Preprint
Article

This version is not peer-reviewed.

The Restricted Chen Sieve: Explicit Main-Term Bounds, Kernel Growth, and Factorable Supports

Submitted:

03 November 2026

Posted:

06 November 2026

You are already at the latest version

Abstract
We investigate a restricted Chen sieve in which Rosser's lower weights are used at level \(D = \lfloor N^{\theta} \rfloor\), with \(3/4 < \theta < 1\), and switching is confined to a prescribed interval of large prime divisors. We obtain an unconditional lower bound for the main term with an explicit coefficient, determine its exact positivity threshold, and evaluate the coefficient as \(\frac{50}{49}\log\frac{2544}{2405}\) at the principal parameters. The threshold for this coefficient is strictly greater than \(2/3\). For admissible parameters with \(c \le 1/3\), we construct exact negative values of the sieve kernel on its arithmetic summation set and prove that its minimum is at most \(-\exp\left(c_{\theta}\sqrt{\log N}\right)\) for infinitely many even \(N\). We also characterize the union of the supports of all \(k\)-factorable sequences of a given level by a generalized Schinzel–Szekeres function. For squarefree moduli, the corresponding strict sublevel set identifies the intersection of the upper and lower Rosser supports. This criterion excludes every switched modulus from the support of triply factorable weights at the sieve level. Further results establish the optimal constant \(\log 3\) in the dimension-one condition for the Goldbach density and an unconditional mean-value estimate for the full remainder, yielding the weighted lower bound for almost all even integers. A companion construction using parity-weighted cofactor minorants admits an exact reweighting envelope, an explicit gain function, and unconditional mean-value estimates.
Keywords: 
;  ;  ;  ;  ;  ;  ;  

1. Introduction

1.1. Background and Objectives

Let N be a sufficiently large even integer. Chen’s theorem [Che73] gives a representation N = p + n, where p is prime and Ω(n) ≤ 2. Its proof combines a lower-bound linear sieve for the sequence {N − p} with weighted upper-bound sieves for subsequences divisible by a prime, together with Chen’s switching principle; see [HR74, Ch. 11] and [FI10]. The Bombieri–Vinogradov theorem supplies the underlying distribution estimate at exponents below 1/2. The distinction between prime and semiprime values is governed by the parity phenomenon [Sel49], which makes the structure of the weights central to the analysis.
The interaction between the sieve level, the sifting limit, and the support of the coefficients becomes particularly explicit when switching is restricted to large prime divisors. We study this interaction for a family in which the lower sieve has level D = ⌊Nθ⌋, with 3/4 < θ < 1, and the switched primes lie in A α , B . Our analysis has three principal components: an explicit lower bound for the main term, quantitative growth of negative kernel values, and an exact characterization of the moduli compatible with factorable decompositions. Mean-value estimates connect these structural results to the weighted counting function. Section 8 develops a related construction based on parity-weighted cofactor minorants.

1.2. The Construction

A parameter vector p = (c,a,β,θ) is admissible if, with α = (a − 1)/a,
0 < c < 1 / 2 ,     1 / 3 < β < α < 1 / 2 ,     max 2 β , 3 / 4 < θ < 1 .
The principal vector is p 0 = 1 4 , 19 10 , 9 20 , 49 50 , for which α = 9 19 . For an even integer N put A = ⌈cN⌉, B = ⌊(1 − c)N⌋, z = Nβ and D = ⌊Nθ⌋. The quantity of interest is
J N = # p   prime :   N − B ≤ p ≤ N − A ,   P - N − p > z ,   b a N − p = 1 ,
where ba(n) = 1 if n is prime or n = rq with primes r < q and r ≤ qa−1, and ba(n) = 0 otherwise. Thus JN > 0 means that N = p + n where n is a prime or a product of two primes both exceeding Nβ; at p0 the smaller factor then lies in (N0.45,N0.474]. The sieve kernel (Definition 3.1) is
K N n = ∑ d ∣ n , P N z λ D - d   − ∑ r ∈ R N r ∣ n ∑ e ∣ n / r ,   P N w r λ D / r + e ,
where λ± are Rosser’s weights, PN(y) is the product of the primes p ≤ y with p∤N, R N is the set of primes r ∈ A α , B   with r∤N, and w r = B / r . Summing KN over SN = {n ∈ [A,B]:N − n prime} produces a main term MN and a remainder EN with MN + EN ≤ JN (Corollary 3.8). We measure all quantities against
R N = S N   N log 2 N ,     S N = 2 C 2 ∏ p ∣ N p > 2 p − 1 p − 2 ,     C 2 = ∏ p > 2 1 − 1 p − 1 2 .

1.3. Main Results

The first theorem gives an explicit unconditional lower bound for the main term and identifies the positivity region of its coefficient.
Theorem A (Explicit main-term lower bound). For every admissible p, as N → ∞ through even integers,
M N   ≥   C p − o 1 R N ,     C p = 2 1 − 2 c θ   log θ / β − 1   α 2 θ − 1 θ − α .
Moreover C p > 0 if and only if θ > θ*(α,β), where θ * α , β = u + u 2 − 16 α 2 β   / 4 α and u = α + β + 2αβ. At p0 one has θ * = 57 + 3 41 / 80 = 0.9526 … and C p 0 = 50 49 log 2544 2405 = 0.05733 … > 3475 62328 .
The proof (Section 4) combines Iwaniec’s bounds for the main terms of Rosser’s sieve [Iw80a] with Mertens’ theorem and the uniform dimension condition of Theorem E below.
Theorem B (Reduction to the high-modulus remainder). Let QN=⌊N1/2(logN)−8⌋ and split E N = E N low + E N high according to whether the modulus is at most QN or exceeds it (see (5.1)). For every admissible p we have E N low = O N log N − 3 , and consequently
J N   ≥   C p − o 1 R N + E N high     N → ∞ ,   N   even .
In particular, if E N high ≥ − κ R N for all large even N with some κ < C p , then J N ≥ C p − κ − o 1 R N ; and under the Elliott–Halberstam conjecture, J N ≥ C p − o 1 R N .
The next two theorems describe the kernel and the admissible supports of factorable coefficient sequences.
Theorem C (The kernel is unbounded below). Let p be admissible with c ≤ 1/3.
(i) For every integer k ≥ 3 with θ k ∉ 2 Z there are infinitely many even N and n ∈ SN with ω(n) = Ω(n) = k such that K N n = − k − 1 2 ⌊ θ k / 2 ⌋ − 1 .
(ii) There is cθ > 0 such that min n ∈ S N K N n ≤ − exp c θ log N for infinitely many even N.
(iii) For all large N and all n ∈ SN one has |KN(n)| ≤ 3τ(n).
For k = 3, the value in part (i) is −2. The exact binomial formula and part (ii) quantify the growth of negative kernel values on SN, even though the coefficients satisfy |WN(m)| ≤ 1 (Proposition 3.3).
Theorem D (Universal support of factorable sequences). Let k ≥ 1 be an integer and Q ≥ 1 be real. Call a : N → C  k−factorable of level Q if for all reals Q1,…,Qk ≥ 1 with Q1 … Qk = Q there are a1,…,ak, with ai supported on [1,Qi], such that a = a1*…*ak. Then every k−factorable sequence of level Q vanishes outside
S k Q = m ≥ 1 :   p k − 1 m ≥ p ≤ Q     for   every   prime   p ∣ m ,     m ≥ p = ∏ q e   ∥   m q ≥ p q e ,
and for every m ∈ S k Q the point mass δm is k−factorable of level Q with factors taking values in {0,1}.
For m > 1, the maximum of pk−1m≥p over primes p∣m equals the generalized Schinzel–Szekeres function Fk−1(m) = maxd∣m, d>1dP−(d)k−1; set Fk−1(1) = 1. Thus S k Q = m : F k − 1 m ≤ Q . The functions Fβ and their counting theory are studied in [SW26]. When k = 2, this set consists precisely of the (Q/m)−densely divisible integers [Ten86, Sai97] and coincides with the well-factorable support considered in [Lic23, §6]. Theorem D supplies an exact universal-support criterion for every integer k ≥ 1, together with a factorable point mass at each admissible modulus.
For k = 3 and squarefree m, the inequalities defining S 3 D are the Rosser inequalities imposed at every index, with non-strict endpoints. The intersection of the upper and lower Rosser supports is the strict sublevel set {m:μ2(m) = 1, F2(m) < D} (Proposition 7.5). This relation gives an explicit support obstruction to triple factorability. In conjunction with [SW26, Theorem 5], it also yields # S 3 Q ≍ Q log Q − 2.462 … as Q → ∞. For the restricted Chen sieve, Proposition 7.8 shows that every switched modulus lies outside S 3 D and that the high-modulus coefficient sequence fails to be 2−factorable at level D.
Theorem E (Optimal constant in the dimension condition). For every even N and all real 2 ≤ u < v,
∏ u ≤ p < v p ∤ N 1 − 1 p − 1 − 1   ≤   log v log u 1 + log 3 log u ,
and the constant log3 cannot be replaced by any smaller number.
Theorem E determines the optimal uniform dimension constant, including the contribution of the smallest admissible prime. Its proof combines the estimates of Rosser and Schoenfeld [RS62] with a finite calculation over the primes below 286. The resulting bound supplies the dimension hypothesis in Theorem A uniformly in the even integer N.
The full remainder also admits an unconditional estimate on average over the additive parameter.
Theorem F (Almost all N). Let p be admissible and A > 0. Then ∑ X < N ≤ 2 X E N ≪ A , p X 2 log X − A . Consequently J N ≥ C p − o 1 R N for all even N ∈ (X,2X] outside a set of OA,p(X(logX)−A) elements.
Theorem F controls the signed sieve remainder attached to the coefficients WN, thereby transferring the explicit main-term lower bound to JN outside a logarithmically sparse exceptional set. The circle-method background for almost-all representations is discussed in Remark 5.10. Separately, Proposition 4.6 proves that positivity of the coefficient C requires θ > 2/3. This is a threshold for the derived lower bound; the admissibility conditions additionally require θ > max(2β,3/4).
The companion construction in Section 8 uses the Liouville function and Rosser’s lower sieve to define a pointwise minorant for a prime cofactor over a rectangle of prime pairs. Theorems 8.2–8.7 establish its minorant property, exact decompositions, an optimal reweighting envelope, and a correction of the lower Rosser kernel that preserves the sieve inequality. Proposition 8.9 gives an explicit main-term lower bound for the envelope. Propositions 8.13 and 8.14 determine the positivity region of its gain function G a , b η and provide explicit parameter bounds. The unconditional mean-value estimates of Proposition 8.11 yield the almost-all conclusion in Theorem 8.12.

1.4. Contributions and Relation to Earlier Work

The results distinguish the arithmetic evaluation of a sieve from the geometry of its coefficients. For the restricted family, the explicit coefficient C , its algebraic positivity threshold, and the exact loss identities describe the quantitative effect of restricting the switching interval. The construction in Theorem C further identifies negative values of unbounded magnitude on the actual set SN, rather than on an auxiliary set of integers. These statements give a quantitative account of the difference between bounded divisor coefficients and their associated sieve kernel.
The support theorem provides a second contribution: it characterizes, for each k and each level, every modulus at which a k−factorable sequence can be nonzero. Its constructive proof realizes each such modulus by a point mass with factors bounded by 1. The functions defining these moduli are those of the dense-divisibility theory [Ten86, Sai97, SW26]; the factorability criterion connects that theory to the support of Rosser weights and to the switched coefficients of the present sieve. In particular, the exclusion in Proposition 7.8 is a support statement, independent of the size of an error term in a distribution theorem.
The analytic inputs are the linear-sieve estimates [Iw80a], the Bombieri–Vinogradov theorem [Bom65, Vin65], and the Barban–Davenport–Halberstam theorem [DH66, Dav00]. Distribution estimates with well-factorable and triply well-factorable weights [BFI86, May25, Lic23] give the broader context; their level and residue ranges are compared with this construction in Remark 4.7. Within the present family, Theorem B is a one-sided transfer criterion from the high-modulus remainder to the weighted count, and its Elliott–Halberstam specialization is stated in Corollary 5.4. Theorem F instead follows unconditionally by averaging over N. The parity-weighted construction adds an exact optimization over the reweighting function and an explicit positivity criterion for the resulting gain.

1.5. Organisation and Notation

Section 2 recalls Rosser’s weights and the linear sieve, establishes the Goldbach density estimates, and proves Theorem E. Section 3 defines the restricted Chen sieve and derives its exact identities. Section 4 proves Theorem A and the positivity threshold, while Section 5 proves Theorems B and F. Section 6 and Section 7 establish Theorems C and D. Section 8 develops the parity-weighted cofactor minorants, and Section 9 presents further questions. Appendix A supplies the explicit prime-sum estimates used in the parameter calculations.
The letters p , q , r , l denote primes unless otherwise specified; the variables u,v,w are used as indicated in each statement. We write P−(n) and P+(n) for the least and largest prime factors of n, with P−(1) = ∞ and P+(1) = 1. The functions ω(n) and Ω(n) count prime factors without and with multiplicity, respectively. We use τ for the divisor function, μ for the Möbius function, φ for Euler’s function, λL(n) = (−1)Ω(n) for the Liouville function, and 1 P for the prime indicator. Put li x = ∫ 2 x d t / log t and π(x;m,a) = #{p ≤ x:p≡a (mod m)}. For y > 0, let P y = ∏ p ≤ y p and P N y = ∏ p ≤ y ,   p ∤ N p . Unless otherwise specified, o(1) is taken as N → ∞ through even integers with the parameter vector fixed. Uniformity in additional variables is stated where it is used, and sufficiently large N means N ≥ N0(p).

2. Preliminaries

2.1. Rosser’s Weights

Definition 2.1. Let D > 1 be real. For a squarefree integer d = p1 … pk with p1 > p2>…>pk and 1 ≤ j ≤ k put T j d = p 1 … p j − 1 p j 3 . Set λ D ± 1 = 1 and λ D ± d = 0 if d is not squarefree. For squarefree d > 1 let λ D + d = μ d if Tj(d) < D for every odd j ≤ k, and λ D − d = μ d if Tj(d) < D for every even j ≤ k; in all other cases λ D ± d = 0 . For a squarefree integer P and n ≥ 1 we write F D ± n ; P = ∑ d ∣ n , P λ D ± d .
These are Rosser’s weights for the linear sieve [Iw80a]; see also [KT05, (4.17)–(4.18)] and [FI10]. Since Tj(d) involves only the j largest prime factors of d, the defining conditions depend only on the set of prime factors and not on any enumeration of them. If λ D + d ≠ 0 then d = 1 or d < D, and if λ D − d ≠ 0 then d = 1, d is prime, or d < D: indeed Tk(d) ≥ d, and for k ≥ 2 we have T k − 1 d = p 1 … p k − 1   p k − 1 2 > d .
The first-failure identities below express Rosser’s sifting functions as signed counts of minimal violations of the defining inequalities. This form will be used to evaluate the kernel in Section 6.
Lemma 2.2 (First-failure identities). Let D > 1 and let P be squarefree. Let T − (resp. T + ) be the set of divisors t > 1 of P such that k = ω(t) is even (resp. odd), Tj(t) < D for every j < k with j≡k (mod 2), and Tk(t) ≥ D. For n ≥ 1 write nP = (n,P). Then
F D − n ; P = 1 n P = 1 − # t ∈ T − :   t ∣ n P ,   P − t = P − n P , F D + n ; P = 1 n P = 1 + # t ∈ T + :   t ∣ n P ,   P − t = P − n P .
Proof. We treat F−; the argument for F+ is the same with the parities interchanged and μ(t) = −1 in place of μ(t) = 1. Let D − be the set of squarefree d with λ D − d = μ d . Then
F D − n ; P = ∑ d ∣ n P μ d − ∑ d ∣ n P d ∉ D − μ d = 1 n P = 1 − ∑ d ∣ n P d ∉ D − μ d .
Let d∣nP with d ∉ D − , write d = p1 … pk with p1>…>pk, and let m be the least even index with Tm(d) ≥ D. Put t = p1 … pm and e = d/t. Since Tj(d) = Tj(t) for j ≤ m, we have t ∈ T − , and every prime factor of e is smaller than P−(t) = pm. Conversely, if t ∈ T − divides nP and e∣nP/t has all its prime factors below P−(t), then d = te ∉ D − , the first failure of d occurs at m = ω(t), and the m largest prime factors of d form t. This correspondence is bijective, so
∑ d ∣ n P d ∉ D - μ d = ∑ t ∈ T - t ∣ n P μ t ∑ e ∣ n P / t P + e < P - t μ e .
Here μ(t) = 1, and the inner sum equals 1 if nP/t has no prime factor below P−(t), that is, if P−(t) = P−(nP), and 0 otherwise. ◻
Corollary 2.3 (Rosser’s inequality). For all n ≥ 1 we have F D − n ; P ≤ 1 n , P = 1 ≤ F D + n ; P and F D ± n ; P ≤ τ n , P .
Lemma 2.4 (Products of primes of comparable size). Let D > 1, let P be squarefree, and let n ≥ 1 satisfy (n,P) = q1 … qk with q1<…<qk and k ≥ 2. Let m be an even integer with 2 ≤ m ≤ k. If q k m < D ≤ q 1 m + 2 , then F D − n ; P = − k − 1 m − 1 .
Proof. For t∣(n,P) and j ≤ ω(t), Tj(t) is a product of j + 2 of the primes qi, counted with multiplicity, so q 1 j + 2 ≤ T j t ≤ q k j + 2 . For even j ≤ m − 2 this gives T j t ≤ q k m < D , and for j = m it gives T m t ≥ q 1 m + 2 ≥ D . Hence the elements of T − dividing (n,P) are exactly the divisors with m prime factors, and those with least prime factor q1 are obtained by choosing the remaining m − 1 factors among q2,…,qk. Lemma 2.2 gives the claim, since (n,P) ≠ 1. ◻
For instance, if (n,P) = q1q2q3 with q 3 2 < D ≤ q 1 4 , then F D − n ; P = − 2 .

2.2. The Linear Sieve

We use Iwaniec’s bounds for the main terms of Rosser’s sieve in the following form.
Theorem 2.5 (Iwaniec). Let P be a set of primes and g a multiplicative function supported on squarefree integers composed of primes of P , with 0 < g(p) < 1 for p ∈ P , such that for some K ≥ 0
∏ w 1 ≤ p < w 2 p ∈ P ( 1 − g p ) − 1 ≤ log w 2 log w 1 1 + K log w 1     2 ≤ w 1 < w 2 .
Let z ≥ 2, D ≥ 2, P z = ∏ p ∈ P ,   p < z p , V z = ∏ p ∣ P z 1 − g p and s = logD/logz. Then
∑ d ∣ P z λ D + d g d ≤ V z F s + O K e − s log D − 1 / 3     s ≥ 1 , ∑ d ∣ P z λ D - d g d ≥ V z f s − O K e − s log D − 1 / 3     s ≥ 2 ,
where F and f are the functions of the linear sieve; in particular sF(s) = 2eγ for 0 < s ≤ 3 and sf(s) = 2eγlog(s − 1) for 2 ≤ s ≤ 4.
This formulation follows from Iwaniec’s estimates [Iw80a]; see [KT05, (4.15)–(4.20)]. The dimension condition there is stated with the closed range w1 ≤ p ≤ w2. That formulation follows from (2.1) by approaching the upper endpoint from above, since the prime product is constant between consecutive primes.

2.3. The Goldbach Density

For even N let gN(p) = 1/(p − 1) if p∤N and gN(p) = 0 if p∣N, extended multiplicatively to squarefree integers; thus gN(d) = 1/φ(d) for squarefree d coprime to N. Put
V N y = ∏ p ≤ y 1 − g N p = ∏ p ≤ y p ∤ N p − 2 p − 1 .
Lemma 2.6 (Euler factorisation of the density). For even N ≥ 2 and y ≥ 2,
V N y = ∏ 2 < p ≤ y 1 − 1 p ⋅ ∏ 2 < p ≤ y 1 − 1 p − 1 2 ⋅ S N 2 C 2 ⋅ T N y ,     T N y = ∏ p ∣ N p > y p − 2 p − 1 ,
and 1 − log N / log y y − 1 ≤ T N y ≤ 1 .
Proof. For p > 2 one has 1 − 1 p 1 − 1 p − 1 2 = p − 2 p − 1 , which is the factor of VN(y) at a prime p∤N. The primes 2 < p ≤ y dividing N are absent from VN(y) but present in the two complete products; the compensating factor ∏ p ∣ N ,   2 < p ≤ y p − 1 p − 2 equals S N 2 C 2 T N y . Since N is even, the prime 2 is excluded from VN(y). If N has j prime factors exceeding y, then yj < N, so j ≤ logN/logy; each factor of TN(y) is at least 1 − 1/(y − 1), and the lower bound follows from ∏ 1 − x i ≥ 1 − ∑ x i . ◻
Corollary 2.7. Fix δ∈(0,1]. Uniformly for even N and Nδ ≤ y ≤ N, V N y = 1 + o 1   e − γ S N / log y .
Proof. By Mertens’ theorem [Mer74], ∏ 2 < p ≤ y 1 − 1 / p = 2 ∏ p ≤ y 1 − 1 / p = 1 + o 1   2 e − γ / log y . Further ∏ 2 < p ≤ y 1 − p − 1 − 2 = C 2 1 + O 1 / y and TN(y) = 1 + O(δ−1y−1) by Lemma 2.6. ◻
We now prove Theorem E. For real 2 ≤ u < v let
G u , v = ∏ u ≤ p < v p ≥ 3 p − 1 p − 2 .
Lemma 2.8. For every even N and 2 ≤ u < v, ∏ u ≤ p < v ,   p ∤ N 1 − g N p − 1 ≤ G u , v , with equality if N is a power of 2.
Proof. Factors with p∣N, in particular the factor at p = 2, equal 1, and every other factor equals p − 1 p − 2 ≥ 1 . ◻
Lemma 2.9 (The range u ≥ 286). For 286 ≤ u < v one has G u , v ≤ log v log u exp 1 log 2 u + 2 u − 2 .
Proof. We may assume that [u,v) contains a prime. Since log 1 + 1 p − 2 ≤ 1 p − 2 = 1 p + 2 p p − 2 ,
log G u , v ≤ ∑ u ≤ p < v 1 p + ∑ m ≥ ⌈ u ⌉ 1 m − 2 − 1 m ≤ ∑ u ≤ p < v 1 p + 2 u − 2 .
By [RS62, Theorem 5], with B1 the Meissel–Mertens constant, ∑ p ≤ x 1 / p > loglog x + B 1 − 1 2 log 2 x for x > 1 and ∑ p ≤ x 1 / p < loglog x + B 1 + 1 2 log 2 x for x ≥ 286. Let p* be the largest prime below v; then p* ≥ u ≥ 286. For every x < u such that no prime lies in [x,u) we get
∑ u ≤ p < v 1 p = ∑ p ≤ p * 1 p − ∑ p ≤ x 1 p ≤ log log v log x + 1 2 log 2 p * + 1 2 log 2 x ,
and letting x → u− gives ∑ u ≤ p < v 1 / p ≤ log log v / log u + 1 / log 2 u . ◻
Lemma 2.10 (Reduction to the primes). Let q0 = 2 < q1 = 3 < q2 = 5<… be the primes, P0 = 1 and P j = ∏ t = 1 j q t − 1 q t − 2 for j ≥ 1. Put I1 = (2),3 and Ii = (qi−1,qi] for i ≥ 2. Let K > 0 and i ≥ 1. If
P j log q j ≤ P i − 1   log q i + K log 2 q i     for   every   j ≥ i ,
then G u , v ≤ log v log u 1 + K log u for all u ∈ Ii and all v > u.
Proof. Let u ∈ Ii and v > u. If v ≤ qi, then G(u,v) = 1 < logv/logu. Otherwise let j ≥ i be such that qj < v ≤ qj+1. The primes p ≥ 3 in [u,v) are qi,…,qj, so G(u,v) = Pj/Pi−1. Since t ↦ (t + K)/t2 is decreasing on t > 0, logu ≤ logqi and logv > logqj, the right-hand side is at least logqj (logqi + K)/log2qi, and (2.2) gives the claim. ◻
Proof of Theorem E. By Lemma 2.8 it suffices to prove G u , v ≤ log v log u 1 + log 3 log u .
Optimality. If 3∤N, u = 3 and 3 < v < 5, the product in Theorem E equals 2, while the right side with log3 replaced by K tends to 1 + K/log3 as v → 3+. Hence no K < log3 is admissible uniformly in N.
The range u ≥ 286. Let L = logu ≥ log286 > 5.65 and x = 1/L2 + 2/(u − 2) < 0.04. Since ex ≤ 1 + 1.05x for 0 ≤ x ≤ 0.04, Lemma 2.9 shows that it suffices to have 1.05(1/L2 + 2/(u − 2)) ≤ log3/L, that is, 1.05/L + 2.1 L/(u − 2) ≤ log3. Here 1.05/L < 0.19, and L/(u − 2) = logu/(u − 2) is decreasing for u ≥ 286, with value less than 0.02 at u = 286; so the left side is less than 0.24 < log3. For 283 < u < 286 there are no primes in [u,286), so G(u,v) = G(286,v) for v > 286 and G(u,v) = 1 otherwise, while the right-hand side decreases in u; these u are therefore covered as well.
The range 2 ≤ u ≤ 283. Here u ∈ Ii with 1 ≤ i ≤ 60, since q60 = 283. Write Fj = Pj/logqj and Hi = Pi−1(logqi + log3)/log2qi. By Lemma 2.10 it suffices to show that Fj ≤ Hi for 1 ≤ i ≤ 60 and j ≥ i. For j ≥ 61 we have qj > 286 and Pj = P60 G(286,v) for v ∈ (qj,qj+1], so Lemma 2.9, on letting v → q j + , gives
F j ≤ P 60 log 286 exp 1 log 2286 + 2 284 .
A direct calculation with the primes up to 283 gives P60 < 7.75, hence Fj < 1.43 for j ≥ 61, and
F 1 = H 1 = 2 log 3 ,     F 2 = 8 3 log 5 < 1.66 ,     F 3 = 16 5 log 7 < 1.65 ,     F j < 1.52   4 ≤ j ≤ 60 ,
H 2 = 2 log 15 log 25 > 2.09 ,     H 3 = 8 log 21 3 log 27 > 2.14 ,     H i > 1.62   4 ≤ i ≤ 60 .
Since 2/log3 > 1.82, these inequalities give Fj ≤ Hi for all 1 ≤ i ≤ 60 and j ≥ i, with equality only for i = j = 1. In that case u ∈ (2),3 and 3 < v ≤ 5, so G(u,v) = 2 while the right-hand side exceeds 2logv/log3 > 2; thus the inequality of Theorem E is strict. ◻
Remark 2.11. The extremal configuration is u = 3, v → 3+. If the range is restricted to u ≥ u0, Lemma 2.9 bounds the optimal constant by O(1/logu0). For the full range, Theorem E gives the exact constant K = log3 in (2.1), uniformly for the Goldbach densities gN with even N.

