Preprint
Article

This version is not peer-reviewed.

Uniformity, Local Structure and Boundary Correlations in Goldbach Representations

Submitted:

28 October 2026

Posted:

30 October 2026

You are already at the latest version

Abstract
We study uniformity in Goldbach representations and the signed correlations arising in a reflected-prime comparison. From classical analytic inputs we derive a common exceptional set for all compatible residue prescriptions modulo sufficiently small fixed powers of the location parameter. We also prove weighted Type I distribution, convex factor regularity, a mixed prime–rough estimate approaching the square-root scale, a balanced rough-pair asymptotic and an uncentered two-prime upper bound. For a sharp bilinear boundary component, we compute the exact unit mean and coefficient moments, remove a repeated-prime contribution with its variance cross terms, and obtain a restricted Möbius product identity. A lower bound for the rowwise absolute character majorant shows why that particular relaxation cannot deliver arbitrary logarithmic savings. Multiplicative prime-factor markers and a rational-target CRT kernel clarify the remaining distribution requirements. Additional results treat almost-all prime-product representations, a conditional implication for real zeros, and finite paired-congruence criteria. All exceptional sets and unproved hypotheses are retained. The results do not establish the required full signed correlation estimate or resolve the binary Goldbach conjecture.
Keywords: 
;  ;  ;  ;  ;  

MSC:  11P32; 11N13; 11N36; secondary 11L05; 11N37

1. Introduction and Statements

The circle method separates the binary Goldbach problem into a local arithmetic main term and an analytic remainder. In an almost-all theorem, a mean-square estimate can absorb the minor arcs. Prescribing the residue classes of both primes introduces a second uniformity question: whether a single exceptional set can accommodate all the prescriptions under consideration. If the allowed moduli grow with the location of the primes, the cost of taking this common set must be included in the error budget.
Montgomery and Vaughan [8] developed an effective treatment of Goldbach exceptional sets which retains a possible exceptional real character. Kawada [6] studied prime tuples in arithmetic progressions and, in particular, already treated the geometry in which both primes lie in the same short interval. Halupczok [5] obtained related distribution results with short prime intervals and arithmetic progressions. Thus the same-interval geometry itself is classical. We give the precise comparison with Kawada’s norm in Section 6; that norm implies a common relative asymptotic for polylogarithmic moduli, whereas it does not by itself imply the power-saving simultaneous conclusion formulated below.
The first part of this paper gives an explicit derivation of the latter conclusion. The subsequent parts examine what is still required to pass from averaged representations and rough comparisons to a signed estimate at the balanced scale. The argument keeps the exceptional character fixed while refining the requested modulus, proves a finite arithmetic norm for the prescribed residue fibres, and distinguishes the location T from the short interval length Y . It gives a lower bound, rather than a relative asymptotic, uniformly in the possible exceptional-character signs. The underlying analytic inputs are classical. We make no assertion of an optimal exponent or of an exhaustive priority comparison with the literature on exceptional sets.

Organization and Status of the Additional Results

The principal short-interval theorem is proved in Section 2, Section 3, Section 4 and Section 5; its classical comparison and prime-power normalization appear in Section 6 and Section 7. The appendices make the other obtained results part of the same mathematical framework. Theorems B.1, B.2 and B.4 concern the actual rough comparison, while Theorem C.2 is a quantitative subendpoint consequence of maximal Bombieri–Vinogradov. Theorems D.2 and D.4 provide, respectively, a rough benchmark and a two-prime upper bound at the balanced scale. Propositions D.5 and E.3 are sufficient conditions with unproved arithmetic hypotheses, rather than assertions that these hypotheses hold. Theorem F.2 and Corollary F.3 retain an exceptional set. Theorem G.2 assumes an external positive-density band for the actual Goldbach convolution. Appendix H proves finite congruence criteria. Appendix I, Appendix J, Appendix K and Appendix L study an explicitly defined sharp boundary component: its coefficient moments and exact mean, a repeated-prime error, the resulting Möbius product weight, a failed absolute-value method, and a restricted rational-target kernel estimate. These different logical statuses are essential to the interpretation of the results. Proofs below are complete deductions from the explicitly stated classical analytic inputs; they do not reprove the general analytic machinery of the cited sources. No exhaustive novelty or priority certification is claimed for these deductions.
Put e ( u ) = e x p ( 2 π i u ) . We use φ , μ , Λ , τ ( n ) and ω ( n ) for Euler’s totient, the Möbius function, the von Mangoldt function, the divisor function and the number of distinct prime divisors, respectively. Parameters in subscripts on O and ≪ are fixed before the asymptotic limit. Unless stated otherwise, implied constants are independent of the residue classes. Pairs of primes in this paper are ordered, and equal primes are permitted. For 0 < τ < 1 define
I τ ( T , Y ) = [ 2 T + τ Y ,   2 T + ( 2 − τ ) Y ] ∩ 2 Z .
For reduced classes a , b   ( m o d   m ) , let
r ( H ; m , a , b ; T , Y ) = # { ( p , r ) : p + r = H ,   T < p , r ≤ T + Y ,   p ≡ a ,   r ≡ b   ( m o d   m ) } .
Theorem 1.1. Fix 0 < τ < 1 . There exist positive constants γ , α , β , η , c τ , C τ and T τ such that, for every real T ≥ T τ and Y = T 1 − γ , there is a set E τ ( T ) ⊂ I τ ( T , Y ) with the following properties:
| E τ ( T ) | ≤ C τ Y T − α − β ,
# { H ∈ E τ ( T ) : H ≡ c   ( m o d   m ) } ≤ C τ Y m T − β ( 2 ∣ m ,   2 ≤ m ≤ T α ) .
For every such m , every ordered pair of units a , b   ( m o d   m ) , and every H ∈ I τ ( T , Y ) \ E τ ( T ) with H ≡ a + b   ( m o d   m ) , one has
r ( H ; m , a , b ; T , Y ) ≥ c τ Y T − η m ( l o g T ) 4 .
One may take
t = s 2 ,     α = s 4 = t 2 ,     γ = t / 100 , β = t / 8 ,     η = 2 t ,
where s > 0 is sufficiently small in terms of τ and the absolute constants in the classical input estimates.
The set in the theorem is common to all the requested moduli and residue pairs at a fixed anchor T . It is not one set for all real anchors. The representing pair can change with the prescription. Since 1 − γ − η − α > 0 , the lower bound tends to infinity uniformly over the allowed moduli. Also | p − H / 2 | , | r − H / 2 | ≤ Y / 2 . Neither these observations nor (1.2) removes the exceptional set.
The finite kernel estimate in Section 3 is the arithmetic step used to preserve both small factors when the exceptional character suppresses both residue classes. Section 2, Section 3, Section 4 and Section 5 give the proof of Theorem 1.1. Section 6 records the classical comparison, and Section 7 gives a normalization result for proper prime powers. Section 8 identifies the separate prime-correlation estimate which remains open in a reflected-prime comparison approach. Appendix A proves a classical mixed prime–rough average for every fixed row exponent below 1 / 2 , while keeping the balanced endpoint unresolved.

2. Classical Analytic Inputs and Interval Localization

We use the following form of Montgomery–Vaughan [8, Lemmas 4.1–4.3]. There are absolute positive constants c 2 , c 4 , c , C with the stated properties. Let
e x p l o g N ≤ R 1 ≤ N c 4 .
At level R 1 select the possible exceptional primitive real nonprincipal character χ , of conductor f ≥ 3 , and its real zero 1 − ε , according to the source convention. If it is selected,
ε ≥ c 2 f ( l o g f ) 2 .    
For each primitive character ψ of conductor at most R 1 define the finitely supported coefficients, for 1 ≤ n ≤ N , by
Δ ψ ( n ) = ψ ( n ) l o g n   1 n   p r i m e − 1 ψ = 1 + 1 ψ = χ n − ε .
The last term is absent if no exceptional character is selected; 1 denotes the primitive principal character of conductor one. Set
B ψ = s u p 0 < u < v ≤ N | ∑ u < n ≤ v Δ ψ ( n ) | v − u + N / R 1 ,     G = ∑ ψ   p r i m i t i v e c o n d ψ ≤ R 1 B ψ .
The supremum is inside the primitive-character sum. The source estimate gives
G ≤ C e x p − c l o g N l o g R 1 1 , no   exceptional   character , min 1 , ε l o g R 1 , otherwise .
In particular the exceptional-zero factor multiplies the whole aggregate. Half-open intervals follow from the source’s endpoint convention by limits, without adding a logarithmic endpoint error. We cite this analytic theorem, including its zero-density input; its proof is not part of the present argument.
We also use Gallagher’s finite-frequency inequality [3] in the form of [8, Lemma 4.2]. If h = ( 2 A ) − 1 and ( u n ) has finite support, then
∫ − A A | ∑ n u n e n z | 2   d z ≪ h − 2 ∫ R | ∑ v < n ≤ v + h u n | 2   d v .  
The version for complex coefficients follows by applying the real version to their real and imaginary parts.
Lemma 2.1. Let N = ⌈ 2 T ⌉ , R 1 = T s , and Y = T 1 − γ , where 0 < γ < s and Y ≤ T . Put E ψ ( z ) = ∑ T < n ≤ T + Y Δ ψ ( n ) e ( n z ) . Then
∫ z ≤ R 1 T | E ψ z | 2   d z ≪ B ψ 2 Y ,     E ψ 0 ≪ B ψ Y .       
Proof. Take h = T / ( 2 R 1 ) in (2.4), after truncating the coefficients to ( T , T + Y ] . Each sliding interval intersects that box in an interval of length at most h , and hence has corrected sum at most B ψ ( h + N / R 1 ) ≪ B ψ h . The set of sliding positions has length at most Y + h . Thus the integral is ≪ B ψ 2 ( Y + h ) ≪ B ψ 2 Y , since γ < s . The zero-frequency assertion follows directly from the interval norm. All endpoints here remain bounded by N , the location-scale parameter. ▫
Our final analytic input is Vaughan’s estimate [9, Theorem 17.1]: if ( a , q ) = 1 and | θ − a / q | ≤ q − 2 , then
∑ n ≤ Z Λ ( n ) e ( n θ ) ≪ ( Z q − 1 2 + Z 4 5 + ( Z q ) 1 2 ) ( l o g Z ) 5 2 .
We apply it to two prefixes of size comparable to T . Their difference does not replace the factor T by Y .

3. A Character-Kernel Norm on Prescribed Residue Fibres

Let M be even, A , B units modulo M , and H > 0 with H ≡ A + B   ( m o d   M ) . Write c d ( h ) for the Ramanujan sum. Define
S M H = ∏ p ∤ M H 1 − 1 ( p − 1 ) 2 ∏ p ∣ H p ∤ M p p − 1 .
All primes in these products are odd, and S M ( H ) ≥ C 2 = ∏ p > 2 ( 1 − ( p − 1 ) − 2 ) > 0 .
For a primitive character ψ whose conductor divides M d , put
K ψ A ( c ; M d ) = 1 φ ( M d ) ∑ v   m o d   M d ( v , M d ) = 1 v ≡ A   ( M ) ψ ( v ) ¯ e c v M d .
Set this coefficient to zero if the conductor does not divide M d . For two primitive characters define
Γ d H ; ψ , ω = ∑ c   m o d   M d c , d = 1 K ψ A c ; M d K ω B c ; M d e − H c M d
The second coefficient is not conjugated: this kernel belongs to an additive convolution.
Proposition 3.1. Uniformly in the above data and in both primitive characters,
∑ d ≥ 1 | Γ d ( H ; ψ , ω ) | ≪ M φ ( M ) 2 S M ( H ) .  
Proof. Writing c = j d + u and first summing over j   ( m o d   M ) gives
Γ d = M φ ( M d ) 2 ∑ v , w   m o d   M d   u n i t s v ≡ A ,   w ≡ B   ( M ) ψ ( v ) ω ( w ) ¯   c d v + w − H M .  
The quotient is integral and is defined modulo d . The Chinese remainder theorem factors the original finite sum into local kernels; the additive CRT multipliers are units and can be absorbed into the local summation variables.
Put a = v p ( M ) , b = v p ( d ) , and let t 1 , t 2 be the local conductor exponents. Suppose first that a ≥ 1 . For b = 0 a nonzero factor requires t 1 , t 2 ≤ a ; the residue fibres are singletons and its modulus is p a / φ ( p a ) 2 . For b > 0 , let n = a + b . A fibre Fourier coefficient vanishes if its character has conductor exponent less than n : summation over the last p lifts cancels its nontrivial additive phase. Thus a nonzero product requires t 1 = t 2 = n . Expanding c p n ( z ) = p n 1 p n ∣ z − p n − 1 1 p n − 1 ∣ z in the local version of (3.2), the second term vanishes by primitive last-lift cancellation. The remaining factor is
p n φ ( p n ) 2 ∑ v   m o d   p n v ≡ A   ( p a ) ψ p ( v ) ω p ( H − v ) ¯ .
Both arguments are units, and there are p b terms. Its modulus is at most p n p b / φ ( p n ) 2 = p a / φ ( p a ) 2 . For fixed t 1 , t 2 only one value of b can contribute, including the case b = 0 . This argument also applies to p = 2 .
Now suppose a = 0 , so p is odd. If both local characters are principal, the only supported exponents are b = 0 , 1 , with factors 1 and c p ( H ) / ( p − 1 ) 2 . If at least one local character is nonprincipal, there is at most one supported exponent, b = m a x ( t 1 , t 2 ) . For b ≥ 2 , both characters must be primitive at that exponent; for b = 1 , one may be principal. These statements again follow by summing the last additive lifts. The nonzero kernel is
p b φ ( p b ) 2 ∑ v   m o d   p b p ∤ v ( H − v ) ψ p ( v ) ω p ( H − v ) ¯ .
For b = 1 the discarded Ramanujan term vanishes by the complete sum of the nonprincipal character; for b ≥ 2 use last-lift cancellation. The number of terms in the displayed sum is φ ( p b ) if p ∣ H , and p b − 1 ( p − 2 ) otherwise. Its modulus is consequently at most p / ( p − 1 ) or p ( p − 2 ) / ( p − 1 ) 2 , respectively.
Thus in every case the local sum of absolute values at a prime not dividing M is bounded by 1 + | c p ( H ) | / ( p − 1 ) 2 . Multiplication and monotone convergence yield
∑ d | Γ d | ≤ M φ ( M ) 2 ∏ p ∤ M 1 + | c p ( H ) | ( p − 1 ) 2 .
At p ∣ H this Euler factor agrees with that of S M ( H ) . At the other primes the ratio is at most [ 1 + ( p − 1 ) − 2 ] / [ 1 − ( p − 1 ) − 2 ] . Their product over p > 2 converges. This proves (3.3). ▫

4. The Major Arcs at the Short Interval Scale

Fix the parameters in (1.4), initially with
0 < s ≤ m i n 1 100 , c 4 2 .
Put L = l o g T , Q = T α , R 0 = T t , R 1 = T s and N = ⌈ 2 T ⌉ . For large T , (2.1) holds. Make the exceptional-character selection once at R 1 . For a requested even m ≤ Q choose
M , W , D = m , T t / 2 , T t / 2 , χ   absent   or   f > R 0 , l c m m , f , T 8 t , T 8 t , f ≤ R 0 .
The first branch has M D < R 0 , so a selected character with f > R 0 induces no character on its major-arc moduli. In the second branch M ≤ m T t and M D < R 1 ; no new character is selected.
The following inequalities all hold with a fixed power reserve:
α + t / 2 < t ,   α + 9 t < s ,   8 t + γ < s ,   α + 17 t + γ < 1 , 16 t + γ < 1 ,   24 t + 2 γ < 2 .
For instance ( α + 9 t ) / s = s 3 + 9 s < 1 . Hence, eventually,
4 M W D < Y ,   W 2 < Y ,   M D < T ,   W Y ≤ R 1 T < 1 2 ,   Y ≫ T R 1 .
For units A , B   ( m o d   M ) put
P A ( θ ) = ∑ T < p ≤ T + Y p ≡ A   ( M ) l o g p   e ( θ p ) .
In the first branch let w A ( u ) = 1 and k A = 1 . In the second set
δ = ε L ,   κ = m i n ( δ , 1 ) ,   w A ( u ) = 1 − χ ( A ) u − ε , k A = κ , χ ( A ) = 1 , 1 , χ ( A ) = − 1 .
Use the analogous definitions for B , and put λ = k A k B . On [ T , T + Y ] ,
c w k A ≤ w A u ≤ C w k A ,     ∥ w A ∥ ∞ + T V w A ≪ k A .
This follows from 1 − e − δ ≍ m i n ( δ , 1 ) , monotonicity, and l o g ( T + Y ) ≤ 2 L . Equation (2.3) gives
G ≪ m i n k A , k B e x p − c s .              
Lemma 4.1. Uniformly in these residue classes,
P A 0 ≪ k A Y φ M ,     ∫ 0 1 | P A θ | 2   d θ ≪ k A Y L φ M .
Proof. Character orthogonality at zero has coefficient modulus 1 / φ ( M ) for each inducing primitive character. The corrected model is φ ( M ) − 1 ∑ T < n ≤ T + Y w A ( n ) , and Lemma 2.1 bounds the aggregate error by O ( G Y / φ ( M ) ) . Apply (4.5) and (4.6). Parseval and ( l o g p ) 2 ≤ l o g ( 2 T ) l o g p prove the second bound. ▫
Define disjoint major charts on R / Z by
θ = j d + u M d + z , j   m o d   M ,   u   m o d   d ,   ( u , d ) = 1 ,   d ≤ D ,   | z | ≤ W d Y .
For d = 1 use u = 0 . Distinct centres have distance at least 1 / ( M d d ′ ) , while the sum of their radii is at most 2 W D / ( d d ′ Y ) ; (4.4) proves disjointness, also across the circle boundary. Let their union be M . Write
D A ( z ) = ∑ T < n ≤ T + Y w A ( n ) e ( z n ) ,     I A ( z ) = ∫ T T + Y w A ( v ) e ( z v )   d v .
Since every box prime exceeds M d , exact character orthogonality gives
P A c M d + z = K 1 A c ; M d D A z + ∑ ψ K ψ A c ; M d E ψ z .
In the refined branch K χ A = χ ( A ) K 1 A , which accounts for the signed weight. The same primitive function E ψ occurs on every chart which induces it.
For H ≡ A + B   ( m o d   M ) let R M ( H ) be the Fourier coefficient of 1 M P A P B , and let T A B ( H ) be its two-model counterpart. Summing the centre phases before taking absolute values gives the kernels (3.2). Each chart window is contained in | z | ≤ R 1 / T . Cauchy–Schwarz, Lemma 2.1, and discrete Parseval for D A bound the linear and quadratic error integrals by O ( B ψ k B Y ) and O ( B ψ B ω Y ) . Applying Proposition 3.1 before summing the primitive characters yields
| R M ( H ) − T A B ( H ) | ≪ M Y S M ( H ) φ ( M ) 2 ( ( k A + k B ) G + G 2 ) ≪ λ M Y S M H φ ( M ) 2 e x p − c s .
There is no multiplicity loss from induction in this estimate.

Completion of the Signed Model

The principal coefficient is supported on squarefree d coprime to M . Summing the aliases in (4.8) gives exactly
T A B ( H ) = M φ ( M ) 2 ∑ d ≤ D ( d , M ) = 1 μ ( d ) 2 c d ( H ) φ ( d ) 2 × ∫ | z | ≤ W / ( d Y ) D A ( z ) D B ( z ) e ( − H z )   d z .
In particular, the Ramanujan factor is c d ( H ) , not c d ( H − 2 T ) . Factoring an analytic phase from the interval model does not change this arithmetic factor.
Unit-cell quadrature, with the two partial endpoint cells separated, and integration by parts give
| D A − I A | ≪ k A ( 1 + Y | z | )   ( | z | ≤ 1 / 2 ) , | I A ( z ) | ≪ k A Y 1 + Y | z |   ( z ∈ R ) .
Indeed the total variation of w A ( v ) e ( z ( v − T ) ) is O ( k A ( 1 + Y | z | ) ) ; the phase e ( z T ) has modulus one. It follows for 0 < A 0 ≤ Y / 2 that
∫ | z | ≤ A 0 / Y | D A D B − I A I B |   d z ≪ λ ( A 0 + A 0 3 / Y ) .  
The continuous product tail outside W / ( d Y ) is O ( λ Y m i n ( 1 , d / W ) ) .
Let
J H = [ m a x ( T , H − T − Y ) ,   m i n ( T + Y , H − T ) ] , K A B ( H ) = ∫ J H w A ( v ) w B ( H − v )   d v .
For H ∈ I τ ( T , Y ) the overlap length is at least τ Y . Continuous Fourier inversion gives the coefficient K A B ( H ) for I A I B . For fixed 0 < ϑ < 1 , the Euler product also gives
∑ ( d , M ) = 1 μ ( d ) 2 | c d ( H ) | d ϑ φ ( d ) 2 ≤ e x p ( C ϑ ( l o g ( 2 H ) ) ϑ ) .
To see this, nondivisor primes give a convergent product. The extra logarithm at primes dividing H is O ϑ ( ∑ p ∣ H p ϑ − 1 ) . Splitting at l o g ( 2 H ) and bounding the number of larger divisors by l o g H / l o g l o g ( 2 H ) proves (4.14).
Consequently, with M A B ( H ) = M S M ( H ) K A B ( H ) / φ ( M ) 2 ,
T A B H − M A B H λ M Y / φ ( M ) 2 ≪ ϑ e C ϑ L ϑ W Y + W 3 Y 2 + W − ϑ + D − ϑ .
Here (4.13) handles the discrete-to-continuous change, m i n ( 1 , d / W ) ≤ ( d / W ) ϑ handles the integral tails, and (4.14) handles the omitted moduli. For ϑ = 1 / 4 and either branch the right side is at most
C e C L 1 / 4 ( T 8 t − 1 + γ + T 24 t − 2 + 2 γ + T − t / 8 ) ≤ T − t / 16
for sufficiently large T .
By (4.5), M A B ( H ) ≥ c * τ λ M Y S M ( H ) / φ ( M ) 2 for an absolute c * > 0 . Decrease the fixed s in (4.1) so that the constant in (4.10) times e − c / s is at most c * τ / 8 ; then increase T τ to absorb (4.15). We obtain
R R M H ≥ 3 c * τ 4 λ M Y S M H φ ( M ) 2 .
The source allowance e − c / s is a fixed small constant, whereas the completion error tends to zero as a power of T .

5. Minor Arcs and the Common Exceptional Set

Let m = ( R / Z ) \ M . For any θ ∈ m and any additive-filter shift θ + j / M , Dirichlet approximation with K = ⌈ Y / W ⌉ gives a reduced u / k such that
k ≤ K ≤ 2 Y / W ,     | θ + j / M − u / k | ≤ 1 / ( k K ) ≤ W / ( k Y ) ,     | θ + j / M − u / k | ≤ k − 2 .
The last inequality follows from k ≤ K , separately from the width bound. If k ≤ D , reducing M u / k produces a denominator d = k / ( k , M ) ≤ D and places θ in (4.8). This is impossible. Hence D < k ≤ 2 Y / W for every shift.
Apply (2.6) to the prefixes at T and T + Y . Proper powers contribute at most O ( T L 2 ) , which is absorbed by T 4 / 5 . Since D = W and T k ≪ T / W , the additive residue filter yields
s u p θ ∈ m P A θ ≪ T L 5 2 W − 1 2 + T − 1 5 .
The filter is an average of M shifted sums; it supplies no additional 1 / φ ( M ) factor in this supremum estimate.
Use Lemma 4.1 on the coordinate with smaller k A and (5.1) on the other. With F Y = M Y 3 / φ ( M ) 4 , we obtain
∫ m | P A P B | 2 ≤ λ 2 F Y ρ ,     ρ ≪ φ ( M ) 3 M L 6 κ − 3 ( T / Y ) 2 ( W − 1 + T − 2 5 ) ,
where κ = 1 in the first branch. The worst case is two suppressed classes, for which the mass factor divided by λ 2 costs κ − 3 . The multiplier ( T / Y ) 2 is essential. In the refined branch (2.2) gives
κ − 3 ≪ 1 + T 3 t / 2 L 3 .
Using φ ( M ) 3 / M ≤ M 2 , the powers of T in (5.2) are
2 α + 2 γ − t / 2 ,   2 α + 2 γ − 2 / 5 ,
2 α + 2 γ − 6 t ,   2 α + 2 γ − 9 t / 2 ,
2 α + 2 γ + 2 t − 2 / 5 ,   2 α + 2 γ + 7 t / 2 − 2 / 5 .
All are strictly smaller than − t / 4 for the parameters (1.4), 0 < t ≤ 10 − 4 . For the closest one the reserve is 0.23 t − 2 t 2 > 0 . The logarithmic factors are at most L 9 , so eventually
ρ ≤ T − t 4 .  
Let R m ( H ) be the Fourier coefficient of 1 m P A P B . For a fixed refined pair define the bad set by the threshold
| R m ( H ) | > c * τ 4 λ M Y S M ( H ) φ ( M ) 2 ,     H ∈ I τ ( T , Y ) ,   H ≡ A + B   ( m o d   M ) .
Full Parseval, (5.2) and (5.3) give
B M , A , B ≪ τ Y M T − t 4 .
Restriction to a target residue class does not yield an extra 1 / M . Outside this set, (4.16) implies that the actual log-weighted prime-pair count is at least
c τ λ M Y S M ( H ) φ ( M ) 2 .
For a coarse pair a , b   ( m o d   m ) in the refined branch, consider each class c   ( m o d   M ) above a + b   ( m o d   m ) . Choose one unit lift A c ≡ a   ( m o d   m ) for which B c = c − A c is a unit above b . Such a choice exists by CRT. At a prime already dividing m both classes are units; at a new prime, necessarily odd, excluding 0 , c leaves at least p − 2 ≥ 1 choices. Lift to the relevant prime powers and fix one pair for the entire class c . There are M / m such classes. The union of their restricted bad sets has size
B m , a , b ≪ τ Y m T − t 4 .
There is no union over all refined residue pairs. In the dense branch M = m and the same estimate is immediate.
Take the union over all requested coarse pairs and all even m ≤ Q . It is one set E τ ( T ) , and
E τ T ≪ τ Y T − t 4 ∑ m ≤ Q 2 ∣ m φ ( m ) 2 m ≪ τ Y T − t 4 + 2 α .
Since t / 4 − 3 α > β = t / 8 , this proves (1.1). Any grid intersection is at most the total size, and m ≤ T α then proves (1.2).
Finally (2.2) implies κ ≫ T − t / 2 / L in the refined branch, hence λ ≫ T − t / L 2 . The dense value λ = 1 satisfies this weaker bound too. As M ≤ m T t , φ ( M ) ≤ M and S M ( H ) ≥ C 2 , (5.5) is at least c τ Y T − 2 t / ( m L 2 ) . Dividing by the upper bound l o g 2 ( 2 T ) ≪ L 2 for the weight of one ordered pair proves (1.3). This completes the proof of Theorem 1.1.

6. The Classical Polylogarithmic Comparison

For even H let
S ( H ) = 2 C 2 ∏ p ∣ H p > 2 p − 1 p − 2 , l ( H ) = m a x { 0 , m i n ( T + Y , H − T ) − m a x ( T , H − T − Y ) } .
For admissible a   ( m o d   q ) , meaning ( a ( H − a ) , q ) = 1 , set
σ q H = S H q ∏ p ∣ q 1 − ρ p H p − 1 ,     ρ p H = 1 , p ∣ H , 2 , p ∤ H .
Write σ ( H ; a , q ) = 0 in the inadmissible case and σ ( H ; a , q ) = σ q ( H ) otherwise. The local product is the specialization of Kawada’s singular series to the forms u , H − u . In particular σ q ( H ) ≥ 2 C 2 / q for even admissible targets.
Proposition 6.1. Classical corollary of Kawada. Fix 2 / 3 < θ < 1 , 0 < τ < 1 and C , B , D > 0 . Put Y = T θ , L = l o g T , Q = L C . There is a common set E ⊂ I τ ( T , Y ) with
| E | ≪ Y L − B − C ,     # { H ∈ E : H ≡ c   ( m o d   m ) } ≪ ( Y / m ) L − B
for all even m ≤ Q and all c . Outside E , simultaneously for all compatible reduced pairs modulo these moduli,
r ( H ; m , a , b ; T , Y ) = σ m ( H ) l ( H ) L 2 ( 1 + O ( L − D ) ) .
Proof. In [6, Theorem 2], set the source location x = T + Y , length y = Y , and use the fixed coefficient forms u , H − u . Its mean-value estimate gives, for every fixed A > 0 ,
∑ q ≤ Q m a x a   m o d   q ∑ H ∈ Z | Ψ ( H ; q , a ) − σ ( H ; a , q ) l ( H ) | ≪ Y 2 L − A ,  
where Ψ is the two- Λ convolution in the common box and Z is the set with positive real overlap. The source gates Y ≥ x 2 / 3 ( l o g x ) C 0 and Q ≤ Y x − 1 / 2 ( l o g x ) − B 0 hold eventually for our fixed parameters. Restricting the nonnegative error sum to the interior is legitimate.
Declare an admissible triple bad if its error exceeds Y / ( q L D ) . For a fixed q , Markov’s inequality and a union over at most φ ( q ) classes give at most ( q φ ( q ) L D / Y ) m a x a ∑ H | e r r o r | bad targets. Summing (6.3) yields | E | ≪ Q 2 Y L D − A . Take A ≥ B + 3 C + D + 2 . The grid estimate follows from the total bound and m ≤ Q . There is no second independent residue union: b = H − a is determined.
Since l ( H ) ≥ τ Y and σ m ( H ) ≫ 1 / m , the good weighted sums have relative error O ( L − D ) . At a fixed H , proper powers in either coordinate contribute at most O ( T L 3 ) : each such coordinate determines the other. This is smaller than Y / ( q L D ) for q ≤ L C . Finally l o g p l o g r = L 2 ( 1 + O ( Y / ( T L ) ) ) throughout the box; as θ < 1 , its relative error is smaller than every fixed inverse logarithm. This proves (6.2). ▫
Remark 6.2. The same-box geometry and Proposition 6.1 are classical consequences and are not claimed as new. The parameters in Theorem 1.1 concern a different modulus/error tradeoff: fixed-power moduli and power-small common exceptions, at the cost of a much weaker lower count and a nonoptimized interval exponent. The larger range of moduli in (6.3) alone is insufficient to make that tradeoff.
For completeness, the following logical example makes the last point precise. Fix 0 < α < θ − 1 / 2 and let q 0 ≍ T α be a power of two. Start with the nonnegative main-term array F ( H ; q , a ) = σ ( H ; a , q ) l ( H ) , and set it to zero only for q = q 0 , H ≡ 0   ( m o d   q 0 ) in the interior, and a = 1 , − 1   ( m o d   q 0 ) . This preserves pair-swap symmetry. There are ≍ τ Y / q 0 affected targets. Since σ q 0 ( H ) = 2 S ( H ) / q 0 ≪ L / q 0 , the entire norm change in (6.3) is
O ( Y 2 L / q 0 2 ) = o ( Y 2 L − A )
for each fixed A . Nevertheless, all these targets fail for the pair ( 1 , − 1 ) modulo q 0 , contradicting any common exceptional-set bound of order ( Y / q 0 ) T − β with fixed β > 0 . Here S ( H ) ≪ l o g ( 2 H ) follows, for example, by ordering the odd prime divisors and bounding their product of 1 + 1 / ( p − 2 ) by one plus their number. This example is an abstract real array, not a prime convolution. It proves a non-implication between norms, not a counterexample about primes or a priority claim.

7. Uniform Removal of Proper Prime Powers

The following elementary estimate is useful when translating a progression statement between Λ and ordinary prime log weights.
Lemma 7.1. Fix positive c , C , ϵ . Uniformly in integers c x ≤ q ≤ C x , reduced classes a   ( m o d   q ) , and intervals I ⊂ [ 0 , C x ] ,
∑ p j ∈ I ,   j ≥ 2 p j ≡ a   ( q ) l o g p ≪ c , C , ϵ x ϵ .      
Proof. Choose a fixed integer J ≥ m a x ( 2 , 4 / ϵ ) . For 2 ≤ j ≤ J , reducedness forces the base to be a unit modulo q . The fibre of the j th-power homomorphism of ( Z / q Z ) * is either empty or a coset of its kernel. Odd prime-power factors give at most j roots each. The 2 -primary unit group has at most two cyclic factors, one of order at most two, and gives at most 2 j . CRT therefore bounds the number of roots by 2 j ω ( q ) . For fixed j and ν > 0 , j ω ( q ) ≪ j , ν q ν : the finitely many primes below j 1 / ν contribute a constant, and every larger prime absorbs its factor j into p ν .
Each root class has O c , C ( 1 ) positive representatives below ( C x ) 1 / j ≪ C x , because q ≥ c x . Taking ν = ϵ / 4 shows that the contribution of j ≤ J is O c , C , ϵ ( x ϵ / 4 l o g x ) . For j > J , count all integer bases up to ( C x ) 1 / ( J + 1 ) ; there are O C ( l o g x ) possible exponents and each weight is O C ( l o g x ) . The result is O C , ϵ ( x ϵ / 4 ( l o g x ) 2 ) = O ( x ϵ ) . Dropping membership in I only increases these bounds. ▫
Corollary 7.2. Let M 0 ≥ 1 be a fixed function with M 0 = x o ( 1 ) , and let A x be any set of distinct reduced classes ( q , a ) with c x ≤ q ≤ C x M 0 . For every fixed F , δ > 0 ,
∑ q , a ∈ A x τ ( q ) F s u p x ≤ T ≤ 3 x 2 ∑ x ≤ p j < T ,   j ≥ 2 p j ≡ a   q l o g p 2 ≪ F , c , C , δ x 1 + δ .
The eventual threshold may depend on the chosen function M 0 . Consequently, an anchored maximal-variance estimate of size O B ( x 3 / 2 ( l o g x ) − B ) for every fixed B transfers between Λ and ordinary prime log weights on this family.
Proof. There are at most ∑ q ≤ C x M 0 q = O ( x M 0 2 ) classes. Apply Lemma 7.1 with a sufficiently small fixed ϵ , square, and use τ ( q ) F = x o ( 1 ) . For δ < 1 / 2 , the resulting error is smaller than the asserted variance scale at every fixed logarithmic precision. The two normalizations have the same interval-length main term; apply | u + v | 2 ≤ 2 | u | 2 + 2 | v | 2 in either direction. This transfers an estimate if it is proved; it proves neither variance. ▫

8. A Separate Mixed-Correlation Question

The simultaneous representation theorem above does not close a pointwise reflected-prime comparison. We state the distinction precisely. Let N = 2 x , L = l o g x , l = l o g L , z = e x p ( L / l 2 ) , P z = ∏ p ≤ z p , and V = ∏ p ≤ z ( 1 − 1 / p ) . On x / 2 < n ≤ x , put
a n = l o g ( N − n ) 1 N − n   p r i m e ,     b n = V − 1 1 ( N − n , P z ) = 1 ,     w n = a n − b n ,
and set these sequences to zero elsewhere. For distinct d , e ≍ x , ( d e , N ) = 1 , take a real interval I d , e = ( u , v ] , u ≤ v , such that both d k , e k ∈ ( x / 2 , x ] throughout it, and define
A B d , e = V − 1 ∑ k ∈ I d , e l o g ( N − d k ) 1 N − d k   p r i m e 1 ( N − e k , P z ) = 1 .
Bonferroni truncation of the last factor involves squarefree m ∣ P z with ω ( m ) ≤ r + 1 , where r ≍ K l o g l o g x and K is fixed after the desired logarithmic precision. Thus m ≤ R = z r + 1 = x o ( 1 ) .
For a retained reduced term the conditions are ( m , e N ( e − d ) ) = 1 . If k 0 ≡ N e − 1   ( m o d   m ) , the exact prime-coordinate progression is
q = d m ,     a d , e , m ≡ N − d k 0   ( m o d   d m ) , N − d k ∈ [ N − d v , N − d u ) .
Indeed k = k 0 + m h gives the product modulus d m , also when ( d , m ) > 1 . The class is reduced: at primes dividing d , use ( d , N ) = 1 , and at primes dividing m , use N − d k 0 ≡ N e − 1 ( e − d ) ≢ 0 .
Lemma 7.1 removes proper powers from these actual retained sums at a total cost O F , K , δ ( x 1 + δ ) , even with fixed weights τ ( d ) F τ ( e ) F and the factor V − 1 . There are O ( x ) slope pairs and at most x o ( 1 ) choices of m ; all other displayed factors are subpower. This is negligible against x 3 / 2 ( l o g x ) − A for every fixed A when δ < 1 / 2 .
What remains is an estimate for the actual prime discrepancy in (8.1), including its selected residues and pair-dependent endpoints. The modulus exponent alone does not provide it. For example, a triply well-factorable weight of level Q vanishes on any integer with a prime factor greater than Q 1 / 3 : use the factorization into three equal support lengths. Thus the prime-row instance d ≍ x cannot be represented directly by the triply well-factorable weights in [10, Theorem 1.3]. For the signed two-factor theorem of [7, Theorem 1.1], the conditions are Q 1 2 Q 2 < x 1 − ϵ and Q 1 7 Q 2 12 < x 4 − ϵ . A prime d ≍ x must divide one factor: placement in the first violates the first condition, and placement in the second violates the second. These are restrictions on those direct applications, not lower bounds for the actual errors or impossibility results for other methods.
In particular, no bound for every separated pair of coefficients | ξ d | , | κ k | ≤ τ ( ⋅ ) F in
∑ ( x / 2 ) 1 / 4 < d ≤ x ξ d ∑ x / 2 < d k ≤ x κ k w d k
is established here. Even a mixed prime–rough estimate would leave the prime–prime part of the centred quadratic form to be controlled. This separation is necessary when interpreting the scope of an almost-all representation result.

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.

Use of Artificial Intelligence

Generative AI tools were used substantially in developing and checking arguments, comparing sources, drafting the text, and preparing the document. These activities do not constitute independent human peer review or formal proof verification. No computer experiment is used as a proof of an asymptotic statement.

Data and Reproducibility

The results are analytical. Definitions, endpoint conventions, parameter dependencies, and proofs of the deductions are included in the text; the general analytic inputs are identified in the cited primary sources. No empirical dataset is required for the stated deductions.

Appendix A. A Classical Mixed Theorem Below the Square-Root Scale

Let N = 2 x be even, L = l o g x , l = l o g L , z = e x p ( L / l 2 ) , P z = ∏ p ≤ z p , and V = ∏ p ≤ z ( 1 − 1 / p ) . Fix 0 < c < C . Take distinct integers d , e ∈ ( c D , C D ] with ( d e , N ) = 1 , and any interval I d , e = ( u , v ] , u ≤ v , on which d k , e k ∈ ( x / 2 , x ] . Put H = v − u = O ( x / D ) and
A B d , e = V − 1 ∑ k ∈ I d , e l o g ( N − d k ) 1 { N − d k   p r i m e } 1 { ( N − e k , P z ) = 1 } .
Intervals and row subsets may depend on d , e , N ; empty intervals contribute zero. No primality restriction is imposed on k . Let ν p count roots of ( N − d k ) ( N − e k ) modulo p , and set
C z = ∏ p ≤ z 1 − ν p / p ( 1 − 1 / p ) 2 ,     C ( N ; d , e ) = ∏ p 1 − ν p / p ( 1 − 1 / p ) 2 .
If p ∣ N , then ν p = 1 . Otherwise its value is 0 if p divides both rows, 1 if it divides exactly one, and, when both are units, 1 or 2 according as d = e   ( m o d   p ) or not. Thus the p = 2 factor is positive. For distinct rows the product converges positively, since outside the divisors of N d e ( d − e ) its factors are 1 − ( p − 1 ) − 2 .
Theorem A.1. Fix 0 < γ < 1 / 2 and set D = x γ . For every fixed A > 0 , F ≥ 0 , uniformly for each even N = 2 x , all allowed row subsets and all the above intervals,
∑ d , e τ ( d ) F τ ( e ) F | A B d , e − H d , e C ( N ; d , e ) | ≪ A , F , γ , c , C D x ( l o g x ) A .  
The same assertion holds with B A d , e = A B e , d , retaining the same common interval.
Proof. We first keep the exact sieve-divisor progressions. For squarefree m ∣ P z , let T d , e ( m ) denote the ordinary-prime log sum with m ∣ N − e k . If ( m , e ) > 1 , the congruence is impossible. Otherwise choose k 0 = N e − 1   ( m o d   m ) . The relation k = k 0 + m j gives, in both directions,
q = d m ,     a d , e , m = N − d k 0   ( m o d   d m ) , N − d k ∈ [ N − d v , N − d u ) .
This is the product modulus, also when ( d , m ) > 1 . Since a = N   ( m o d   d ) and e a = N ( e − d )   ( m o d   p ) for p ∣ m , the class is reduced precisely when ( m , N ( e − d ) ) = 1 . Inadmissible classes contribute zero: their prime values would be divisible by a prime at most z , whereas all values are in [ x , 3 x / 2 ) . The complete condition is ( m , e N ( e − d ) ) = 1 .
For reduced classes write
Δ < ( t ; q , a ) = ∑ p < t p ≡ a   ( q ) l o g p − t φ ( q ) .
Literal endpoints then give
T d , e ( m ) = d H φ ( d m ) + R d , e ( m ) , R d , e ( m ) = Δ < ( N − d u ; d m , a ) − Δ < ( N − d v ; d m , a ) .  
For empty intervals put the remainder equal to zero.
Let r be the least even integer at least K l , with K a sufficiently large fixed precision reserve. For s = r , r + 1 , put
E d , e , s = ∑ m ∣ P z ,   ω ( m ) ≤ s ( m , e N ( e − d ) ) = 1 μ ( m ) R d , e ( m ) .
Even and odd Bonferroni inequalities, multiplied by the nonnegative prime weight, bracket A B by the two signed remainders and their local main terms. The squarefree multiplicative coefficient g d ( m ) = φ ( d ) / φ ( d m ) has g d ( p ) = 1 / p for p ∣ d and 1 / ( p − 1 ) otherwise. The full finite main coefficient is exactly
d φ ( d ) V ∏ p ≤ z p ∤ e N ( e − d ) ( 1 − g d ( p ) ) = C z ∏ p ∣ d p > z ( 1 − 1 / p ) − 1 .
To check this identity, primes dividing N , exactly one row, or e − d with both rows units contribute ( 1 − 1 / p ) − 1 ; primes dividing both rows contribute ( 1 − 1 / p ) − 2 ; the remaining primes contribute ( 1 − 2 / p ) / ( 1 − 1 / p ) 2 . These are exactly the root factors above.
The included g d ( p ) sum to O ( l ) , and d / ( φ ( d ) V ) ≤ L C 0 . The elementary-symmetric tail after order r is at most 2 ( e C 1 l / ( r + 1 ) ) r + 1 . Choosing K after any required logarithmic precision makes this tail negligible. Also C z ≤ L C 2 , C / C z = 1 + O ( l 2 / z ) , and the extra product in (A.3) is 1 + O ( L / ( z l o g z ) ) . For the tail ratio, the exceptional integer N d e ( d − e ) has polynomial size: at most O ( L / l o g z ) of its prime divisors exceed z , while the other prime-square tail is O ( 1 / z ) . Consequently, for every fixed reserve A ′ > 0 ,
A B d , e − H C N ; d , e ≤ V − 1 m a x s = r , r + 1 E d , e , s + O A ′ x D L − A ′ .
Every retained m is at most R = z r + 1 = e x p ( O K ( L / l ) ) = x o ( 1 ) . No single fixed K is required to serve every precision.
The classical maximal Bombieri–Vinogradov theorem [12, printed pp. 1,3] implies, whenever Q ≤ x 1 / 2 − ϵ with fixed ϵ > 0 ,
∑ q ≤ Q τ ( q ) J m a x ( a , q ) = 1 s u p x ≤ t ≤ 3 x / 2 | Δ < ( t ; q , a ) | ≪ B , J , ϵ x L − B
for all fixed B , J . Here the fixed divisor weights cost only logarithmic reserves: for the corresponding ψ -error the individual trivial bound is O ( x [ L + τ ( q ) ] / q ) ; divisor harmonic moments and Cauchy–Schwarz with arbitrary-precision unweighted maximal BV give the weighted statement. Removing proper powers costs O ( Q x L C J ) , and changing endpoint convention costs O ( Q L C J ) . Both are power smaller than x in this fixed-gap range, so the displayed form concerns ordinary prime log weights.
Since γ < 1 / 2 is fixed, Q = C D R ≤ x 1 / 2 − ϵ eventually for some fixed ϵ > 0 . Write E q ( x ) for the maximum in the last display. From (A.2), | R d , e ( m ) | ≤ 2 E d m ( x ) . Summing the other row costs O F ( D L C F ) . There are at most τ ( q ) decompositions q = d m , and τ ( d ) F ≤ τ ( q ) F . Hence
∑ d , e τ ( d ) F τ ( e ) F m a x s | E d , e , s | ≪ F D L C F ∑ q ≤ C D R τ ( q ) F + 1 E q ( x ) .
There is no multiplicity factor R after grouping the modulus. Apply maximal BV with adequate reserves, absorb V − 1 ≤ L C 3 , and sum the deterministic term in (A.4) against the two divisor envelopes. This proves (A.1). Exchanging the rows proves the B A assertion. ▫
Remark A.2. This is a classical-method absolute slope average for each fixed γ < 1 / 2 , with no average over N . It is neither a per-pair asymptotic nor a theorem at γ = 1 / 2 . At the balanced endpoint d m may reach x   x o ( 1 ) ; the proof does not place those moduli inside the arbitrary-accuracy BV range. Failure of this sufficient maximal-error relaxation is not a lower bound for the actual signed errors. The theorem estimates neither the two-prime A A term nor every separated coefficient test, and does not imply binary Goldbach.

Appendix B. Distribution and Factor Regularity of the Rough Comparison

The sequences a , b , w are those of Section 8. This section concerns actual distributional properties of b and a Type I comparison with a . None of these statements estimates Type II.
Theorem B.1. Weighted Type I with interval maxima. For fixed 0 < γ < 1 / 2 , F ≥ 0 and A > 0 ,
∑ d ≤ x γ τ ( d ) F s u p I   interval ∑ k ∈ I w d k ≪ γ , F , A x L − A .
No coprimality restriction on d is imposed in this assertion.
Proof. On the finite common support, an interval selects an integer block K 0 < k ≤ K 1 ; write M = K 1 − K 0 and h = d M ≤ x . For ( d , N ) = 1 , the exact reflected ordinary-prime progression has real main term h / φ ( d ) . Weighted maximal BV and the proper-power conversion in Appendix A bound its summed maximal error by O ( x L − A ) . If ( d , N ) > 1 , an ordinary prime N − d k ≥ x would have to equal a prime divisor of d < x , a contradiction. Define g N ( d ) = 1 ( d , N ) = 1 / φ ( d ) .
If a prime p ≤ z divides both d and N , the rough count is zero. Otherwise primes at most z not dividing d each exclude one class of k . Finite Bonferroni through adjacent orders r , r + 1 , r ≍ K l , counts the block with error O ( M L − B + z r + 1 ) , uniformly in its position and length. This follows by counting one residue class for each squarefree divisor and bounding the reciprocal tail by 2 ( e C l / ( r + 1 ) ) r + 1 . After dividing by V , its density in the h variable is
g z , N ( d ) = 0 , some   p ≤ z   divides   ( d , N ) , d − 1 ∏ p ∣ d , p ≤ z ( 1 − 1 / p ) − 1 , otherwise .
The total rough-count error is O ( x L C F − B + x γ + o ( 1 ) ) with adequate fixed log reserves. It remains to match densities without declaring rows with only large common prime factors to be zero. For ( d , N ) = 1 ,
0 ≤ g N ( d ) − g z , N ( d ) ≤ τ ( d ) d ∑ p ∣ d , p > z 1 p .
Sum with τ ( d ) F , put d = p m , and use τ ( p m ) ≤ 2 τ ( m ) and divisor harmonic moments. This costs O F ( L C F / z ) . For ( d , N ) > 1 and no small common factor, the corresponding upper bound is O F ( L C F ∑ p ∣ N , p > z p − 1 ) = O F ( L C ′ F / z ) . Multiplication by h ≤ x makes either density correction O A , F ( x L − A ) . Choosing B , K after the requested precision and subtracting the two actual block sums proves the result. ▫
Theorem B.2. Prime mass and local factor. Uniformly in each even N = 2 x and all real x / 2 ≤ t ≤ x , for every fixed A > 0 ,
∑ x / 2 < p ≤ t b p = S ( N ) ∫ x / 2 t d u l o g u + O A ( x L − A ) .
In particular B N = ∑ p b p ≍ S ( N ) x / L > 0 for large x .
Proof. On prime support p > x / 2 > z , each sieve prime l ∤ N excludes one reduced class p = N   ( m o d   l ) ; primes l ∣ N exclude none. Bonferroni truncation at order O K ( l ) gives distinct squarefree moduli at most e x p ( O K ( L / l ) ) , all inside any fixed sub-square-root BV range. Weighted maximal BV, ordinary prime-power removal and the elementary-symmetric tail estimate give
V − 1 ∑ x / 2 < p ≤ t l o g p   1 ( N − p , P z ) = 1 = S z ( N ) ( t − x / 2 ) + O A ( x L − A ) ,
where S z ( N ) = V − 1 ∏ l ≤ z , l ∤ N ( 1 − 1 / ( l − 1 ) ) . Partial summation removes the log weight. Its local factors are those of S ( N ) ; the good-prime tail is O ( 1 / z ) and the primes dividing N above z contribute O ( l o g N / ( z l o g z ) ) to its logarithm. Thus S z / S = 1 + O ( l 2 / z ) uniformly. Both factors are polynomially bounded in L , and S ( N ) ≥ 2 C 2 > 0 , proving the theorem. ▫
Definition B.3. Sorted factor coordinates. Fix 0 < ν ≤ 1 / 2 and an integer 2 ≤ k ≤ ⌊ 1 / ν ⌋ . Let
Δ k , ν = { u ∈ R k : ∑ u i = 1 , ν ≤ u 1 ≤ … ≤ u k } .
Write an integer with exactly k prime factors, counted with multiplicity, uniquely as n = p 1 … p k with p 1 ≤ … ≤ p k , and put v ( n ) = ( l o g p i / l o g n ) i . For a convex subset T ⊂ Δ k , ν and a bounded one-Lipschitz f : T → C , | f | ≤ 1 , define
Q b ( T , f ) = ∑ n ∈ ( x / 2 , x ] Ω ( n ) = k ,   v ( n ) ∈ T b n f ( v ( n ) ) ,     I T ( f ) = ∫ T f ( u ) u 1 … u k   d u 1 … d u k − 1 .
Here u k = 1 − u 1 − … − u k − 1 , so the integral uses projection measure, not Euclidean surface measure. Degenerate convex sets have zero measure; the discrete sums retain their literal boundary atoms.
Theorem B.4. Uniform convex factor regularity. For each fixed ν , uniformly in the above k , T , f and even N = 2 x ,
Q b ( T , f ) = B N { I T ( f ) + O ν ( L − 2 ) } .
Repeated prime factors and arbitrary convex endpoint conventions are included. The constants need not be uniform as ν → 0 .
Proof. First fix a rough integer m with ( m , P z ) = 1 and let Y = x / m lie in [ x η , x ] , with fixed η > 0 . The proof of Theorem B.2 works at this last-prime scale: the active classes are q = N m − 1   ( m o d   l ) for l ∤ N , and every truncated modulus is at most e x p ( O K ( L / l ) ) ≤ Y 1 / 4 eventually. After partial summation and PNT it gives, for every fixed A ,
s u p Y / 2 ≤ t ≤ Y ∑ Y / 2 < q ≤ t , q   prime ( b m q − S z ( N ) ) ≪ η , A Y L − A .
No coprimality between m and N is required.
Fix a sorted initial tuple p 1 , … , p k − 1 , including repeats, put m = ∏ p i , c i = l o g p i and C = l o g m . On its last-prime section the exact factor vector is ( c 1 ρ , … , c k − 1 ρ , 1 − C ρ ) , where ρ = 1 / l o g ( m q ) . Convexity makes the section an interval in q , also after sorting and product restrictions. Nonempty sections have p i > ( x / 2 ) ν , Y ≥ ( x / 2 ) ν , hence all initial factors are rough and (R1) applies with η = ν / 2 . The path length as q traverses ( Y / 2 , Y ] is O ( 1 / L ) , so f ( v ( m q ) ) has bounded variation there. Discrete Abel summation using (R1), including left limits, gives a row error O ( Y L − A ) . Its sum over all initial tuples is O ν , k ( x L − A ) , since
∑ Y ≤ x ∑ ( x / 2 ) ν ≤ p ≤ x 1 p k − 1 ≪ ν , k x .
Writing Q 0 for the sum with b n replaced by one, we have proved Q b = S z Q 0 + O ν , k , A ( x L − A ) .
For completeness we evaluate Q 0 uniformly in the convex region. Lift T to the convex logarithmic slab
U = { u ∈ R k : 1 − ( l o g 2 ) / L < s = ∑ u i ≤ 1 ,   u / s ∈ T } .
Convexity follows by writing the normalized vector of a convex combination as the convex combination with weights proportional to the two positive radial coordinates. For large x , U ⊂ [ ν / 2 , 1 ] k . Partition the ambient cube into cubes of side M − 1 with M = ⌈ L 3 ⌉ . A convex subset of a bounded k -cube meets O k ( M k − 1 ) boundary cubes: assign supporting normals by their largest coordinate and its sign; within a column perpendicular to that coordinate, the supporting inequality bounds the vertical separation by ( k − 1 ) / M . There are O ( M k − 1 ) columns and finitely many assignments. The same argument on the closure covers lower dimensional sets and all boundary choices.
For a cube meeting U , its coordinate upper endpoints b i satisfy ∑ b i ≤ 1 + k / M , hence x ∑ b i ≤ 2 x eventually. Ordinary PNT with fixed error exponent ten gives in each coordinate
π ( x b ) − π ( x b − 1 / M ) = ∫ b − 1 / M b x u u   d u + O ν ( x b L − 10 ) .
The integral is comparable to x b / M . Multiplying the coordinate counts and summing the errors over at most M k cubes costs O ν , k ( x M / L 10 ) . Boundary cubes cost O ν , k ( x / M ) ; the variation of the uniformly Lipschitz function f ( u / s ) on interior cubes costs the same. Thus
Q 0 = ∫ U f ( u / s ) x s u 1 … u k   d u + O ν , k ( x / L 3 ) .
Ties between sorted factors are contained in the boundary cover and are counted once, including actual repeated primes. Under u i = s v i the Jacobian is s k − 1 , so the integral equals exactly
I T ( f ) ∫ 1 − ( l o g 2 ) / L 1 x s s   d s = I T ( f ) ∫ x / 2 x d t l o g t .
Combine this with Q b = S z Q 0 + O ( x L − A ) and B N = S z ∫ x / 2 x d t / l o g t + O ( x L − A ) . The positive lower bound for S z and B N ≍ S z x / L make the normalized error O ν ( L − 2 ) . Only finitely many k occur for fixed ν , completing the assertion. ▫

Appendix C. A Mixed Estimate with an Accuracy Dependent Moving Endpoint

The fixed exponent in Theorem A.1 cannot simply be allowed to vary. The quantitative form of the input does, however, give the following more explicit classical consequence. The notation and the local product are those of Appendix A.
Lemma C.1. Weighted maximal distribution in a logarithmically shortened range.
For every fixed a > 0 and J ≥ 0 there is a fixed b = b ( a , J ) > 0 such that, whenever Q ≤ x ( l o g x ) − b ,
∑ q ≤ Q τ ( q ) J m a x ( h , q ) = 1 s u p x ≤ t ≤ 3 x / 2 θ < ( t ; q , h ) − t φ ( q ) ≪ a , J x ( l o g x ) − a .
Proof. Write L = l o g x and let E q ψ denote the corresponding von Mangoldt maximum, enlarged to 0 ≤ t ≤ 2 x . The quantitative maximal theorem in [12, p. 3] gives
∑ q ≤ Q E q ψ ≪ B x L − B + x Q L 6 .
For q ≤ x , integer counting and q / φ ( q ) ≤ τ ( q ) give E q ψ ≪ x L τ ( q ) / q . The elementary divisor inequalities
∑ q ≤ Q τ ( q ) t q ≪ t L 2 ⌈ t ⌉ ,     ∑ q ≤ Q τ ( q ) t ≪ t Q L 2 ⌈ t ⌉ − 1
follow by dominating τ t by the 2 ⌈ t ⌉ -fold divisor function and expanding its ordered factors. Put C J = 1 + 2 ⌈ 2 J + 1 ⌉ . Cauchy’s inequality yields
∑ q ≤ Q τ ( q ) J E q ψ ≪ J x L ( C J − B ) / 2 + x L ( C J + 6 − b ) / 2 .
Choose B ≥ 2 a + C J + 2 and b ≥ 2 a + C J + 8 . The total proper-prime-power log weight up to 2 x is O ( x L 2 ) . Its weighted aggregate over the moduli is O ( x L 2 ⌈ J ⌉ + 1 − b ) ; enlarge b to absorb it. The change from ≤ to < costs at most O ( L ) per modulus, whose weighted sum is O ( x L 2 ⌈ J ⌉ ) . These estimates prove the stated ordinary-prime version. ▫
Theorem C.2. Moving subendpoint mixed estimate. For every fixed A > 0 , F ≥ 0 and 0 < c < C there is a fixed κ > 0 such that, uniformly for
2 ≤ D ≤ x e x p − κ l o g x l o g l o g x ,
the conclusion (A.1) holds, with constants depending only on A , F , c , C . It holds for each even N = 2 x , all the allowed row subsets, and all pair dependent interval endpoints.
Proof. The exact reduction (A.4) holds for arbitrary 2 ≤ D ≤ x ; only its subsequent distribution estimate used a fixed exponent. Fix its local precision A ′ sufficiently large in terms of A , F . Choose its fixed Bonferroni constant K and let r be the least even integer at least K l , where l = l o g l o g x . The two truncations have orders r , r + 1 , so
R = z r + 1 ,     l o g R ≤ K L l + 3 L l 2 .
Take κ = K + 1 . For l ≥ 6 the total modulus range satisfies
Q = C D R ≤ C x e x p − L 2 l ≤ x L − b
eventually, for every fixed b , since L / l 2 → ∞ . Apply Lemma C.1 with J = F + 1 and sufficiently large accuracy to the modulus grouping in Appendix A:
∑ d , e τ ( d ) F τ ( e ) F m a x s | E d , e , s | ≪ F D L C F ∑ q ≤ C D R τ ( q ) F + 1 E q ( x ) .
The factor V − 1 ≪ L is absorbed by the logarithmic reserve. The local error in (A.4), summed over the two row envelopes, is O ( D x L C ′ F − A ′ ) and is absorbed by our choice of A ′ . All constants are now fixed independently of D ; in particular no constant from a fixed- γ theorem has been passed to a limit. Exchanging the two rows proves the B A version. ▫
Remark C.3. At the largest permitted D , the exponent l o g D / l o g x = 1 / 2 − κ / l o g l o g x tends to 1 / 2 , while D / x → 0 . The constant κ is chosen after the accuracy. There is no assertion that one κ works for all A , F , or that the original proportional square-root band is covered.

Appendix D. Balanced Rough Pairs, Two Prime Upper Bounds, and a Covariance Criterion

Throughout this section N = 2 x , Y = x , and the rows lie in ( c Y , C Y ] with ( d e , N ) = 1 . For a common-support interval I d , e of length H d , e define
A A d , e = ∑ k ∈ I d , e a d k a e k , A B d , e = ∑ k ∈ I d , e a d k b e k , B A d , e = ∑ k ∈ I d , e b d k a e k , B B d , e = ∑ k ∈ I d , e b d k b e k .
For distinct rows let ρ p count roots of ( N − d t ) ( N − e t ) modulo p , and put
C z ( N ; d , e ) = ∏ p ≤ z 1 − ρ p / p ( 1 − 1 / p ) 2 ,     C ∞ ( N ; d , e ) = ∏ p 1 − ρ p / p ( 1 − 1 / p ) 2 .
Lemma D.1. Root densities and completion. One has ρ 2 = 1 and 0 ≤ ρ p < p for every prime. The product C ∞ converges for d ≠ e . If S y denotes its truncation at y ≥ 3 , then, for large y ,
l o g ( C ∞ / S y ) = O C 1 y + l o g x y l o g y .
In particular C z = C ∞ + O M ( ( l o g x ) − M ) for every fixed M .
Proof. If p ∣ N , the unit slopes have one common root. If p ∤ N , there are no roots when both slopes vanish, one when just one vanishes, and otherwise one or two according as p ∣ d − e or not. This proves the table and positivity. Outside the nonzero integer Δ = N d e ( d − e ) , which has size O C ( x 5 / 2 ) , the factor is 1 − ( p − 1 ) − 2 . Its logarithmic tail is O ( 1 / y ) . The bad-prime tail is at most a constant times ∑ p ∣ Δ , p > y p − 1 ≤ l o g | Δ | / ( y l o g y ) . This proves the displayed bound. Finally C z ≪ L C 0 by the root table and the elementary prime harmonic estimate, and the superlogarithmic size of z absorbs this factor. ▫
Theorem D.2. Uniform rough-pair benchmark. For every fixed M > 0 , uniformly in distinct rows and all endpoints,
B B d , e = H d , e C ∞ ( N ; d , e ) + O M ( Y L − M ) .
Proof. Let F ( k ) = ( N − d k ) ( N − e k ) and j ( k ) = ω ( ( F ( k ) , P z ) ) . Bonferroni inequalities through the adjacent orders r , r + 1 , with r the least even integer at least K l , bracket 1 j ( k ) = 0 . For each squarefree m ∣ P z , CRT gives
# { k ∈ I d , e : m ∣ F ( k ) } = H d , e ρ ( m ) / m + O ( ρ ( m ) ) ,     ρ ( m ) = ∏ p ∣ m ρ p .
All divisors in either bracket are at most z r + 1 , and ρ ( m ) ≤ 2 r + 1 . The total endpoint error is at most 2 r + 1 z r + 1 = e x p ( O K ( L / l ) ) . Also ∑ p ≤ z ρ p / p ≪ l . The tail of the elementary symmetric expansion of ∏ p ≤ z ( 1 − ρ p / p ) is at most
2 e C 0 l r + 1 r + 1
when K is sufficiently large. Choose fixed K after M so that this is smaller than the required logarithmic reserve. Multiplying by V − 2 ≪ L 2 pays both errors by O M ( Y L − M ) . The main term is H d , e C z ; use Lemma D.1. This proof is uniform even when the real interval is very short. ▫
Lemma D.3. A finite Selberg normalizer. Let ρ p be the root densities just defined and fix R > 1 . Let D R be the squarefree integers s ≤ R whose prime factors have ρ p > 0 , and write
g ( s ) = ρ ( s ) / s ,   h ( s ) = ∏ p ∣ s g ( p ) 1 − g ( p ) ,   G ( R ) = ∑ s ∈ D R h ( s ) ,   h ( 1 ) = 1 .
There are real weights λ 1 = 1 , supported on D R , such that
∑ s , t λ s λ t g ( [ s , t ] ) = G ( R ) − 1 ,     | λ s | ≤ s .
Consequently, if R < x , the number P d , e ( I ) of actual prime pairs on I is at most H / G ( R ) + O ( R 6 ) .
Proof. Zero-density primes have been excluded before any division by g . For y r = ∑ r ∣ s λ s g ( s ) the multiplicative identity
g ( [ s , t ] ) = g ( s ) g ( t ) ∑ r ∣ ( s , t ) h ( r ) − 1
diagonalizes the quadratic form as ∑ r y r 2 / h ( r ) . Finite Möbius inversion gives λ 1 = ∑ r μ ( r ) y r . The minimum under λ 1 = 1 is therefore G ( R ) − 1 , attained at y r = μ ( r ) h ( r ) / G ( R ) . Inverting once more gives
λ s = μ ( s ) ∏ p ∣ s ( 1 − g ( p ) ) − 1 ∑ b ≤ R / s ,   ( b , s ) = 1 ,   b ∈ D R h ( b ) G ( R ) .
The fraction is at most one; since p / ( p − ρ p ) ≤ p , | λ s | ≤ s . At a prime pair both prime values are at least x > R , so ∑ s ∣ F ( k ) λ s = 1 . Squaring majorizes the prime-pair indicator. CRT and half-open interval counting give the main term H / G ( R ) and an error at most
∑ s , t ≤ R | λ s λ t | ρ ( [ s , t ] ) ≤ ∑ s , t ≤ R s 2 t 2 ≪ R 6 .
□
Theorem D.4. Balanced two-prime upper bound. For every fixed ϵ > 0 , uniformly in all the above distinct balanced rows and real intervals,
A A d , e ( I ) ≪ ϵ , c , C H C ∞ ( N ; d , e ) + x ϵ .
Proof. Put R = x δ , y = R 1 / 80 , with δ = m i n ( 1 / 100 , ϵ / 12 ) . The full finite product of h over primes at most y is
Z = ∏ p ≤ y ( 1 + h ( p ) ) = ∏ p ≤ y ( 1 − ρ p / p ) − 1 = V ( y ) − 2 / S y ,
where zero-density factors equal one. After normalization by Z , the product has independent inclusion probabilities ρ p / p . It is a probability law on explicit squarefree products, not a model for prime pairs. The elementary estimate ϑ ( t ) ≤ 4 t gives
E l o g s = ∑ p ≤ y ρ p l o g p / p ≤ 20 l o g y = 1 / 4 l o g R .
For completeness, ϑ ( n ) ≤ ϑ ( ⌈ n / 2 ⌉ ) + n l o g 2 follows from n ⌊ n / 2 ⌋ ≤ 2 n ; induction gives the stated bound, and partial summation gives ∑ p ≤ y l o g p / p ≤ 4 + 4 l o g y ≤ 10 l o g y . Markov’s inequality retains at least 3 / 4 of Z on products s ≤ R . These are included in G ( R ) , even though other primes may also occur in G . The Euler expansion of V ( y ) − 1 contains ∑ n ≤ y 1 / n , so V ( y ) − 1 ≥ l o g y . By Lemma D.1, S y ≤ 2 C ∞ eventually, uniformly in the rows. Hence
G ( R ) ≥ 3 8 ( l o g y ) 2 C ∞ .
Lemma D.3 bounds the unweighted prime-pair count by O δ ( H C ∞ / L 2 + R 6 ) . The logarithmic pair weight is at most l o g 2 ( 3 x / 2 ) . Finally 6 δ ≤ ϵ / 2 , so L 2 R 6 ≪ ϵ x ϵ . ▫
Proposition D.5. A sufficient covariance condition. Let D ⊂ ( c Y , C Y ] consist of rows coprime to N , and let K ⊂ ( c Y , C Y ] be a real interval. Define
G N ( d , e ) = ∑ k ∈ K w d k w e k ,     B N ( α , β ) = ∑ d ∈ D , k ∈ K α d β k w d k ,
where support in ( x / 2 , x ] is included in w . Fix F ≥ 0 , choose an integer s ≥ 2 F , and put h = 2 s . If, for a fixed A > 0 ,
∑ d ≠ e τ ( d ) F τ ( e ) F | G N ( d , e ) | ≪ A , F x 3 / 2 L − ( 2 A + h + 2 ) ,
then | B N ( α , β ) | ