor S(N)/log⁡2N\mathcal{S}(N)/\log^2 N cancels. Since κθ≤k\kappa_\theta\leq k, take θ=1/(4+ε)\theta=1/(4+\varepsilon). The bounded remaining values of NN are absorbed into the implied constant.

For a sequence of zeros with N→∞N\to\infty, one also has k→∞k\to\infty and

c=O(log⁡klog⁡log⁡k),c=ω(gcd⁡(P,N/2)).(57)c=O\left(\frac{\log k}{\log\log k}\right),\qquad c=\omega(\gcd(P,N/2)). \tag*{(57)}

Indeed, (c+1)!≤gcd⁡(P,N/2)≤N/2(c+1)!\leq\gcd(P,N/2)\leq N/2, whereas (5.11) gives log⁡N=O(log⁡k)\log N=O(\log k).

Consequences for the first-zero structure

The arithmetic estimates also control the localized moments. Their role here is a consequence of the survivor mass, independently of the extremal characterization.

Lemma 5.6 (Mass for long intervals). Fix a>4a>4. There is p1(a)p_1(a) such that, for every prime p≥p1(a)p\ge p_1(a), every product QQ of primes in [5,p)[5,p), and every even N≥paN\ge p^a,

sQ(N)≫aNlog⁡2p.(58)s_Q(N) \gg_a \frac{N}{\log^2 p}. \tag*{(58)}

Proof. Let U≥5U\ge5 be an absolute constant, chosen below. If pa≤N<pUp^a\le N<p^U, put θ=1/a<1/4\theta=1/a<1/4. For pp sufficiently large in terms of aa, Lemma 5.1 applies. Every h∈Tθ(N)h\in\mathcal{T}_\theta(N) satisfies P−(h)>Nθ≥pP^-(h)>N^\theta\ge p, and N−hN-h is a prime exceeding N/2>pN/2>p. Thus both orientations survive QQ, and

sQ(N)≥2∣Tθ(N)∣≫aNlog⁡2N≫aNlog⁡2p,s_Q(N)\ge2|\mathcal{T}_\theta(N)|\gg_a \frac{N}{\log^2 N}\gg_a \frac{N}{\log^2 p},

using S(N)≥1\mathfrak{S}(N)\ge1 and log⁡N<Ulog⁡p\log N<U\log p.

Now let N≥pUN\ge p^U. Sieve the parameters 0≤z<ℓ0\le z<\ell by the νq\nu_q excluded values at each q∣Qq\mid Q. Their densities satisfy νq/q≤2/q\nu_q/q\le2/q. The inverse-product condition for a sieve of dimension two holds uniformly by comparison with two excluded values at every prime q≥5q\ge5. For squarefree d∣Qd\mid Q, the divisibility remainder obeys

∣Rd∣≤∏q∣dνq≤2ω(d).|R_d|\le\prod_{q\mid d}\nu_q\le2^{\omega(d)}.

Take D=ℓ1/2D=\ell^{1/2} and s0=log⁡D/log⁡ps_0=\log D/\log p. For sufficiently large pp, one has s0≥U/3s_0\ge U/3. The lower-bound part of [3 Theorem 9.1], with lower sieve function f2f_2, gives

CQ(N)≥ℓρQ(f2(s0)−o(1))−O(∑d<Dd∣Q4ω(d)∣Rd∣).C_Q(N)\ge\ell\rho_Q\bigl(f_2(s_0)-o(1)\bigr)-O\left(\sum_{\substack{d<D\\d\mid Q}}4^{\omega(d)}|R_d|\right).

The sieve error o(1)o(1) is O((log⁡log⁡D)2/(log⁡D)1/6)O((\log\log D)^2/(\log D)^{1/6}), with an absolute implied constant because the inverse-product condition is uniform. Since f2(s)f_2(s) increases to 11, choose the absolute UU so that f2(U/3)>0f_2(U/3)>0. The remainder is bounded by

∑d<D8ω(d)≪D(1+log⁡D)7≪ℓ1/2(1+log⁡ℓ)7.\sum_{d<D}8^{\omega(d)}\ll D(1+\log D)^7\ll\ell^{1/2}(1+\log\ell)^7.

Also,

ρQ≥∏5≤q<p(1−2q)≫1log⁡2p.\rho_Q\ge\prod_{5\le q<p}\left(1-\frac{2}{q}\right)\gg\frac{1}{\log^2 p}.

After division by ℓ/log⁡2p\ell/\log^2 p, the remainder is O(ℓ−1/2(1+log⁡ℓ)9)=o(1)O(\ell^{-1/2}(1+\log\ell)^9)=o(1) uniformly for N≥pUN\ge p^U as p→∞p\to\infty. Hence CQ(N)≫ℓ/log⁡2pC_Q(N)\gg\ell/\log^2 p, and (11) proves the claim. ∎

Corollary 5.7 (Uniformity of increasing prefixes). Under (54), U(N,Q,p)\mathrm{U}(N,Q,p) holds at every increasing prefix QpQp of PP for all sufficiently large NN, uniformly in PP. Moreover, for every fixed a>4a>4,

pVpsQ(N)2≪alog⁡2pp.(59)\frac{pV_p}{s_Q(N)^2}\ll_a\frac{\log^2 p}{p}. \tag*{(59)}

whenever N≥paN\ge p^a, pp is sufficiently large in terms of aa, and QQ is supported on primes in [5,p)[5,p).

Proof. Under (54), the surviving prime family gives sQ(N)≫θN/log⁡2Ns_Q(N)\gg_\theta N/\log^2 N for every Q∣PQ\mid P. Capacity yields

pVpsQ(N)2≤4MN,psQ(N)≪θlog⁡2N(1p+1N).\frac{pV_p}{s_Q(N)^2}\le\frac{4M_{N,p}}{s_Q(N)}\ll_\theta\log^2 N\left(\frac{1}{p}+\frac{1}{N}\right).

This proves the strict criterion for p≥Kθlog⁡2Np\ge K_\theta\log^2 N, with a sufficiently large fixed KθK_\theta. For the smaller steps, QQ is supported below pp and N≥p5N\ge p^5 uniformly when NN is large. Lemma 5.6 with a=5a=5 gives sQ(N)≫N/log⁡2ps_Q(N)\gg N/\log^2 p, so the normalized variance is O(log⁡2p/p)O(\log^2 p/p) and satisfies the strict criterion for all sufficiently large pp. The remaining bounded primes and their finitely many preceding supports are covered by (36).

For the second assertion, Lemma 5.6 and capacity give

pVpsQ(N)2≤4MN,psQ(N)≪alog⁡2pp.\frac{pV_p}{s_Q(N)^2}\le\frac{4M_{N,p}}{s_Q(N)}\ll_a\frac{\log^2 p}{p}.

Since the right side tends to zero, U\mathrm{U} holds for sufficiently large pp.

Remark 5.8 (Almost all NN). For every A>0A>0, there is a set of OA(X/log⁡AX)O_A(X/\log^A X) even integers in [26,X][26,X] outside which sQ(N)>0s_Q(N)>0 and U(N,Q,p)\mathrm{U}(N,Q,p) hold simultaneously for every prime 5≤p<N5\le p<\sqrt{N} and every QQ supported in [5,p)[5,p). To see this, apply the classical weighted abundance estimate in [1 proof of Theorem 4.3] on dyadic intervals. Proper prime powers and primes at most N\sqrt{N} contribute only O(Nlog⁡3N)=o(N)O(\sqrt{N}\log^3 N)=o(N) to the weighted count. Dividing the remaining weight by log⁡2N\log^2 N gives ≫N/log⁡2N\gg N/\log^2 N prime pairs with both summands exceeding N\sqrt{N}. These pairs survive every natural-cutoff support. Capacity handles p>N1/5p>N^{1/5}, and (59) handles the smaller unbounded primes. The finitely many remaining primes are covered by (36). The exceptional set is therefore common to all supports for the same NN.

Minimum prime covers and a residual problem

The preceding section bounds from below the noncolliding cost of a zero. We now reduce the minimum total covering problem by selecting one prime from each prime pair. This will give the lower bound for J(k)J(k) along powers of two.

EN={1≤h≤N/2:gcd⁡(h(N−h),6)=1},Ph={q≥5:q prime, q∣h(N−h)}.E_N=\{1\le h\le N/2:\gcd(h(N-h),6)=1\},\qquad\mathcal{P}_h=\{q\ge5:q\text{ prime},\ q\mid h(N-h)\}.

Every Ph\mathcal{P}_h is nonempty. Define

τ(N)=min⁡{ω(P):gP(N)=0}.(60)\tau(N)=\min\{\omega(P):g_P(N)=0\}. \tag*{(60)}

This minimum is finite: taking all primes in [5,N)[5,N) covers every Ph\mathcal{P}_h. Equivalently, τ(N)\tau(N) is the minimum size of a set of primes meeting every Ph\mathcal{P}_h. Primes at least NN can be omitted.

Let G(N)=gΠ(N)(N)G(N)=g_{\Pi(N)}(N) be the count in (6), and put

M0(N)=π(N)−2,(61)M_0(N)=\pi(\sqrt{N})-2, \tag*{(61)}
H(N)=1N−1 prime+#{5≤h≤N/2:h,N−h prime},(62)H(N)=\mathbf{1}_{N-1\ \mathrm{prime}}+\#\{5\le h\le N/2:h,N-h\ \mathrm{prime}\}, \tag*{(62)}
Hsm(N)=#{5≤q≤N:q,N−q prime}.(63)H_{\mathrm{sm}}(N)=\#\{5\le q\le\sqrt{N}:q,N-q\ \mathrm{prime}\}. \tag*{(63)}

Then H=G+HsmH=G+H_{\mathrm{sm}}. Distinct prime pairs use disjoint sets of primes, because each prime determines its complement. The endpoint contributes another disjoint singleton when N−1N-1 is prime.

Define the following set of primes:

FN={h:5≤h≤N/2, h,N−h prime}∪{N−1 if N−1 is prime}.\mathcal{F}_N=\{h:5\le h\le N/2,\ h,N-h\ \mathrm{prime}\}\cup\{N-1\ \mathrm{if}\ N-1\ \mathrm{is\ prime}\}.

Thus ∣FN∣=H(N)|\mathcal{F}_N|=H(N). The residual rows are

EN0={h∈EN:Ph∩FN=∅}.E_N^0=\{h\in E_N:\mathcal{P}_h\cap\mathcal{F}_N=\varnothing\}.

Let τ0(N)\tau_0(N) be their minimum covering number, with value zero if EN0E_N^0 is empty.

Theorem 6.1 (Canonical prime-pair reduction). A minimum cover can be chosen to contain FN\mathcal{F}_N, and

τ(N)=H(N)+τ0(N),0≤τ0(N)≤M0(N)−Hsm(N).(64)\tau(N)=H(N)+\tau_0(N),\qquad0\le\tau_0(N)\le M_0(N)-H_{\mathrm{sm}}(N). \tag*{(64)}

In particular,

G(N)+Hsm(N)≤τ(N)≤G(N)+M0(N).(65)G(N)+H_{\mathrm{sm}}(N)\le\tau(N)\le G(N)+M_0(N). \tag*{(65)}

Proof. Consider a prime pair 5≤λ≤ϱ5\le\lambda\le\varrho with λ+ϱ=N\lambda+\varrho=N. If λ<ϱ\lambda<\varrho, then ϱ>N/2\varrho>N/2. The prime ϱ\varrho occurs in only one row of the covering problem, namely the row h=λh=\lambda: its only positive multiple below NN is ϱ\varrho. A cover containing ϱ\varrho but not λ\lambda can therefore replace ϱ\varrho by λ\lambda without increasing its size or uncovering any row. If λ=ϱ\lambda=\varrho, the row forces the prime λ\lambda. The endpoint, when present, forces N−1N-1.

The replacements for distinct prime pairs do not interfere. Hence a minimum cover can be chosen to contain FN\mathcal{F}_N. Its additional primes must cover exactly EN0E_N^0, proving the identity. Every residual row has a prime factor at most N\sqrt{N}, since otherwise it would be a surviving prime pair or the endpoint already covered by FN\mathcal{F}_N. None of the Hsm(N)H_{\mathrm{sm}}(N) small primes in FN\mathcal{F}_N occurs in a residual row. The other M0(N)−Hsm(N)M_0(N)-H_{\mathrm{sm}}(N) small primes cover all residual rows, proving the upper bound.

The set FN\mathcal{F}_N specifies a canonical choice within some minimum cover. The theorem does not require every minimum cover to contain it.

In particular, τ(N)>M0(N)\tau(N)>M_0(N) implies G(N)>0G(N)>0. The converse has not been established here. The unconditional relation 0≤τ(N)−G(N)≤M0(N)0\le\tau(N)-G(N)\le M_0(N) explains why τ\tau is close to the large-prime representation count whenever that count is large. It does not imply that G+M0−τG+M_0-\tau is bounded, or that all minimum covers have the same shape.

Fractional covers and disjoint products

Let ν(N)\nu(N) be the maximum cardinality of a family of rows whose products h(N−h)h(N-h) are pairwise coprime. The fractional covering number is

τ∗(N)=min⁡{∑qxq:xq≥0, ∑q∈Phxq≥1 for every h}(66)\tau^*(N)=\min\left\{\sum_q x_q:x_q\geq0,\ \sum_{q\in\mathcal{P}_h}x_q\geq1\text{ for every }h\right\} \tag*{(66)}
=max⁡{∑hyh:yh≥0, ∑h:q∈Phyh≤1 for every q}.(67)=\max\left\{\sum_h y_h:y_h\geq0,\ \sum_{h:q\in\mathcal{P}_h}y_h\leq1\text{ for every }q\right\}. \tag*{(67)}

The equality is finite-dimensional linear programming duality. Integer covers and disjoint products satisfy

H(N)≤ν(N)≤τ∗(N)≤τ(N).(68)H(N)\leq\nu(N)\leq\tau^*(N)\leq\tau(N). \tag*{(68)}

The duality does not assert equality of the two integer optima.

Proposition 6.2. With τ0∗\tau^*_0 and ν0\nu_0 defined on the residual rows,

τ∗(N)=H(N)+τ0∗(N),ν(N)=H(N)+ν0(N).(69)\tau^*(N)=H(N)+\tau^*_0(N),\qquad\nu(N)=H(N)+\nu_0(N). \tag*{(69)}

Moreover, for every 0<θ<1/40<\theta<1/4,

τ∗(N)≫θNθ(70)\tau^*(N)\gg_\theta N^\theta \tag*{(70)}

for all sufficiently large even NN.

Proof. For a prime pair λ<ϱ\lambda<\varrho, transfer the fractional weight at ϱ\varrho to λ\lambda and truncate the resulting weight at 1. The row constraint ensures that this leaves weight 1 at λ\lambda. No other row uses ϱ\varrho, so feasibility is preserved and cost cannot increase. The diagonal and endpoint force unit weights as before. This proves the first identity in (6.10).

A residual family of disjoint products can be joined to all the prime-pair rows and the endpoint. The primes in FN\mathcal{F}_N do not occur in residual rows, and their larger prime partners occur only in their own rows. Conversely, any family of disjoint products has at most H(N)H(N) rows using a prime from FN\mathcal{F}_N, since different rows cannot use the same such prime. Its other rows form a residual disjoint family. This proves the second identity.

For (6.11), give each row of Tθ(N)\mathcal{T}_\theta(N) the weight log⁡2N/(CθS(N)N1−θ)\log^2 N/(C_\theta\mathcal{S}(N)N^{1-\theta}) and every other row weight zero. Lemma 5.3 makes these weights feasible for (6.8). Summing them and using Lemma 5.1 proves the claim.

Example 6.3 (A fractional gap at N=352N=352). Here EN={5+6j:0≤j≤28}E_N=\{5+6j:0\leq j\leq28\}, FN={5,41,59,71,83,89,101,113,173}\mathcal{F}_N=\{5,41,59,71,83,89,101,113,173\}, and H(N)=9H(N)=9. Removing the rows in FN\mathcal{F}_N and those with 5∣h(N−h)5\mid h(N-h) leaves the following nine rows. Their displayed prime factors avoid FN\mathcal{F}_N, so the list is exactly EN0E^0_N:

hhPh\mathcal{P}_hhhPh\mathcal{P}_h
11{11,31}\{11,31\}119{7,17,233}\{7,17,233\}
23{7,23,47}\{7,23,47\}131{13,17,131}\{13,17,131\}
29{17,19,29}\{17,19,29\}143{11,13,19}\{11,13,19\}
53{13,23,53}\{13,23,53\}149{7,29,149}\{7,29,149\}
161{7,23,191}\{7,23,191\}

Table 6.3.

The primes {7,11,17,23}\{7, 11, 17, 23\} cover these rows. A fractional cover gives weight 1 to 11 and weight 1/21/2 to each of 7, 13, 17, 23, 29, with total cost 7/27/2. A dual solution of the same value gives weight 1 to row 11 and weight 1/21/2 to each of rows 23, 29, 53, 131, 149. The table verifies every constraint exactly. Thus τ0∗=7/2\tau_0^* = 7/2 and τ0=4\tau_0 = 4. Rows 11, 131, 161 have pairwise disjoint prime factors, so ν0=3\nu_0 = 3, using ν0≤⌊τ0∗⌋\nu_0 \le\lfloor\tau_0^* \rfloor. Therefore

ν(352)=12,τ∗(352)=252,τ(352)=13.\nu(352) = 12,\qquad\tau^*(352) = \frac{25}{2},\qquad\tau(352) = 13.

The displayed factorizations and rational weights prove a gap between integer covering and packing. No external data or numerical optimization is needed.

A lower bound for the extremal size

The cover reduction also constructs vanishing supports and therefore complements the upper bound for J(k)J(k).

Proposition 6.4. As k→∞k \to\infty, one has J(k)≫k(log⁡k)2J(k) \gg k(\log k)^2.

Proof. For N=2jN = 2^j, the local factor is S(N)=1\mathfrak{S}(N) = 1. Apply Lemma 5.2 with q=1q = 1 and a fixed θ\theta, for example θ=1/8\theta= 1/8. Every prime pair whose smaller summand exceeds w=N1/100w = N^{1/100} is counted by that upper sieve, apart from the possible diagonal. The pairs with smaller summand at most ww, the diagonal, and the endpoint contribute O(w+1)O(w + 1). Hence

H(N)≪Nlog⁡2N.H(N) \ll\frac{N}{\log^2 N}.

By Theorem 6.1,

τ(N)≤H(N)+M0(N)≪Nlog⁡2N,\tau(N) \le H(N) + M_0(N) \ll\frac{N}{\log^2 N},

since M0(N)≤N=o(N/log⁡2N)M_0(N) \le\sqrt{N} = o(N/\log^2 N). Choose a sufficiently small fixed γ>0\gamma> 0 and, for each large kk, take the largest power of two N≤γklog⁡2kN \le\gamma k\log^2 k. Then N>γklog⁡2k/2N > \gamma k\log^2 k/2, log⁡N∼log⁡k\log N \sim\log k, and τ(N)≤k\tau(N) \le k. A minimum cover yields gP(N)=0g_P(N) = 0. Every prime factor of PP is noncolliding because NN is a power of two. Thus N≤J(k)N \le J(k), proving the claim. □\square

Together with Corollary 5.5, this proves Theorem 1.1. The lower bound constructs zeros for unrestricted supports. These covers need not satisfy pmax⁡2<Np_{\max}^2 < N.

Remark 6.5 (Forced noncolliding cost). Every vanishing support satisfies the sharper inequality

k(N,P)≥H(N)−1N/2 prime≥H(N)−1.(71)k(N,P) \ge H(N) - \mathbf{1}_{N/2\ \mathrm{prime}} \ge H(N) - 1. \tag*{(71)}

Indeed, a prime pair λ<ϱ\lambda< \varrho with λ+ϱ=N\lambda+\varrho= N and λ≥5\lambda\ge5 forces PP to contain λ\lambda or ϱ\varrho. Neither divides nn: one has ϱ>n\varrho> n, and λ∣n\lambda\mid n would imply λ∣ϱ\lambda\mid\varrho. Distinct pairs use distinct primes. The endpoint, when present, forces the noncolliding prime N−1N - 1. Only the diagonal prime nn, when present, collides. Consequently,

J(k)≤sup⁡({0}∪{N≥26:N even, H(N)≤k+1}).(72)J(k) \le\sup\left(\{0\} \cup\{N \ge26 : N\ \text{even},\ H(N) \le k + 1\}\right). \tag*{(72)}

If H(N)≫N/log⁡2NH(N) \gg N/\log^2 N for all sufficiently large even NN, as predicted by the Hardy–Littlewood conjecture [6], then (6.12) gives N≪k(log⁡k)2N \ll k(\log k)^2 for every zero with at most kk noncolliding factors, for all sufficiently large kk. Proposition 6.4 would then yield

J(k)≍k(log⁡k)2.J(k) \asymp k(\log k)^2.

Thus the conjectural exponent of kk is 1. Proposition 7.1 uses exponent 2 with a leading constant below 1, while Theorem 1.1 gives 4+ε4+\varepsilon.

The natural cutoff and the remaining gap

The estimates for J(k)J(k) quantify the size of a vanishing instance, while the cover reduction identifies its forced prime-pair component. Neither bound reaches the scale needed to settle positivity at the complete natural cutoff.

Proposition 7.1 (A sufficient quadratic bound). If there is a fixed 0<η<10<\eta<1 such that

J(k)≤(1−η)(klog⁡k)2(73)J(k) \le(1-\eta)(k\log k)^2 \tag*{(73)}

for all sufficiently large kk, then gΠ(N)(N)>0g_{\Pi(N)}(N)>0 for every sufficiently large even NN.

Proof. Put kN=ω(Π(N)/gcd⁡(Π(N),N/2))k_N=\omega\left(\Pi(N)/\gcd(\Pi(N),N/2)\right). The prime number theorem and ω(N)≪log⁡N\omega(N)\ll\log N give

kN∼2Nlog⁡N,(kNlog⁡kN)2∼N.k_N \sim\frac{2\sqrt{N}}{\log N}, \qquad(k_N\log k_N)^2\sim N.

If gΠ(N)(N)=0g_{\Pi(N)}(N)=0, the definition of JJ gives N≤J(kN)N\le J(k_N), contradicting (73) for large NN. □\square

The proposition is a sufficient criterion, not a converse or an established estimate for JJ. The exponent 4+ε4+\varepsilon in Theorem 1.1 is separated from this quadratic target. The analogy with the ordinary Jacobsthal bound does not transfer its proof to the two classes tied to NN.

The weighted criterion locates another precise gap. For the complete natural support, Mertens’ theorem gives

Wθ(N,Π(N))=∑Nθ<q≤Nq∤N1q=log⁡12θ+oθ(1).(74)W_\theta(N,\Pi(N))=\sum_{\substack{N^\theta<q\le\sqrt{N}\\q\nmid N}}\frac{1}{q}=\log\frac{1}{2\theta}+o_\theta(1). \tag*{(74)}

The factors dividing NN contribute only Oθ(N−θ)O_\theta(N^{-\theta}) in this range. Thus the upper bound for the total loss now has the same order as the supplied family, S(N)N/log⁡2N\mathcal{S}(N)N/\log^2 N. The constants obtained here do not prove cθ>Cθlog⁡(1/(2θ))c_\theta>C_\theta\log(1/(2\theta)). Finite measurements of the actual losses cannot replace a uniform strict inequality. The gap is consistent with the parity obstruction in classical sieve theory. Our estimates give no lower bounds for intersections of losses. Type II, or bilinear, information is one possible route beyond these bounds, as illustrated by the prime-producing sieve framework in [4]. No such estimates are proved here.

For completeness, the local question remains:

Problem 7.2. Does every even N≥26N\ge26, every prime 5≤p<N5\le p<\sqrt{N}, and every product QQ of primes in [5,p)[5,p) satisfy sQ(N)>0s_Q(N)>0 and U(N,Q,p)U(N,Q,p)?

An affirmative answer would exclude every first zero at the natural cutoff. It would also imply, by (6), that every even N≥26N\ge26 with N−1N-1 composite is a sum of two primes exceeding N\sqrt{N}. Because a large central occupancy can preserve a representation without satisfying UU, this local question contains more information than positivity alone.

Order the factors of a hypothetical vanishing support increasingly and let P=QpP = Q_p be its first vanishing prefix. The preceding set is nonempty. At the natural cutoff and for N>2780460N > 2780460, Theorems 4.6 and 4.8 and Proposition 4.3 require

ω(P)≥12,k(N,P)≥8,p2<N<6Q,ℓ<Q.(75)\omega(P) \ge12,\qquad k(N,P) \ge8,\qquad p^2 < N < 6Q,\qquad\ell< Q. \tag*{(75)}

The explicit results also require the profile inequalities (4.17)–(4.18). For each fixed θ<1/4\theta< 1/4, sufficiently large instances must have κθ(N,P)≫θNθ\kappa_\theta(N,P) \gg_\theta N^\theta. Corollary 5.7 excludes N≥p4+εN \ge p^{4+\varepsilon} for sufficiently large pp, with any fixed ε>0\varepsilon> 0. The first-zero configuration is therefore constrained both by its residue concentration and by the cost of the support that produces it. Excluding every remaining instance is the open step.

References

  1. [1]Gautami Bhowmik and Lasse Grimmelt. The exceptional set of the Goldbach problem. Analysis Mathematica, 52:767–808, 2026. doi: 10.1007/s10476-026-00191-z.DOI
  2. [2]Matteo Bordignon, Daniel R. Johnston, and Valeriia Starichkova. An explicit version of Chen’s theorem and the linear sieve, 2025. URL https://arxiv.org/abs/2207.09452. arXiv:2207.09452v6.
  3. [3]Harold G. Diamond and Heini Halberstam. A Higher-Dimensional Sieve Method, volume 177 of Cambridge Tracts in Mathematics. Cambridge University Press, Cambridge, 2008. With procedures for computing sieve functions by William F. Galway.
  4. [4]Kevin Ford and James Maynard. On the theory of prime producing sieves, 2024. URL https://arxiv.org/abs/2407.14368. Preprint, arXiv:2407.14368.
  5. [5]Henryk Iwaniec. On the problem of Jacobsthal. Demonstratio Mathematica, 11(1):225–232, 1978. doi: 10.1515/dema-1978-0121.DOI
  6. [6]Angel V. Kumchev and Dimitar I. Tolev. An invitation to additive prime number theory. Serdica Mathematical Journal, 31:1–74, 2005. URL https://tigerweb.towson.edu/akumchev/a17.pdf.
  7. [7]Tien Tuan Khiem Nguyen. Finite-window nonconvering on primordial wheels: Higher-order CRT bounds and shift correlations, 2026. URL https://www.preprints.org/manuscript/202608.1299. Preprint, version 1.DOI
  8. [8]Andrés M. Salazar. Exact formulas for restricted coprime representations of even integers. Communications on Number Theory and Combinatorial Theory, 7, 2026. doi: 10.70013/k94n5fzj. Article 2.DOI
  9. [9]Andrés M. Salazar. Restricted coprime representations modulo 6P: Exact CRT formulas and positivity thresholds, 2026. URL https://arxiv.org/abs/2609.22305v2. Preprint, arXiv:2609.22305v2.
  10. [10]Mario Ziller and John F. Morack. Divisibility in paired progressions, Goldbach’s conjecture, and the infinitude of prime pairs, 2017. URL https://arxiv.org/abs/1706.00317. Preprint, arXiv:1706.00317.

Paper details

vlnc:W_kktm7kr8
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