2.4. Primes in Arithmetic Progressions

Theorem 2.12 (Bombieri–Vinogradov). For every C > 0,
∑ m ≤ x 1 / 2 log x − C − 5 / 2 max a , m = 1   max y ≤ x π y ; m , a − li y φ m ≪ C x log x C .
This is due to Bombieri [Bom65] and A. I. Vinogradov [Vin65]; the exponent C + 5/2 is due to Vaughan, see [KT05, Theorem 8 and the remark following it].
Conjecture 2.13 (Elliott–Halberstam [EH70]). For every ε > 0 and C > 0,
∑ m ≤ x 1 − ε   max a , m = 1   max y ≤ x π y ; m , a − li y φ m ≪ ε , C x log x C .

3. The Restricted Chen Sieve

3.1. Definitions

Definition 3.1. Let p = (c,a,β,θ) be admissible (1.1) and α = (a − 1)/a. For an integer N ≥ 2 put A=⌈cN⌉, B=⌊(1 − c)N⌋, z = Nβ, D=⌊Nθ⌋,
R N = r   prime :   A α < r ≤ B ,   r ∤ N ,     w r = B / r   r ∈ R N ,
S N = n ∈ Z :   A ≤ n ≤ B ,   N − n   prime ,     S N r = n ∈ S N :   P − n > z ,
J N = ∑ n ∈ S N r b a n ,
with ba as in the introduction. The kernel and the truncated kernel are
K N n = F D - n ; P N z − ∑ r ∈ R N r ∣ n F D / r + n / r ; P N w r , K N * n = F D - n ; P N z − ∑ r ∈ R N ,   r ∣ n A r ≤ n ≤ B r F D / r + n / r ; P N w r ,
where Ar = max(A,r2) and Br = min(B,⌈r1/α⌉−1).
For an integer n the condition n ≤ Br means n ≤ B and n < r1/α. For a prime r ≤ Aα the window [Ar,Br] would be empty, since then r1/α ≤ A; so K N * would not change if the sum were extended over all primes r ∈ z , B   not dividing N.
Lemma 3.2 (Size conditions). Let p be admissible. There is N0 = N0(p) such that for all N ≥ N0:
(a) z < Aα and A3α > N;
(b) B3/2 < A2 (hence B < A ), z < B and B < z3;
(c) B3/2 < D2;
(d) for every r ∈ R N : wr < z < r and wr < D/r;
(e) if n ∈ SN and p ≤ B is a prime dividing n, then p∤N; in particular, for n ∈ SN we have (n,PN(z)) = 1 if and only if P−(n) > z.
Proof. Since A≍B≍N, (a)–(c) follow from the strict inequalities β < α, 3α > 1, 3 2 < 2 , β < 1 2 , 3β > 1 and 3 2 < 2 θ . For (d), wr < z is equivalent to B < z2r, which holds because r > Aα and 2β + α > 1; z < r follows from (a); and wr < D/r is equivalent to Br < D2, which follows from r ≤ B and (c). For (e), if p∣n and p∣N then p divides the prime N − n ≥ N − B ≥ cN, so p = N − n > B , a contradiction. Since z < B , every prime p ≤ z dividing n then divides PN(z). ◻

3.2. Coefficients

Proposition 3.3 (Coefficients of the kernel). For every n ≥ 1, K N n = ∑ m ∣ n W N m , where W N = W N low − W N high with
W N low m = λ D - m   1 m ∣ P N z ,     W N high m = ∑ r ∈ R N r ∣ m λ D / r + m / r   1 m / r ∣ P N w r .
For admissible p and N ≥ N0 the following hold for every m ≥ 1.
(i) If WN(m) ≠ 0 then m is squarefree, (m,N) = 1, m < D and WN(m) = μ(m).
(ii) W N low m   W N high m = 0 .
(iii) If W N high m ≠ 0 then r : = P + m ∈ R N , r ∥ m, every prime factor of e:=m/r is at most wr, and W N high m = λ D / r + e .
Proof. The divisor-sum identity follows by interchanging summations: the term of r in KN(n) is ∑ e ∣ n / r , P N w r λ D / r + e , and with m = re the condition e∣n/r becomes m∣n.
(iii) If the term of r in W N high m is nonzero, then e = m/r divides PN(wr), so its prime factors are at most wr < r by Lemma 3.2(d). Hence r = P+(m), r ∥ m, and at most one r contributes.
(i) For the high family, λ D / r + e ≠ 0 forces e = 1 or e < D/r, so m < D (if e = 1 then m = r ≤ B < D ); moreover (m,N) = 1 because r∤N and e∣PN(wr), and WN(m) = −μ(e) = μ(r)μ(e) = μ(m). For the low family, m∣PN(z) is squarefree and coprime to N, and λ D − m ≠ 0 forces m = 1, m prime (so m ≤ z < D), or m < D; the value is μ(m).
(ii) If W N low m ≠ 0 then P+(m) ≤ z < Aα, so no element of R N divides m. ◻

3.3. The Truncated Kernel

Proposition 3.4 (Structure of the truncated kernel). Let p be admissible, N ≥ N0 and n ∈ SN. Then
(i) K N n ≤ K N * n ;
(ii) K N * n = b a n if n ∈ S N r ;
(iii) K N * n ≤ 0 if n ∉ S N r .
In particular K N n ≤ K N * n ≤ b a n   1 S N r n .
Proof. (i) The terms omitted in K N * are values of F+, which are nonnegative by Corollary 2.3.
(iii) If n ∉ S N r then (n,PN(z)) ≠ 1 by Lemma 3.2(e), so F D − n ; P N z ≤ 0 by Corollary 2.3, and the subtracted terms are nonnegative.
(ii) Let n ∈ S N r . Then (n,PN(z)) = 1, so the first term of K N * n equals 1. Since n ≤ B < z3 and P−(n) > z, either n is prime, or n = rq with primes z < r < q, or n = r2 with r > z prime. A prime r ' ∈ R N contributes to K N * n only if r'∣n and Ar' ≤ n < r'1/α.
If n is prime, r'∣n forces r ' = n ≥ A > B , which is impossible; so K N * n = 1 = b a n .
If n = rq with r < q, then r' = q is excluded because Aq ≥ q2 > n. For r' = r we have n ≥ Ar, and n < r1/α is equivalent to q < r1/(a−1), that is, to r > qa−1. Suppose r > qa−1. Then r1/α = r ⋅ r1/(a−1) > rq = n ≥ A, so r > Aα; also r ≤ n ≤ B and r∤N by Lemma 3.2(e), so r ∈ R N . Moreover q = n / r ≥ A / r > B / r = w r because Br ≤ B3/2 < A2, so (q,PN(wr)) = 1 and F D / r + q ; P N w r = 1 . Hence K N * n = 1 − 1 r > q a − 1 = b a n .
If n = r2, then r ≥ A > A α , r ≤ B and r∤N, so r ∈ R N ; further Ar ≤ n, n = r2 < r1/α since α < 1 2 , and F D / r + r ; P N w r = 1 since r > wr. Hence K N * n = 0 = b a r 2 . ◻
Corollary 3.5 (Exact loss identity). For N ≥ N0 the quantity L N = ∑ n ∈ S N ∖ S N r − K N * n is nonnegative and ∑ n ∈ S N K N * n = J N − L N .

3.4. Main Terms and Remainders

For m ≥ 1 let Am(N) = #{p≡N (mod m): N − B ≤ p ≤ N − A},
Y N = li N − A − li N − B − 1 ,     ρ N m = A m N − Y N φ m ,
and for r ∈ R N define A m r N , YN,r and ρ N r m in the same way with [A,B] replaced by [Ar,Br] (all three being 0 if Ar > Br). Put
Σ N - = ∑ d ∣ P N z λ D - d φ d ,     Σ N , r + = ∑ e ∣ P N w r λ D / r + e φ e ,
M N = Y N Σ N - − Y N ∑ r ∈ R N Σ N , r + r − 1 , M N * = Y N Σ N - − ∑ r ∈ R N Y N , r Σ N , r + r − 1 , E N = ∑ d ∣ P N z λ D - d ρ N d − ∑ r ∈ R N ∑ e ∣ P N w r λ D / r + e ρ N re , E N * = ∑ d ∣ P N z λ D - d ρ N d − ∑ r ∈ R N ∑ e ∣ P N w r λ D / r + e ρ N r re ,
and Δ N = M N * − M N .
Proposition 3.6 (Count-to-main expansion). For N ≥ N0, ∑ n ∈ S N K N n = M N + E N and ∑ n ∈ S N K N * n = M N * + E N * .
Proof. The map n ↦ N − n is a bijection from SN onto the primes in [N − B,N − A], and m∣n if and only if N − n≡N (mod m). Hence #{n ∈ SN:m∣n} = Am(N) and # n ∈ S N ∩ A r , B r : m ∣ n = A m r N . Interchanging summations,
∑ n ∈ S N K N n = ∑ d ∣ P N z λ D - d A d N − ∑ r ∈ R N ∑ e ∣ P N w r λ D / r + e A re N .
By Lemma 3.2(d), (r,e) = 1, so φ(re) = (r − 1)φ(e), and inserting Am(N) = YN/φ(m) + ρN(m) gives the first identity. The second is identical. ◻
Proposition 3.7 (Four identities). Let
H N = ∑ n ∈ S N ∖ S N r − K N n ,     G N = ∑ n ∈ S N K N * n − K N n ,     G N r = ∑ n ∈ S N r K N * n − K N n .
For N ≥ N0 the quantities H N , L N , G N , G N r are nonnegative, and
H N = L N + G N − G N r ,   E N = J N − H N − G N r − M N ,   E N * = J N − L N − M N * ,   E N * − E N = G N − Δ N .
Proof. Nonnegativity follows from Proposition 3.4. The first identity is a splitting of a finite sum. For the second, ∑ S N K N = ∑ S N r K N * − G N r − H N = J N − G N r − H N by Proposition 3.4(ii); compare with Proposition 3.6. The third combines Corollary 3.5 with Proposition 3.6, and the fourth follows from the second, the third and the first. ◻
Corollary 3.8. For N ≥ N0: MN + EN ≤ JN and E N * ≥ E N − Δ N .
Remark 3.9 (Odd N). For large odd N one has JN = 0: if n ∈ S N r then N − n ≥ cN > 2 is an odd prime, so n is even and P−(n) = 2 ≤ z. Accordingly VN(y) = 0 for odd N and y ≥ 2, and all statements about positivity concern even N.

4. The Main Term

Lemma 4.1 (Sieve parameters). Let p be admissible. As N → ∞, logD/logz = θ/β + o(1) with 2 < θ/β < 3, and, uniformly for r ∈ R N with tr = logr/logN,
s r : = log D / r log w r = 2 θ − t r 1 − t r + o 1 ∈   4 θ − 2 − o 1 ,   2 θ − α 1 − α + o 1 ⊂ 1 , 2 .
Proof. We have logD = θlogN + o(1), log(D/r) = (θ − tr)logN + o(1) and log w r = 1 2 1 − t r log N + O 1 with 1 − t r ≥ 1 2 − o 1 , and α − o 1 < t r ≤ 1 2 + o 1 . The map t ↦ 2(θ − t)/(1 − t) is decreasing, its derivative being −2(1 − θ)/(1 − t)2, with value 4θ − 2 > 1 at t = 1 2 (as θ > 3 4 ) and 2(θ − α)/(1 − α) < 2 at t = α (as θ < 1). Finally θ/β > 2 because θ > 2β, and θ/β < 3 because θ < 1 < 3β. ◻
At p0 these values are θ/β = 98/45 = 2.177… and sr ∈ [1.92,1.924] + o(1).
Lemma 4.2 (A prime sum). For admissible p,
∑ r ∈ R N 1 r − 1 log D / r = 1 θ log N log θ − α α 2 θ − 1 + o 1 .
Proof. Let x0 = Aα, x 1 = B and θN = logD/logN = θ + o(1). At most three primes in (x0,x1] divide N, since each exceeds N1/3 for large N, and their contribution is O(N−1/3); replacing 1/(r − 1) by 1/r costs O ∑ r > x 0 r − 2 = O x 0 − 1 . Hence the sum equals (logN)−1Σ + o((logN)−1) with Σ = ∑ x 0 < p ≤ x 1 h p / p and h(x) = 1/(θN − logx/logN). On [x0,x1] the function h is increasing and bounded by 1 / θ N − 1 2 − o 1 ≤ 4 . Write S x = ∑ p ≤ x 1 / p = loglog x + B 1 + ϵ x , where ϵ(x) → 0 by Mertens’ theorem. Integrating by parts,
Σ = ∫ x 0 x 1 h x   dx x log x + h   ϵ x 0 x 1 − ∫ x 0 x 1 ϵ x   dh x ,
and the last two terms are O sup x ≥ x 0 ϵ x = o 1 because h is bounded and monotone. The substitution t = logx/logN turns the first integral into ∫ t 0 t 1 dt t θ N − t with t0 → α and t 1 → 1 2 , which tends to ∫ α 1 / 2 dt t θ − t = 1 θ log θ − α α 2 θ − 1 . ◻
Theorem 4.3. For every admissible p, M N ≥ C p − o 1 R N as N → ∞ through even integers.
Proof. Let N be even and let P be the set of primes not dividing N; in particular 2 ∉ P . The function g = gN satisfies 0 < g(p) < 1 on P and, by Theorem E, condition (2.1) with K = log3, uniformly in N. Fix ε > 0.
Lower bound for Σ N − . With z'=⌊z⌋+1 we have P N z = ∏ p ∈ P ,   p < z ' p and logD/logz' = θ/β + o(1) ∈ (2,3) by Lemma 4.1. Theorem 2.5 and the continuity of f give, for large N,
Σ N - ≥ V N z f θ / β − ε ,     f θ / β = 2 e γ β θ log θ β − 1 .
Upper bound for Σ N , r + . Similarly, with sifting limit ⌊wr⌋+1 and level D/r, Lemma 4.1 gives a sieve parameter sr + o(1) ∈ (1,2) uniformly in r, and log D / r ≥ θ − 1 2 log N . Hence, for large N and uniformly in r ∈ R N ,
Σ N , r + ≤ V N w r 2 e γ s r + ε .
Asymptotics. Since z = Nβ and N1/5 ≤ wr ≤ N for large N, Corollary 2.7 gives V N z = 1 + o 1 e − γ S N / β log N and V N w r = 1 + o 1 e − γ S N / log w r uniformly in r. Using srlogwr = log(D/r) and log w r ≥ 1 5 log N we obtain
Σ N - ≥ 2 S N θ log N log θ β − 1 − O ε − o 1 ,     Σ N , r + r − 1 ≤ 1 + o 1 2 S N r − 1 log D / r + O ε   S N r − 1 log N .
Summing over r ∈ R N , using Lemma 4.2 and ∑ r ∈ R N 1 / r − 1 = O 1 , we get
Σ N - − ∑ r ∈ R N Σ N , r + r − 1   ≥   2 S N θ log N log θ / β − 1 α 2 θ − 1 θ − α − O ε − o 1 .
Finally Y N = ∫ N − B − 1 N − A d t / log t = 1 + o 1 1 − 2 c N / log N . Multiplying, and letting ε → 0, we obtain M N ≥ C p − o 1 R N . ◻
Proposition 4.4 (Sign of the coefficient). Let p be admissible and u = α + β + 2αβ. Then u2 > 16α2β, and C p > 0 if and only if θ > θ * α , β = u + u 2 − 16 α 2 β / 4 α . For α = 9 19 , β = 9 20 one has θ * = 57 + 3 41 / 80 = 0.95261 … , and
C p 0 = 50 49 log 2544 2405 = 0.057334 … > 3475 62328 > 11 200 .
Proof. Since 2(1 − 2c)/θ > 0 and θ > α, C p > 0 is equivalent to (θ − β)α(2θ − 1) > β(θ − α), that is, to q(θ):=2αθ2 − uθ + 2αβ > 0. The discriminant Δ(α,β) = u2 − 16α2β satisfies ∂ Δ / ∂ β = 2 u 1 + 2 α − 16 α 2 ≥ 1 3 − 28 α 2 + 14 α + 2 > 0 for 1 3 ≤ α ≤ 1 2 (using u ≥ 5 3 α + 1 3 for β ≥ 1 3 ), and Δ α , 1 3 = 1 9 − 23 α 2 + 10 α + 1 > 0 on the same interval; hence Δ > 0 on the admissible domain. The product of the two roots of q is β, so the smaller root θ− satisfies θ − < β < 3 4 < θ . Thus q(θ) > 0 if and only if θ > θ*, the larger root. At α = 9 19 , β = 9 20 we get u = 27 20 , u 2 − 16 α 2 β = 729 ⋅ 41 380 2 , and θ * = 57 + 3 41 80 . At p0, θ / β − 1 α 2 θ − 1 θ − α = 53 45 ⋅ 432 481 = 2544 2405 and 2 1 − 2 c θ = 50 49 ; finally log(1 + t) > t/(1 + t) with t = 139 2405 gives C p 0 > 50 49 ⋅ 139 2544 = 3475 62328 . ◻
Remark 4.5. The algebraic threshold θ*(α,β) may lie below 3/4. For example, (α,β) = (0.49,0.34) gives θ* = 0.7039…, so C p > 0 throughout the admissible range of θ. The condition θ > 3/4 has a separate role: it ensures sr > 1 in Lemma 4.1. At exponents θ ≤ 1/2, the lower linear-sieve factor satisfies f(θ/β) = 0 because β > 1/3, and therefore supplies no positive contribution to the derived lower bound.
Proposition 4.6 (Threshold for the explicit lower-bound coefficient). Let 1 3 < β < α < 1 2 .
(i) θ * α , β > 2 3 , and θ * α , β → 2 3 as α , β → 1 2 , 1 3 .
(ii) For 2β < θ < 1 the expression C c , a , β , θ of Theorem A is positive if and only if θ > θ*(α,β); and for 0 < θ ≤ 2β one has f(θ/β) = 0.
Consequently, for a sifting limit z = Nβ with β > 1/3, positivity of the coefficient C requires θ > 2/3. The condition β > 1/3 ensures that the rough elements of SN have at most two prime factors. The estimate of Theorem A applies in the admissible subrange θ > max(2β,3/4).
Proof. With q(θ) = 2αθ2 − uθ + 2αβ as in the proof of Proposition 4.4,
q 2 / 3 = 2 / 9 α 1 + 3 β − 3 β < 2 / 9 1 + 3 β / 2 − 3 β = 1 / 9 1 − 3 β < 0 ,
so 2 3 lies strictly between the two roots of q, and θ * > 2 3 . At α , β = 1 2 , 1 3 one has q θ = θ − 2 3 θ − 1 2 , and the roots depend continuously on (α,β). For (ii), the proof of Proposition 4.4 shows that C > 0 is equivalent to q(θ) > 0 for every θ > α; since θ > 2 β > β and the smaller root of q is less than β , this happens exactly when θ > θ*. Finally f(s) = 0 for 0 < s ≤ 2. ◻
Remark 4.7. The distribution results in [Lic23, Theorem 1.7] permit exponents below 153/256 for triply well-factorable weights uniformly in residues of the size required by Goldbach-type applications. Under the spectral conjectures specified there, the corresponding range extends to exponents below 5/8; for small residues, the exponent is below 66/107. These thresholds are all less than 2/3. Thus these estimates lie below the positivity threshold in Proposition 4.6 and below the admissible levels of the restricted family. Chen-type applications of such distribution estimates [Lic23, Li26] use additional weights and switching to treat integers with three or more prime factors after sifting to smaller powers. Proposition 7.8 identifies a separate restriction arising from the support of the coefficients.

5. The Remainder

5.1. Reduction to High Moduli

Proposition 5.1 (Coefficient form). For N ≥ N0 the remainder satisfies
E N = ∑ m < D W N m ρ N m .
Proof. By Proposition 3.3(iii), the map (r,e) ↦ re is injective on the pairs with r ∈ R N , e∣PN(wr) and λ D / r + e ≠ 0 , since r = P+(re); its image is the support of W N high , and W N high re = λ D / r + e . Hence the second sum in the definition of EN equals ∑ m W N high m ρ N m ; the first equals ∑ m W N low m ρ N m by definition. The range m < D comes from Proposition 3.3(i). ◻
With QN=⌊N1/2(logN)−8⌋ we define
E N low = ∑ m ≤ Q N W N m ρ N m ,     E N high = ∑ Q N < m < D W N m ρ N m .
Proposition 5.2 (Small moduli). For admissible p, E N low = O N log N − 3 .
Proof. If WN(m) ≠ 0 then (m,N) = 1 and |WN(m)| ≤ 1 by Proposition 3.3(i). For such m, ρN(m) = Δ(N − A;m,N)−Δ(N − B − 1;m,N) with Δ(y;m,a) = π(y;m,a) − li(y)/φ(m). Hence E N low ≤ 2 ∑ m ≤ Q N max a , m = 1 max y ≤ N Δ y ; m , a , and Theorem 2.12 with C = 3 applies, since QN ≤ N1/2(logN)−11/2. ◻
Theorem 5.3. For every admissible p, as N → ∞ through even integers,
J N   ≥   C p − o 1 R N + E N high .
Proof. By Corollary 3.8, J N ≥ M N + E N low + E N high . Apply Theorem 4.3 and Proposition 5.2, and note that N(logN)−3 = o(RN) because S N ≥ 2 C 2 > 0 . ◻
Corollary 5.4. Let p be admissible.
(i) If E N high ≥ − κ R N for all large even N, with some κ < C p , then J N ≥ C p − κ − o 1 R N . At p0 and κ = 1 500 this gives JN ≥ RN/20 for all large even N.
(ii) Assume Conjecture 2.13. Then EN = O(N(logN)−3) and J N ≥ C p − o 1 R N . In particular, if θ > θ*(α,β), every sufficiently large even N can be written as N = p + n with p prime, cN ≤ n ≤ (1 − c)N, and n either prime or a product rq of primes with Nβ < r ≤ qa−1.
Proof. (i) is immediate from Theorem 5.3; at p0, C p 0 − 1 500 > 0.0537 > 1 20 by Proposition 4.4. For (ii), E N ≤ ∑ m < D ρ N m , and Conjecture 2.13 with x = N and ε = 1 − θ bounds this by O(N(logN)−3), exactly as in the proof of Proposition 5.2. Then JN ≥ MN + EN, and Theorem 4.3 applies. ◻
Remark 5.5. Corollary 5.4(ii) specializes the Elliott–Halberstam hypothesis to the weighted statistic JN and its prescribed factor-size conditions. The exact identities of Section 3 and the estimate for small moduli reduce this specialization to the high-modulus remainder. Bombieri’s asymptotic sieve [Bom76] provides the broader conditional framework for studying the factorization of the sequence N − p.

5.2. Termwise Bounds for the High-Modulus Remainder

Proposition 5.6. Let p be admissible and fix 0 < d0 < d1 < b0 < b1 < β with b 0 + d 0 > 1 2 and b1 + 3d1 < θ. Let Q N be the set of pairs of primes (u,v) with (uv,N) = 1, N b 0 < u ≤ N b 1 and N d 0 < v ≤ N d 1 . For large N every m = uv with u , v ∈ Q N satisfies QN < m < D and WN(m) = 1; distinct pairs give distinct m; and
∑ u , v ∈ Q N 1 φ uv = log b 1 b 0   log d 1 d 0 + o 1 .
Proof. Since v < u < z and (uv,N) = 1, we have uv∣PN(z), and the only test in λ D − uv is T 2 uv = u v 3 ≤ N b 1 + 3 d 1 < D ; hence W N low uv = μ uv = 1 , while W N high uv = 0 because P+(uv) < z < Aα. Further uv > N b 0 + d 0 > Q N and uv < D. The two windows are disjoint, so m determines (u,v). The asymptotic formula is Lemma A.4. ◻
Since Am(N) ≥ 0, the termwise estimate ρN(m) ≥ −YN/φ(m) gives
∑ u , v ∈ Q N W N uv ρ N uv ≥ − log b 1 b 0 log d 1 d 0 + o 1 Y N .
Here YN∼(1 − 2c)N/logN, whereas R N = S N N / log 2 N . The ratio of these scales is Y N / R N ∼ 1 − 2 c log N / S N ; in particular, it is of order logN along families with bounded singular series. The calculation isolates the loss arising from termwise nonnegativity, and motivates estimates that use the distribution of primes across the high-modulus family. At p0, one admissible choice is (d0,d1,b0,b1) = (0.11,0.17,0.40,0.44).

5.3. The Remainder on Average Over N

We write ϑ t ; m , a = ∑ p ≤ t ,   p ≡ a   mod   m log p and ψ t ; m , a = ∑ n ≤ t ,   n ≡ a   mod   m Λ n , and ∑ a *   denotes summation over the reduced residues a modulo m.
Lemma 5.7 (Barban–Davenport–Halberstam for prime counts). For every A > 0, uniformly for y ≥ 2 and M ≥ 1,
∑ m ≤ M ∑ a * π y ; m , a − li y φ m 2 ≪ A My log 2 2 My + y 2 log y A .
Proof. Let Eψ(t;m,a) = ψ(t;m,a) − t/φ(m). By the Barban–Davenport–Halberstam theorem [DH66], [Dav00, Ch. 29],
∑ m ≤ Q ∑ a * E ψ t ; m , a 2 ≪ A tQ log t     t log t − A − 1 ≤ Q ≤ t ;
as the left side increases with Q, it is ≪t2(logt)−A for smaller Q. For Q > t we use ψ(t)≪t and maxaψ(t;m,a) ≤ (t/m + 1)logt, which give ∑ a * E ψ t ; m , a 2 ≪ t t / m + 1 log t + t 2 / φ m . Hence, for all t ≥ 2 and Q ≥ 1,
∑ m ≤ Q ∑ a * E ψ t ; m , a 2 ≪ A tQ log 2 2 tQ + t 2 log t − A .  
Since ∑ a * ψ − ϑ t ; m , a 2 ≤ ψ t − ϑ t 2 ≪ t , the bound (5.2) also holds for Eϑ(t;m,a) = ϑ(t;m,a) − t/φ(m). By partial summation, for y ≥ 2,
π y ; m , a − li y φ m = E ϑ y ; m , a log y + ∫ 2 y E ϑ t ; m , a t log 2 t   dt + 2 φ m log 2 ,
and by the Cauchy–Schwarz inequality
∫ 2 y E ϑ t ; m , a t log 2 t   dt 2 ≤ 1 log 2 ∫ 2 y E ϑ t ; m , a 2 t log 2 t   dt .
Summing over m ≤ M and a, inserting (5.2) with Q = M, and using ∫ 2 y d t / log 2 t ≪ y , ∫ 2 y t log t − A − 2 dt ≪ A y 2 log y − A − 2 and ∑ m ≤ M 1 / φ m ≪ log 2 M , we obtain the lemma. ◻
Lemma 5.8 (Averaging over N). Let 0 < θ0 < 1, C ≥ 1 and A > 0. For X ≥ 2 and integers N ∈ (X,2X] let uN ≤ vN be integers with uN ≥ X/C, vN ≤ 2X, and uN+1 − uN, vN+1 − vN ∈ {0,1}. Put
ρ N ∘ m = π v N ; m , N − π u N − 1 ; m , N − li v N − li u N − 1 φ m .
Then
∑ X < N ≤ 2 X ∑ m ≤ X θ 0 m , N = 1 ρ N ∘ m ≪ A , C , θ 0 X 2 log X A .
Proof. We may assume that X is large. Put M = X θ 0 and H = ⌈ X 1 + θ 0 / 2 ⌉ , so that M ≤ H, and split (X,2X] into at most X/H + 1 blocks of at most H consecutive integers. Write Δ(t;m,a) = π(t;m,a) − li(t)/φ(m), so that ρ N ∘ m = Δ v N ; m , N − Δ u N − 1 ; m , N . Fix a block B with first element N0, and let wN denote either vN or uN − 1. Then w N 0 ≤ w N ≤ w N 0 + H for N ∈ B , so for m ≤ M,
Δ w N ; m , N − Δ w N 0 ; m , N ≤ H m + 1 + H φ m ≤ 3 H φ m ,
by counting the integers ≡N (mod m) in w N 0 , w N 0 + H and using li'(t) ≤ 1 for t ≥ e. Summed over N ∈ B and m ≤ M, the error caused by replacing wN by w N 0 is ≪H2logX. Next, as N runs over B , each reduced residue class modulo m ≤ M ≤ H occurs at most H/m + 1 ≤ 2H/m times. By the Cauchy–Schwarz inequality and Lemma 5.7 with y = w N 0 ≤ 2 X and 2A + 1 in place of A,
∑ m ≤ M ∑ N ∈ B m , N = 1 Δ w N 0 ; m , N ≤ 2 H ∑ m ≤ M φ m m 2 1 / 2 ∑ m ≤ M ∑ a * Δ w N 0 ; m , a 2 1 / 2 ≪ HX log X A .
Summing over the blocks gives a total
≪ X H + 1 H 2 log X + HX log X A ≪ X 3 + θ 0 / 2 log X + X 2 log X A .
◻
Proposition 5.9 (The remainder on average). Let p be admissible. For every A > 0,
∑ X < N ≤ 2 X E N ≪ A , p X 2 log X A .
Proof. For large N, Propositions 3.3(i) and 5.1 give E N ≤ ∑ m < D N ,   m , N = 1 ρ N m , and D N ≤ 2 X θ ≤ X θ 0 with θ0 = (1 + θ)/2 for large X. With uN = N − BN and vN = N − AN we have ρ N m = ρ N ∘ m , uN ≥ cN ≥ cX, vN ≤ 2X, and uN, vN increase by 0 or 1 when N increases by 1. Lemma 5.8 gives the claim. ◻
Proof of Theorem F. The first assertion is Proposition 5.9. If |EN| > RN/logN, then |EN| ≥ 2C2X/log3(2X), so by Proposition 5.9 with A + 3 in place of A the number of such N ∈ (X,2X] is O(X(logX)−A). For every other large even N, Corollary 3.8 and Theorem 4.3 give J N ≥ M N + E N ≥ C p − o 1 R N − R N / log N . ◻
Remark 5.10. The circle method gives almost-all binary Goldbach representations and quantitative bounds for the exceptional set; see [MV75]. Every representation N = p + p' with p' ∈ [A,B] prime is counted by JN. Theorem F addresses the coefficient-dependent remainder of the restricted sieve: it bounds ∑ E N and transfers the explicit lower-bound coefficient C p to JN. Together with Theorem B, it localizes any failure of that lower bound to an exceptional set on which E N high is sufficiently negative. The distinction between such an average estimate and a pointwise transfer criterion is made precise in Section 5.4.

5.4. From Averages to Individual N

The following identities formulate one-sided conditions for transferring a lower bound between nearby even integers. For h ≥ 0, write AM,BM,zM for the quantities in Definition 3.1 attached to M, and put
I h N = ∑ A N ≤ n ≤ B N 1 P N + h − n − 1 P N − n   1 P − n > z N   b a n , B h N = ∑ n 1 A N + h , B N + h n − 1 A N , B N n   1 P N + h − n   1 P − n > z N   b a n , K h N = ∑ A N + h ≤ n ≤ B N + h 1 P N + h − n 1 P − n > z N − 1 P − n > z N + h   b a n .
Proposition 5.11 (Shift identity). For integers N ≥ 2 and h ≥ 0: J N + h − J N = I h N + B h N − K h N , K h N ≥ 0 and B h N ≤ h .
Proof. The identity is telescoping: one passes from JN to JN+h by first shifting the prime indicator, then the interval, then the threshold. Since zN ≤ zN+h, K h N ≥ 0 . The positive terms of B h N lie in (BN,BN+h] and the negative ones in [AN,AN+h), each of modulus at most 1; as AN+h − AN≤⌈ch⌉≤h and BN+h − BN≤⌈(1 − c)h⌉≤h, we get B h N ≤ h . ◻
Corollary 5.12 (One-sided transfer). Let m > 0, X ≥ 2 and let E be a set of even integers with JM ≥ m for M ∈ E . Suppose that every even N ∈ X , 2 X ∖ E admits an even h with 0 < h ≤ m/4, N + h ∈ E and I h N ≤ m / 2 . Then JN ≥ m/4 for every even N ∈ [X,2X].
Proof. For N ∉ E , Proposition 5.11 gives J N = J N + h − I h N − B h N + K h N ≥ m − m 2 − m 4 . ◻
Lemma 5.13 (Deficit criterion). Let E ' be a finite set, J , t : E ' → R , and E ⊂ E ' with J > 0 on E . If t ≥ τ > 0 on E ' ∖ E and ∑ N ∈ E ' ∖ E t N − J N + 2 < τ 2 , then J > 0 on E ' .
Proof. If JN ≤ 0 for some N ∈ E ' ∖ E , then t N − J N + ≥ τ . ◻
The explicit threshold τ cannot be dispensed with, as the following example shows.
Example 5.14 (Small mean square does not imply positivity). Let tN = max(N/log2N,1), JN = 0 if N = 2k with k ≥ 2, and JN=⌈tN⌉ otherwise. Every real X ≥ 4 has exactly one power 2k, k ≥ 2, in [X,2X), so ∑ X ≤ N < 2 X ,   N even 1 − J N / t N + 2 = 1 , which is o(X(logX)−C) for every C > 0, although JN = 0 for infinitely many even N.

6. The Kernel Is Unbounded Below

In this section we prove Theorem C.
Proof of Theorem C(i). Let k ≥ 3 with θ k ∉ 2 Z , put m0 = 2⌊θk/2⌋ and δ=θk − m0 ∈ (0,2). Since θ > 3 4 we have θk > 2, hence m0 ≥ 2, and m0 < θk < k. Let ξ be large. By the prime number theorem there are primes ξ < q1<…<qk ≤ 2ξ. Put n = q1 … qk, let q be a prime in (n,2n] (Bertrand’s postulate) and N = n + q. Then ξk < n ≤ (2ξ)k and 2n < N ≤ 3n, so n / N ∈ 1 3 , 1 2 and q / N ∈ 1 2 , 2 3 ; as c ≤ 1 3 , both n and q lie in [A,B]. All primes involved are odd, so N is even, and n ∈ SN because N − n = q is prime. No qi divides N, since otherwise qi∣q with q > n > qi.
For large ξ we have 2ξ ≤ min(z,Aα), because z = N β > 2 ξ k β and A α ≥ c ⋅ 2 ξ k α with kβ > 1 and kα > 1. Hence (n,PN(z)) = n, no element of R N divides n, and K N n = F D − n ; P N z . Moreover 2 ξ k θ − 1 < D ≤ 3 ⋅ 2 k ξ k θ . Since θk − m0=δ>0 and m0 + 2 − θk = 2−δ>0, for large ξ
q k m 0 ≤ 2 ξ m 0 < 2 ξ k θ − 1 < D     and     q 1 m 0 + 2 > ξ m 0 + 2 ≥ 3 ⋅ 2 k ξ k θ ≥ D .
Lemma 2.4 gives K N n = − k − 1 m 0 − 1 . Letting ξ → ∞ produces infinitely many N. ◻
Proof of Theorem C(ii). For k ≥ 1 let xk ∈ [0,2) with xk≡θk (mod 2), and call k good if x k ∈ 1 2 , 3 2 . Among any three consecutive integers at least one is good: if x k < 1 2 then x k + 1 = x k + θ ∈ θ , θ + 1 2 ⊂ 1 2 , 3 2 ; if x k > 3 2 then x k + 1 = x k + θ − 2 ∈ θ − 1 2 , θ , and if moreover x k + 1 < 1 2 then x k + 2 = x k + 1 + θ ∈ 2 θ − 1 2 , θ + 1 2 ⊂ 1 2 , 3 2 .
Let k be good and large, put m0 = 2⌊θk/2⌋, so that δ = θ k − m 0 = x k ∈ 1 2 , 3 2 , and let ξ = 9 ⋅ 4k+1. By the prime number theorem, π(2ξ) − π(ξ) ≥ ξ/(2logξ) ≥ k for large k, and we construct n and N exactly as in the proof of (i). Now
2 ξ m 0 = 2 m 0 ξ θ k − δ ≤ 2 m 0 ξ θ k ξ − 1 / 2 ≤ 1 / 2 ξ θ k < 2 ξ k θ − 1 < D ,
because ξ 1 / 2 = 3 ⋅ 2 k + 1 ≥ 2 m 0 + 1 , and
ξ m 0 + 2 = ξ θ k + 2 − δ ≥ ξ θ k ξ 1 / 2 = 3 ⋅ 2 k + 1 ξ θ k ≥ 3 ⋅ 2 k ξ k θ ≥ D .
As before 2ξ ≤ min(z,Aα) for large k, so Lemma 2.4 gives K N n = − k − 1 m 0 − 1 = − k − 1 k − m 0 . Here 1 − θ k + 1 2 ≤ k − m 0 ≤ 1 − θ k + 3 2 ≤ 3 2 1 − θ k once k ≥ 3/(1 − θ). Using n j ≥ n / j j and k − 1 ≥ k/2,
k − 1 k − m 0 ≥ k − 1 k − m 0 k − m 0 ≥ 1 3 1 − θ 1 − θ k = e c ' θ k ,     c ' θ = 1 − θ log 1 3 1 − θ > 0 ,
the base being larger than 1 because θ > 2 3 . Finally logN ≤ log3 + klog(2ξ) ≤ 2k2 for large k, so k ≥ log N / 2 and K N n ≤ − exp c ' θ log N / 2 . As k runs through the good integers, N → ∞. ◻
Proof of Theorem C(iii). By Corollary 2.3, F D − n ; P N z ≤ τ n and F D / r + n / r ; P N w r ≤ τ n / r ≤ τ n . At most two elements of R N divide n, since the product of three of them exceeds A3α > N > n by Lemma 3.2(a). ◻
Remark 6.1. For θ = 49/50, part (i) gives −(k − 1) for odd k ≤ 49 and − k − 1 2 for even k ≤ 98. The large−k construction in part (ii) uses k ≥ 2/(1 − θ) = 100. On the other hand, part (iii) and the divisor bound give
K N n ≤ exp log 2 + o 1 log N loglog N .
Thus the established scales range from exp c log N along a subsequence to the uniform upper bound exp(ClogN/loglogN). In particular, the boundedness of the coefficients |WN(m)| ≤ 1 does not control the negative values of the divisor-sum kernel KN = WN * 1 on SN.

7. Supports of Factorable Sequences

Definition 7.1. Let k ≥ 1 be an integer and Q ≥ 1 be real. A sequence a : N → C is k−factorable of level Q if for every (Q1,…,Qk) ∈ [1,∞)k with Q1 … Qk = Q there are a 1 , … , a k : N → C with ai(n) = 0 for n > Qi and a = a1*…*ak (Dirichlet convolution). If the ai can always be chosen with |ai| ≤ 1, we call a k−well-factorable. For k = 2 this is Iwaniec’s notion of a well-factorable sequence of level Q [Iw80b], and for k = 3 Maynard’s notion of a triply well-factorable sequence [May25].
Lemma 7.2. Let k,m ≥ 1 be integers and Q ≥ 1 be real. The following are equivalent:
(a) pk−1m≥p ≤ Q for every prime p∣m;
(b) for every (Q1,…,Qk) ∈ [1,∞)k with Q1 … Qk = Q there are positive integers m1,…,mk with m = m1 … mk and mi ≤ Qi for all i.
Proof. For k = 1 both conditions say m ≤ Q, so let k ≥ 2.
(b)⇒(a). Suppose pk−1m≥p > Q for some prime p∣m. If pk−1 ≤ Q, choose h < p so close to p that hk−1m≥p > Q; then hk−1 < Q. If pk−1 > Q, put h = Q1/(k−1) < p. In both cases set Q2=…=Qk = h and Q1 = Q/hk−1 ≥ 1. If m = m1 … mk with mi ≤ Qi, then m2,…,mk ≤ h < p have no prime factor ≥p, so m≥p∣m1 and m1 ≥ m≥p > Q1 (in the second case because Q1 = 1 < p ≤ m≥p), a contradiction.
(a)⇒(b). Write m = p(1)p(2) … p(s) with primes p(1) ≥ p(2)≥…≥p(s), repeated according to multiplicity, and place them one at a time, in this order, into k slots with capacities Q1,…,Qk: at step t the prime p = p(t) is put into a slot i for which the remaining capacity Qi/mi is largest, mi denoting the current product in slot i (initially 1). Let Pt−1 = p(1) … p(t−1). The primes already placed, together with p, form a sub-multiset of the prime factors of m that are ≥p, so Pt−1 p ≤ m≥p. Hence, by (a),
∏ i = 1 k Q i m i = Q P t − 1 ≥ Q   p m ≥ p ≥ p k ,
so the largest remaining capacity is at least p, and the placement keeps mi ≤ Qi. After s steps, m = m1 … mk with mi ≤ Qi. ◻
Theorem 7.3. Let k ≥ 1 be an integer, let Q ≥ 1 be real, and set S k Q = m ≥ 1 :   p k − 1 m ≥ p ≤ Q   for   every   prime   p ∣ m .
(i) If a is k−factorable of level Q, then a(m) = 0 for every m ∉ S k Q .
(ii) If m ∈ S k Q , then δm is k−well-factorable of level Q, with factors of the form δ m 1 , … , δ m k .
In particular, S k Q is the union of the supports of all k−factorable sequences of level Q. Moreover, S k Q ⊆ S k − 1 Q for k ≥ 2, and S k Q ⊆ S k Q ' whenever Q ≤ Q'.
Proof. (i) If m ∉ S k Q , Lemma 7.2 yields (Q1,…,Qk) such that m has no factorisation m = m1 … mk with mi ≤ Qi. For the corresponding decomposition a = a1*…*ak, every term of a m = ∑ m 1 … m k = m a 1 m 1 … a k m k contains a factor ai(mi) with mi > Qi, so a(m) = 0. (ii) For each admissible (Qi), Lemma 7.2 gives mi, and δ m = δ m 1 * … * δ m k . The inclusions follow from the definition, since pk−1m≥p ≥ pk−2m≥p. ◻
Corollary 7.4. Let Q ≥ 1 and m = p 1 e 1 … p s e s with p1<…<ps.
(i) m ∈ S 2 Q if and only if every ratio of consecutive divisors of m is at most Q/m. Consequently, the largest ratio of consecutive divisors of m > 1 equals max j p j / p 1 e 1 … p j − 1 e j − 1 . Moreover P + m ≤ Q for every m ∈ S 2 Q .
(ii) If m ∈ S 3 Q then P+(m) ≤ Q1/3, and u3v ≤ Q for all primes u < v dividing m.
(iii) For primes u < v: uv ∈ S 3 Q if and only if u3v ≤ Q and v3 ≤ Q. For a prime p: p ∈ S k Q if and only if pk ≤ Q.
Proof. (i) By Lemma 7.2, m ∈ S 2 Q if and only if for every R ∈ [1,Q] the integer m has a divisor in [mR/Q,R]. If every ratio of consecutive divisors is at most Q/m, then m ≤ Q (take the ratio m/d for the largest proper divisor d; for m = 1 there is nothing to prove), and for R < m the largest divisor d ≤ R satisfies d ≥ d'/(Q/m) > mR/Q, where d' > R is the next divisor; for R ≥ m take d = m. Conversely, let m ∈ S 2 Q , so that m ≤ Q (take R = 1), and suppose that d < d' are consecutive divisors with d'/d > Q/m. Then every R ∈ (dQ/m,d') lies in [1,Q], and the divisors of m not exceeding R are at most d < mR/Q, so the condition fails. The formula for the largest ratio follows on comparing with p j   m ≥ p j = p j   m / p 1 e 1 … p j − 1 e j − 1 ; it is Tenenbaum’s well-known description [Ten86], see also [Sai97, TW24]. The last claim follows from p m≥p ≥ p2 at p = P+(m). (ii) At p = P+(m) we have p2m≥p ≥ p3, and at p = u we have u2m≥u ≥ u2 ⋅ uv. (iii) is the definition. ◻
In the terminology of the Polymath project [Pol14], Corollary 7.4(i) says that the integers carrying well-factorable weights of level Q are exactly those m that are (Q/m)−densely divisible.
Proposition 7.5 (Supports, the Schinzel–Szekeres function and Rosser’s weights). For β ≥ 0 let Fβ(1) = 1 and Fβ(m) = maxd∣m, d>1d P−(d)β for m ≥ 2.
(i) Fk−1(m) = maxp∣mpk−1m≥p for m ≥ 2; hence S k Q = m ≥ 1 :   F k − 1 m ≤ Q , and S k Q ⊆ 1 , Q .
(ii) Let D > 1 and let d > 1 be squarefree, with prime factors p1>…>pr. Then d ∈ S 3 D if and only if Tj(d) ≤ D for every 1 ≤ j ≤ r. More precisely,
supp λ D + ∩ supp λ D − = d ≥ 1 : μ 2 d = 1 ,   F 2 d < D ⊆ S 3 D .
(iii) If λ D + d ≠ 0 , then d ∈ S 2 D ; if λ D − d ≠ 0 and P+(d)2 ≤ D, then d ∈ S 2 D .
(iv) Let P be squarefree and divisible by two primes in (D1/4,D1/3). Then the sequence d ↦ λ D + d 1 d ∣ P is not a linear combination of 3−factorable sequences of level D.
Proof. (i) If d∣m and P−(d) = p, then d∣m≥p, and d = m≥p itself has P−(d) = p; hence the largest value of d P−(d)k−1 over such d is pk−1m≥p. The inclusion in [1,Q] follows on taking p = P−(m), for which m≥p = m. (ii) For p = pj we have d ≥ p j = p 1 … p j , so p j 2 d ≥ p j = T j d ; By Definition 2.1, membership in both Rosser supports is equivalent to Tj(d) < D at every index. For d > 1, their maximum is F2(d); the case d = 1 follows from F2(1) = 1 < D. This proves the strict-sublevel identity and the inclusion. (iii) For squarefree d and p = pj we have p   d ≥ p = p 1 … p j − 1 p j 2 , which is less than Tj(d) and, for j ≥ 2, less than p 1 … p j − 2 p j − 1 3 = T j − 1 d . If λ D + d ≠ 0 then Tj(d) < D for odd j, which covers odd j directly and even j through Tj−1. If λ D − d ≠ 0 then Tj(d) < D for even j, which covers even j and odd j ≥ 3, while j = 1 is the hypothesis P+(d)2 ≤ D. (iv) Let D1/4 < p2 < p1 < D1/3 be primes dividing P and d = p1p2. The only odd index is j = 1, and T 1 d = p 1 3 < D , so λ D + d = 1 ; but T 2 d = p 1 p 2 3 > D , so d ∉ S 3 D by (ii), and every linear combination of 3−factorable sequences of level D vanishes at d by Theorem 7.3(i). ◻
Remark 7.6. The support inclusions in part (iii) agree with [Lic23, §6.1], subject to the stated largest-prime-factor condition for the lower weights. Theorem 7.3 identifies S 2 D as the universal support for two-factor decompositions. Part (ii) connects triple factorability with the Rosser conditions at all indices: S 3 D uses the non-strict inequalities, while the intersection of the Rosser supports uses their strict counterparts. The relation between generalized Schinzel–Szekeres functions and such intersections is discussed in [SW26, §1.5]. Part (iv) gives an explicit modulus witnessing the failure of triple factorability for upper linear-sieve weights, as considered in [May25, Lic23].
By part (i) and [SW26, Theorem 5],
# S k Q ≍ k Q log 2 Q − λ 1 / k − 1     k ≥ 2 ,   Q ≥ 1 ,
where λ1 = 1 and λ1/2 = 2.46206…. Accordingly, the number of integers admissible for well-factorable weights is ≍Q/logQ as Q → ∞, in agreement with [Sai97], whereas the corresponding count for triply factorable weights is ≍Q(logQ)−2.462….
Lemma 7.7 (A witness prime). Let p be admissible. For large N there is a prime r ∈ R N with r > max Q N , D .
Proof. Let x = max Q N , D , A α . Then x ≥ QN > N1/3, and x = o B since Q N / B ≪ log N − 8 , D / B ≪ N θ − 1 / 2 and A α / B ≪ N α − 1 / 2 . Put n0=⌊x⌋+1, so that 8 n 0 ≤ B for large N, and choose primes p1 ∈ (n0,2n0], p2 ∈ (2n0,4n0] and p3 ∈ (4n0,8n0]. They cannot all divide N, since p1p2p3 > x3 > N. Any pi∤N is a prime r with A α < r ≤ B and r∤N, that is, r ∈ R N , and r > max Q N , D . ◻
Proposition 7.8 (Application to the restricted Chen sieve). Let p be admissible and N large.
(i) No element of the support of W N high lies in S 3 A 3 α ; in particular none lies in S 3 D . Hence every 3−factorable sequence of level at most A3α vanishes identically on the support of the switched part of the coefficients.
(ii) The high-modulus sequence a N m = W N m 1 m > Q N is not 2−factorable (a fortiori neither well-factorable nor 3−factorable) of level D.
Proof. (i) By Proposition 3.3(iii), every m in the support of W N high has r = P + m ∈ R N and m≥r = r, so r2m≥r = r3 > A3α; and A3α > N > D by Lemma 3.2(a). Apply Theorem 7.3(i). (ii) Let r be as in Lemma 7.7. Since r > z, W N low r = 0 , and W N high r = λ D / r + 1 = 1 ; hence aN(r) = WN(r) = −1. But for m = r and p = r we have p m≥p = r2 > D, so r ∉ S 2 D , and Theorem 7.3(i) shows that aN is not 2−factorable of level D. Since S 3 D ⊆ S 2 D , it is not 3−factorable either. ◻
Remark 7.9. Proposition 7.8 is independent of any distribution estimate: every switched modulus lies outside S 3 N , and every triply factorable sequence of level at most N vanishes there. In particular, increasing the level in a theorem whose coefficient class remains triply factorable would preserve this support restriction. The well-factorable estimates of [BFI86] and the triply well-factorable estimates of [May25, Lic23] provide distinct analytic coefficient classes against which this structural criterion can be tested.

8. Parity-Weighted Cofactor Minorants

8.1. Setting

Definition 8.1. A vector v = (r0,r1,s0,s1,c0,c1,η,θ) is admissible if
0 < r 0 < r 1 < s 0 < s 1 ,   r 1 + 2 s 1 < 1 ,   0 < c 0 < c 1 < 1 ,   0 < θ < 1 ,
and, with κ = θ − r1 − s1,
1 / 3 max 1 − r 0 − s 0 ,   θ − r 0 − s 0 < η < min r 0 ,   1 − r 1 − s 1 ,   κ / 2 .
The principal vector is v 0 = 0.291 , 0.299 , 0.321 , 0.329 , 1 2 , 3 4 , 0.13 , 0.98 . We put
σ − = κ η ,   σ + = θ − r 0 − s 0 η ,   Δ = c 1 − c 0 ,   P 0 = log r 1 r 0 log s 1 s 0 ,   h = log σ − − 1 ,   ρ 0 = 1 − h h ,
g 0 = Δ P 0 h 2 η σ − ,     h 0 = Δ P 0 1 − h η σ − .
Admissibility gives 2 < σ−<σ+<3.
For X ≥ 2 let Rect(X) be the set of pairs of primes (u,v) with X r 0 < u ≤ X r 1 and X s 0 < v ≤ X s 1 ; let aX=⌈c0X⌉, bX=⌊c1X⌋, W uv = w ∈ N :   a X ≤ uvw ≤ b X , y = Xη and Quv = Xθ/(uv). Put
F uv ± w = F Q uv ± w ; P y ,     Π w = 1 / 2 1 − λ L w ,     l uv w = Π w   F uv − w ,
so that Π(w) is the indicator of Ω(w) being odd. With ∑ ∑ denoting summation over (u,v) ∈ Rect(X) and w ∈ Wuv, and q any real function of (u,v,w), define for integers N
T X N = ∑ ∑ l uv w 1 P N − uvw , H X N = ∑ ∑ − l uv w + 1 P N − uvw , C X N = ∑ ∑ 1 P w 1 P N − uvw , T X , q N = ∑ ∑ q u , v , w l uv w 1 P N − uvw , A ± N = ∑ ∑ F uv ± w 1 P N − uvw , L ± N = ∑ ∑ λ L w F uv ± w 1 P N − uvw .
The conditions in (8.1) have the following roles, for large X and w ∈ Wuv: η < 1 − r1 − s1 gives w > y; 3η > 1 − r0 − s0 gives w < y3; 3η > θ − r0 − s0 gives σ+<3; η < κ/2 gives σ−>2; and η < r0 gives y < u. Finally r1 + 2s1 < 1 gives w > X s 1 ≥ v .

8.2. The Minorant and Two Decompositions

Theorem 8.2 (Pointwise minorant). Let v be admissible. For X ≥ X0(v), all (u,v) ∈ Rect(X) and w ∈ Wuv,
0 ≤ 1 P w − l uv w ≤ Π w F uv + w − F uv − w ,     1 P w − l uv w = − l uv w + ,
and l uv w = 1 if and only if w is prime. Consequently TX(N) = CX(N) − HX(N) for every N.
Proof. For w ∈ Wuv we have y < w < y3. If w is prime, then Π(w) = 1 and (w,P(y)) = 1, so l uv w = 1 = 1 P w . If Ω(w) is even, then Π(w) = 0 and w is composite, so l uv w = 0 = 1 P w . If Ω(w) ≥ 3 is odd, then w has a prime factor ≤w1/3 < y, so Corollary 2.3 gives F uv − w ≤ 0 ≤ F uv + w , and 1 P w − l uv w = − F uv − w ∈ 0 , F uv + w − F uv − w . Summing 1 P − l = − l + gives the last identity. ◻
Theorem 8.3 (Exact decomposition and cost bound). For X ≥ X0(v) and every N,
T X N = 1 / 2 A − N − L − N ,     0 ≤ H X N ≤ 1 / 2 A + N − A − N − L + N + L − N .
Proof. Expand l uv = 1 2 1 − λ L F uv − in the definition of TX(N). For the second claim, bound H X N = ∑ ∑ 1 P − l uv w 1 P N − uvw by Theorem 8.2 and expand Π in the same way. ◻
Proposition 8.4 (Progression decomposition). Let X ≥ X0(v) and X < N ≤ 2X. Let Rect(X,N) be the set of (u,v) ∈ Rect(X) with (uv,N) = 1, and put YN,X = li(N − aX) − li(N − bX − 1) and ρN,X(m) = π(N − aX;m,N) − π(N − bX − 1;m,N) − YN,X/φ(m). Then A±(N) = B±(N) + Z±(N), where
B ± N = Y N , X ∑ u , v ∈ Rect X , N 1 φ uv ∑ d ∣ P N y λ Q uv ± d φ d , Z ± N = ∑ u , v ∈ Rect X , N ∑ d ∣ P N y λ Q uv ± d   ρ N , X uvd .
Proof. If N − uvw is a prime p', then p' ≥ N − bX > (1 − c1)X, which exceeds u, v and y for large X. Hence pairs with (uv,N) > 1 contribute nothing, and neither do divisors d∣P(y) with (d,N) > 1 in the expansion F uv ± w = ∑ d ∣ w , P y λ Q uv ± d : in both cases a prime dividing N would divide p'. For (uv,N) = 1 and d∣PN(y), the map w ↦ p' = N − uvw is a bijection from w ∈ W uv :   d ∣ w ,   N − uvw ∈ P onto the primes p'≡N (mod uvd) in [N − bX,N − aX], whose number is YN,X/φ(uvd) + ρN,X(uvd). Finally φ(uvd) = φ(uv)φ(d), because the prime factors of d are at most y < u < v. ◻
Remark 8.5 (Support of the moduli). Every modulus uvd occurring in Z±(N) exceeds X r 0 + s 0 . At v0, this lower bound is X0.612 > (2X)1/2 ≥ N1/2 for sufficiently large X. Since every prime factor of d is smaller than u, we also have u 2 uvd ≥ u = u 3 v > X 3 r 0 + s 0 . At v0, the exponent 3r0 + s0 = 1.194 exceeds θ = 0.98. Corollary 7.4(ii) therefore implies that every 3−factorable sequence of level at most (2X)0.98 vanishes at each of these moduli. Proposition 8.11 supplies mean-value control of Z±(N) for this support by averaging over N.

8.3. The Reweighting Envelope and a Kernel Correction

Theorem 8.6 (Exact reweighting envelope). Let D be a finite set and l : D → R with l i ∈ 1 ∪ − ∞ , 0 for every i. For 0 < a ≤ b ,
min q : D → a , b   ∑ i ∈ D q i l i = a ⋅ # i : l i = 1 − b ∑ i ∈ D − l i + ,
the minimum being attained at q = a on l = 1 and q = b elsewhere. Consequently, for admissible v, X ≥ X0(v) and every N,
min a ≤ q ≤ b T X , q N = a   C X N − b   H X N = a   T X N − b − a   H X N .
Proof. On l = 1 we have q i l i = q i ≥ a , and on l ≤ 0 we have q i l i = − q i − l i ≥ − b − l i , with equality in the stated cases. For the application take D = u , v , w : N − uvw ∈ P and l = l uv w ; by Theorem 8.2, # l = 1 = C X N and TX = CX − HX. ◻
At a = 0 , the envelope equals − b H X N . A positive lower constraint a > 0 retains the contribution a C X N and provides the range relevant to positive lower bounds.
Theorem 8.7 (Kernel correction). Let v be admissible and let β',θ' ∈ (0,1) satisfy 1 − r0 − s0 < β' and 2(1 − r0 − s0) < θ' < 4r0. For X < N ≤ 2X put zN = Nβ', DN = Nθ', I N n = 1 n , P z N = 1 and K N R n = F D N − n ; P z N . For q : Rect X × N → 0 , 1 let
K ~ N , q n = K N R n + 2 ∑ u , v ∈ Rect X ,   w ∈ W uv uvw = n q u , v , w   l uv w .
Then for X ≥ X0(v,β',θ'), every X < N ≤ 2X and every n ≥ 1, K ~ N , q n ≤ I N n , and ∑ a X ≤ n ≤ b X K ~ N , q n − K N R n 1 P N − n = 2   T X , q N . At v0 one may take β ' , θ ' = 9 20 , 49 50 .
Proof. The last identity holds by definition, since uvw ∈ [aX,bX] for w ∈ Wuv. If n has no representation n = uvw as above with w prime, all correction terms are ≤0 by Theorem 8.2 and q ≥ 0, so K ~ N , q n ≤ K N R n ≤ I N n by Corollary 2.3. Otherwise n = uvw with w prime, and u < v < w because v > X s 0 > X r 1 ≥ u and w > X s 1 ≥ v . Thus n has exactly three prime factors; since the windows of u and v are disjoint and w lies in neither, the representation is unique, and the correction equals 2q(u,v,w) ≤ 2. All three primes are below X 1 − r 0 − s 0 < z N , so IN(n) = 0. Moreover w 2 < X 2 1 − r 0 − s 0 < D N and u 4 > X 4 r 0 > 2 X θ ' ≥ D N for large X, so Lemma 2.4 with k = 3 and m = 2 gives K N R n = − 2 . Hence K ~ N , q n ≤ 0 = I N n . At v0: 1 − r0 − s0 = 0.388 < 0.45 and 0.776 < 0.98 < 1.164. ◻
The pointwise kernel inequality uses 0 ≤ q ≤ 1. For the summed statistic, Theorem 8.6 expresses the balance between a C X N and b H X N when 0 < a ≤ q ≤ b .

8.4. Main Terms

Lemma 8.8 (Density ratio). Let F(s) = 2eγ/s and f(s) = 2eγlog(s − 1)/s on (2,3], where both are the linear-sieve functions. Then f is increasing, F − f is decreasing, and (F(s) − f(s))/f(s) = (1 − log(s − 1))/log(s − 1) is decreasing; in particular F(s) − f(s) ≤ ρ0f(s) for σ−≤s ≤ 3. If σ − > 1 + e , then ρ0 < 1.
Proof. sf(s)/(2eγ) = log(s − 1) and f'(s) has the sign of s s − 1 − log s − 1 > 0 on (2,3]. Similarly (F − f)(s) = 2eγ(1 − log(s − 1))/s has negative derivative there. The ratio is 1/t − 1 with t = log(s − 1) increasing, and its value at s = σ− is ρ0. If σ − − 1 > e then h > 1 2 and ρ0 = 1/h − 1 < 1. ◻
Proposition 8.9 (Main term of the envelope). Let v be admissible and 0 < a ≤ b with a ≥ b − a ρ 0 , and put
G a , b = a − b − a ρ 0 g 0 = Δ P 0 2 κ b   h − b − a .
For every ε > 0 there is X1 such that for X ≥ X1 and every even N ∈ (X,2X],
a B − N − b − a B + N − B − N   ≥   4 G a , b − ε   S N X log 2 X ,
and consequently
min a ≤ q ≤ b T X , q N   ≥   2 G a , b − ε   S N X log 2 X − b 2 Z + N + Z − N + L + N + L − N .
Moreover, in the same normalisation, the main term of TX(N) is at least 2g0 − ε, and that of the cost 1 2 A + − A − is at most h0 + ε.
Proof. Write Σ uv ± = ∑ d ∣ P N y λ Q uv ± d / φ d and apply Theorem 2.5 to gN, with sifting limit ⌊y⌋+1 and level Quv; condition (2.1) holds with K = log3 by Theorem E. The sieve parameter is suv + o(1) with suv = (θlogX − loguv)/(ηlogX) ∈ [σ-,σ+)⊂(2,3), and logQuv ≥ κlogX. Hence, uniformly in (u,v), Σ uv − ≥ V N y f s uv − ε 1 and Σ uv + ≤ V N y F s uv + ε 1 for large X. By Lemma 8.8, and since a − b − a ρ 0 ≥ 0 ,
  a Σ uv − − b − a Σ uv + − Σ uv − ≥ V N y a − b − a ρ 0 f s uv − 2 b ε 1 ≥ V N y a − b − a ρ 0 f σ − − 2 b ε 1 .
Multiply by YN,X/φ(uv) and sum over Rect(X,N). By Lemma A.4, ∑ Rect X , N 1 / φ uv = P 0 + o 1 ; further YN,X = (1 + o(1))ΔX/logX, V N y = 1 + o 1 e − γ S N / η log X by Corollary 2.7, and f(σ-)=2eγh/σ-. The product of the main factors is
Δ X log X ⋅ e − γ S N η log X ⋅ a − b − a ρ 0 2 e γ h σ − ⋅ P 0 = 4 G a , b   S N X log 2 X ,
which proves the first inequality, using ησ−=κ. For the second, Theorems 8.3 and 8.6 and Proposition 8.4 give
min a ≤ q ≤ b T X , q = a T X − b − a H X ≥ 1 / 2 a B − − b − a B + − B − + 1 / 2 b Z − − b − a Z + + 1 / 2 b − a L + − b L − .
The last assertion follows in the same way from B - ≥ Y N , X V N y f σ - − ε 1 ∑ 1 / φ uv and B + − B - ≤ Y N , X V N y F − f σ - + 2 ε 1 ∑ 1 / φ uv , since F − f is decreasing. ◻
Lemma 8.10 (Davenport’s estimate for the Liouville function). Let e(x) = e2πix. For every A > 0, uniformly for real ξ, K ≥ 2 and every interval I ⊆ [0,K],
∑ k ∈ I λ L k   e k ξ ≪ A K log K A .
Proof. Davenport [Dav37] proved that ∑ n ≤ x μ n e n ξ ≪ A x log x − A uniformly in ξ. Since λ L n = ∑ a 2 b = n μ b , we have ∑ n ≤ x λ L n e n ξ = ∑ a ≤ x ∑ b ≤ x / a 2 μ b e a 2 b ξ . The terms with a ≤ x1/4 contribute ≪ A ∑ a x / a 2 log x − A , because log x / a 2 ≥ 1 2 log x there, and those with a > x1/4 contribute ≪ ∑ a > x 1 / 4 x / a 2 ≪ x 3 / 4 . The claim follows on taking differences. ◻
Proposition 8.11 (Mean-value bounds for the remainders). Let v be admissible and A > 0. Then
∑ X < N ≤ 2 X Z + N + Z − N ≪ A , v X 2 log X A ,
∑ X < N ≤ 2 X L + N 2 + L − N 2 ≪ A , v X 3 log X A .
Proof. The sums Z±. If (u,v) ∈ Rect(X), d∣P(y) and λ Q uv ± d ≠ 0 , then d = 1, or d is a prime ≤y < Quv, or d < Quv; in all cases uvd < Xθ. The prime factors of d are at most y < X r 0 , so the prime factors of m = uvd exceeding y are exactly u and v, one in each window; hence m determines (u,v,d). Therefore Z ± N = ∑ m < X θ ,   m , N = 1 c ± m ρ N , X m with |c±(m)| ≤ 1, and ρN,X(m) is the quantity ρ N ∘ m of Lemma 5.8 with uN = N − bX and vN = N − aX, which satisfy its hypotheses with C = 1 − c 1 − 1 . Lemma 5.8 with θ0 = θ gives the first bound.
The sums L±. For n ∈ Z let a ± n be the sum of λ L w F uv ± w over (u,v) ∈ Rect(X) and w ∈ Wuv with uvw = n, so that L ± N = ∑ n a ± n 1 P N − n , where for X < N ≤ 2X only primes N − n ≤ 2X occur. By Parseval’s identity,
∑ X < N ≤ 2 X L ± N 2 ≤ ∫ 0 1 ∑ n a ± n e n ξ 2 ∑ p ≤ 2 X e p ξ 2 d ξ ≤ π 2 X sup ξ ∑ n a ± n e n ξ 2 .
Expanding F uv ± w = ∑ d ∣ w , P y λ Q uv ± d , writing w = dk and using λL(dk) = λL(d)λL(k), we get
∑ n a ± n e n ξ = ∑ u , v ∈ Rect X ∑ d ∣ P y λ Q uv ± d λ L d ∑ k ∈ I uvd λ L k   e kuvd ξ ,     I uvd = a X uvd , b X uvd .
As above only d with uvd < Xθ occur, so bX/(uvd) ≥ (c1/2)X1−θ for sufficiently large X, and Lemma 8.10 bounds the inner sum by ≪A(X/uvd)(logX)−A uniformly in ξ. Summing over d < Xθ and then over (u,v), using ∑ u , v ∈ Rect X 1 / uv ≪ 1 , the exponential sum is ≪AX(logX)1−A uniformly in ξ, and the second bound follows. ◻
Theorem 8.12 (Almost all N). Let v be admissible and let 0 < a ≤ b satisfy a ≥ b − a ρ 0 and G a , b > 0 . For every A > 0 and ε > 0, all but O A , ε , v , a , b X log X − A even N ∈ (X,2X] satisfy
min a ≤ q ≤ b T X , q N   ≥   2 G a , b − ε S N X log 2 X .
In particular, for all these N there are primes p,u,v,w with N = p + uvw, X r 0 < u ≤ X r 1 and X s 0 < v ≤ X s 1 . At v0 with a / b = 1 2 one may take 2 G a , b − ε ≥ 1.5 ⋅ 10 − 5   b .
Proof. It suffices to establish the assertion for 0 < ε < G a , b : replacing an arbitrary ε > 0 by min ε , G a , b / 2 strengthens the lower bound and preserves the stated exceptional-set order. By Proposition 8.9 with ε/2 in place of ε, an even N ∈ (X,2X] violating the inequality satisfies b 2 Z + + Z - + L + + L - N > ε 2 S N X / log 2 X ≥ ε C 2 X / log 2 X , so |Z+(N)| + |Z−(N)| or |L+(N)| + |L−(N)| exceeds ε C 2 X / b log 2 X . By Proposition 8.11 and Chebyshev’s inequality, the number of such N is
≪ ε log 2 X X ⋅ X 2 log X A + 2 + log 4 X X 2 ⋅ X 3 log X A + 4 ≪ X log X A .
For the second assertion take q ≡ a : then a T X N > 0 , so CX(N) ≥ TX(N) > 0 by Theorem 8.2. At v0 with a = b / 2 we have 2 G a , b = b 1 − ρ 0 g 0 , and Proposition 8.14 gives 1 − ρ 0 = 2 − 1 / h > 70798 546503 and g 0 > 1 8448 , whence 2 G a , b > 1.53 ⋅ 10 − 5   b . ◻
The transfer identities in Section 5.4 give a one-sided framework for pointwise refinements of these mean-value conclusions.

8.5. The Gain Function and Explicit Parameter Bounds

Proposition 8.13 (Dependence on the roughness parameter). Fix (r0,r1,s0,s1,c0,c1,θ) with 0 < r0 < r1 < s0 < s1, r1 + 2s1 < 1, 0 < c0 < c1 < 1, 0 < θ < 1, let κ = θ − r1 − s1 and 0 < a ≤ b . Then:
(i) the vector is admissible exactly for η in the interval (8.1);
(ii) on this interval G a , b η = Δ P 0 2 κ b log κ / η − 1 − b − a ;
(iii) G ' a , b η = − Δ P 0   b / 2 η κ − η < 0 ;
(iv) G a , b η > 0 if and only if η < κ / 1 + e 1 − a / b .
Proof. (i) is the definition. For (ii), substitute σ-=κ/η into a − b − a ρ 0 g 0 = Δ P 0 2 η σ − b h − b − a . For (iii), d d η log κ / η − 1 = − κ / η κ − η , and 0 < η < κ. (iv) is log κ / η − 1 > 1 − a / b . ◻
Thus, at fixed outer geometry, decreasing η within the admissible range always increases the gain.
Proposition 8.14 (Explicit bounds at v0). The vector v0 is admissible, with σ − = 176 65 and σ + = 184 65 ; at its outer geometry the admissible range of η is 97 750 , 22 125 . Moreover
1 1600 < 64 299 ⋅ 329 ≤ P 0 ≤ 64 291 ⋅ 321 < 1 1400 ,     8 15 < 546503 1022208 < h ,
ρ 0 < 7 8 ,     g 0 > 1 8448 ,     h 0 < 1 4224 ,
and 0.13 < κ / 1 + e = 0.13289 … , so that G a , b 0.13 > 0 for a / b = 1 2 . The value η = 0.14 is admissible but gives ρ0 > 1.
Proof. Here κ = 0.352 = 44 125 , and admissibility, σ± and the η−range are rational arithmetic: the range is 0.388 3 , min 0.291 , 0.372 , 0.176 . Using x/(1 + x) < log(1 + x) < x with x = 8 291 and x = 8 321 gives the bounds for P 0 = log 299 291 log 329 321 , and 299 ⋅ 329 = 98371 < 102400 = 64 ⋅ 1600, 291 ⋅ 321 = 93411 > 89600 = 64 ⋅ 1400. For h = log 111 65 write 111 65 = 1 + x 1 − x with x = 23 88 and use log 1 + x 1 − x > 2 x + x 3 / 3 = 546503 1022208 . Then ρ 0 = 1 / h − 1 < 475705 546503 < 7 8 , and g 0 = Δ P 0 h / 2 κ > 1 4 ⋅ 1 1600 ⋅ 8 15 ⋅ 125 88 = 1 8448 , h 0 = Δ P 0 1 − h / κ < 1 4 ⋅ 1 1400 ⋅ 7 15 ⋅ 125 44 = 1 4224 . The inequality 0.13 < κ / 1 + e is equivalent to e < 111 65 2 . For η = 0.14 one has σ − − 1 = 53 35 and 53 35 2 < e , so h < 1 2 and ρ0 > 1. ◻
At v0 one finds g0 = 1.2687…⋅10−4 and, for a / b = 1 2 , G a , b = 8.33 … ⋅ 10 − 6   b .

9. Further Questions

The restricted Chen sieve admits an explicit main-term lower bound, exact identities for its losses and remainders, and unconditional mean-value control. The coefficient C p measures the quantitative effect of the restricted switching range, while Theorem E fixes the optimal uniform dimension constant. Theorems C and D describe two distinct structural features: negative kernel values of unbounded magnitude on SN, and the exclusion of switched moduli from universal factorable supports. The first concerns the divisor sum of the coefficients; the second concerns the moduli at which a factorable sequence can be nonzero.
These structural statements accompany a precise analytic transfer. A one-sided estimate E N high ≥ − κ R N with κ < C p yields the lower bound in Theorem B, whereas averaging gives the unconditional conclusion of Theorem F. The parity-weighted construction has an analogous decomposition, with an exact reweighting envelope and a gain function whose positive range is explicitly determined. Together, these results suggest the following questions.
(1) Can the exceptional set in Theorem F be bounded by O(X1−δ) for some δ>0? A power-saving mean-value estimate for the particular coefficients WN in Proposition 5.9 would provide such a refinement. The exceptional-set estimates for the binary Goldbach problem [MV75] give a natural point of comparison.
(2) What is the true order of the minimum of KN over SN? By Theorem C, it lies between −exp(ClogN/loglogN) and, for infinitely many N, − exp c log N .
(3) For a lower linear sieve with level exponent ρ, positivity of the factor f(ρ/β) requires β < ρ/2. This relation motivates a study of smaller sifting limits together with additional weights for integers having at least three prime factors. Within such Chen-type constructions, how large can β be while obtaining, for every sufficiently large even N, a representation N = p + n in which n is prime or the product of two primes exceeding Nβ?
(4) Theorem D characterizes the union of admissible supports and realizes each admissible modulus by a factorable point mass. Which conditions on the values of a nonnegative sequence supported in S k Q characterize the existence of a k−factorable decomposition for every factorization of the level Q?

Funding

This research was supported by the initial scientific and technological project of the Institute of Information Technology (IOIT), Vietnam Academy of Science and Technology (VAST), “Training a large language model integrating formal logical thinking for solving general mathematics problems,” grant number NVKN02.05/26-27.

Data Availability Statement

Data sharing is not applicable to this article, as no data sets were generated or analysed.

Conflicts of Interest

The author declares no competing interests.

Appendix A. Elementary Explicit Forms of Mertens’ Estimates

We record explicit prime-sum estimates leading to the rectangle-sum formula in Lemma A.4, which is used in Section 5 and Section 8.
Lemma A.1. For every integer N ≥ 2, log N − 6 ≤ ∑ p ≤ N log p p ≤ log N + log 4 .
Proof. By Legendre’s formula, log N ! = ∑ p ≤ N log p ∑ j ≥ 1 ⌊ N / p j ⌋ , and N / p − 1 < ∑ j ≥ 1 ⌊ N / p j ⌋ ≤ N / p − 1 . The lower bound gives N ∑ p ≤ N log p p < log N ! + ϑ N ≤ N log N + N log 4 , using ϑ N = ∑ p ≤ N log p < N log 4 . The upper bound gives log N ! ≤ N ∑ p ≤ N log p p + N ∑ m ≥ 2 log m m m − 1 , and since log m ≤ 2 m and t ↦ t−1/2(t − 1)−1 is decreasing, ∑ m ≥ 2 log m m m − 1 ≤ 2 1 2 + ∫ 2 ∞ dt t   t − 1 = 2 1 2 + log 2 + 1 2 − 1 < 5 . Combined with logN!≥NlogN − N this gives the lower bound. ◻
Lemma A.2. For 2 ≤ y ≤ x, ∑ y < p ≤ x 1 p − log log x log y ≤ 14 log y .
Proof. Let A t = ∑ p ≤ t log p / p and R t = A t − log t . For t ≥ 2 and m=⌊t⌋ we have A t = A m and logm ≤ logt < logm + log2, so Lemma A.1 gives R t < 7 . By partial summation,
∑ y < p ≤ x 1 p = ∫ y x d A t log t = log log x log y + R x log x − R y log y + ∫ y x R t   dt t log 2 t ,
and the last three terms are bounded in absolute value by 7 log x + 7 log y + 7 1 log y − 1 log x = 14 log y . ◻
Lemma A.3. For 2 ≤ y ≤ x, ∑ y < p ≤ x 1 p − 1 − log log x log y ≤ 14 log y + 2 y .
Proof.  1 p − 1 − 1 p = 1 p p − 1 and ∑ m > y 1 m m − 1 = 1 ⌊ y ⌋ ≤ 2 y . ◻
Lemma A.4 (Rectangle sums). Let C ≥ 1 and fix 0 < a0 < a1 ≤ b0 < b1. For X ≥ 2 and every integer M with 1 ≤ M ≤ XC,
∑ X a 0 < u ≤ X a 1 ,   X b 0 < v ≤ X b 1 uv , M = 1 1 φ uv = log a 1 a 0   log b 1 b 0 + O 1 log X ,
where the sum is over primes u,v and the implied constant depends only on C,a0,a1,b0,b1.
Proof. The two windows are disjoint, so u ≠ v and φ(uv) = (u − 1)(v − 1); the sum factorises. By Lemma A.3, ∑ X a 0 < u ≤ X a 1 1 u − 1 = log a 1 a 0 + O 1 a 0 log X . At most C/a0 primes u > X a 0 divide M, and each contributes at most 2 X − a 0 ; the same holds for v. Multiplying the two factors gives the claim. ◻
In Proposition 5.6 the lemma is applied with X = N and M = N, and in Proposition 8.9 with M = N ≤ 2X ≤ X2.

References

  1. [BFI86] E. Bombieri, J. B. Friedlander and H. Iwaniec, Primes in arithmetic progressions to large moduli. Acta Math. 1986, 156, 203–251. [CrossRef]
  2. [Bom65] E. Bombieri, On the large sieve. Mathematika 1965, 12, 201–225. [CrossRef]
  3. [Bom76] E. Bombieri, The asymptotic sieve. Rend. Accad. Naz. XL 1975/76, 1/2((5)), 243–269.
  4. [Che73] J.-R. Chen, On the representation of a larger even integer as the sum of a prime and the product of at most two primes. Sci. Sinica 1973, 16, 157–176.
  5. [Dav00]; Davenport, H. Multiplicative Number Theory, 3rd ed.; Montgomery, H. L., Ed.; Graduate Texts in Mathematics 74; Springer: New York, 2000. [Google Scholar]
  6. [Dav37] H. Davenport, On some infinite series involving arithmetical functions (II). Quart. J. Math. Oxford Ser. 1937, 8, 313–320. [CrossRef]
  7. [DH66] H. Davenport and H. Halberstam, Primes in arithmetic progressions. Michigan Math. J. 1966, 13, 485–489. [CrossRef]
  8. [EH70] P. D. T. A. Elliott and H. Halberstam, A conjecture in prime number theory. In Symposia Mathematica, (INDAM, Rome, 1968/69); Academic Press: London, 1970; Volume IV, pp. 59–72.
  9. [FI10]; Friedlander, J.; Iwaniec, H. Opera de Cribro; American Mathematical Society Colloquium Publications: Providence, RI, 2010; Volume 57. [Google Scholar] [CrossRef] [PubMed]
  10. [HR74] H. Halberstam and H.-E. Richert, Sieve Methods. In London Mathematical Society Monographs; Academic Press: London, 1974; Volume 4.
  11. [Iw80a] H. Iwaniec, Rosser’s sieve. Acta Arith. 1980, 36, 171–202. [CrossRef]
  12. [Iw80b] H. Iwaniec, A new form of the error term in the linear sieve. Acta Arith. 1980, 37, 307–320. [CrossRef]
  13. [KT05] A. Kumchev and D. Tolev, An invitation to additive prime number theory. Serdica Math. J. 2005, 31, 1–74. [CrossRef]
  14. [Li26] R. Li, On Chen’s theorem, Goldbach’s conjecture and almost prime twins II. Math. Reports 2026, arXiv:2405.0572728((78) nos. 1–2), 39–61. [CrossRef]
  15. [Lic23] J. D. Lichtman, Primes in arithmetic progressions to large moduli, and Goldbach beyond the square-root barrier (with an appendix by S. Drappeau). arXiv 2023, arXiv:2309.08522.
  16. [May25] J. Maynard, Primes in arithmetic progressions to large moduli II: Well-factorable estimates. Mem. Amer. Math. Soc. 2025, arXiv:2006.07088306, no. 1543. [CrossRef]
  17. [Mer74] F. Mertens, Ein Beitrag zur analytischen Zahlentheorie. J. Reine Angew. Math. 1874, 78, 46–62. [CrossRef]
  18. [MV75] H. L. Montgomery and R. C. Vaughan, The exceptional set in Goldbach’s problem. Acta Arith. 1975, 27, 353–370. [CrossRef]
  19. [Pol14] D. H. J. Polymath, New equidistribution estimates of Zhang type. Algebra Number Theory 2014, 8, 2067–2199. [CrossRef]
  20. [RS62]; Rosser, J. B.; Schoenfeld, L. Approximate formulas for some functions of prime numbers. Illinois J. Math. 1962, 6, 64–94. [Google Scholar] [CrossRef]
  21. [Sai97] E. Saias, Entiers à diviseurs denses 1. J. Number Theory 1997, 62, 163–191. [CrossRef]
  22. [Sel49] A. Selberg, On elementary methods in prime number theory and their limitations. In Den 11te Skandinaviske Matematikerkongress (Trondheim, 1949); Johan Grundt Tanums Forlag: Oslo, 1952; pp. 13–22.
  23. Sarajian, G.; Weingartner, A. An extension of smooth numbers: multiple dense divisibility. J. Number Theory 2026, arXiv:2504.02699280, 278–317. [Google Scholar] [CrossRef]
  24. [Ten86] G. Tenenbaum, Sur un problème de crible et ses applications. Ann. Sci. École Norm. Sup. (4) 1986, 19, 1–30. [CrossRef]
  25. [TW24] G. Tenenbaum and A. Weingartner, An Erdős–Kac theorem for integers with dense divisors. Q. J. Math. 2024, arXiv:2211.0581975(no. 1), 161–195.
  26. [Vin65] A. I. Vinogradov, The density hypothesis for Dirichlet L−series. Izv. Akad. Nauk SSSR Ser. Mat. (in Russian). 1965, 29, 903–934.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.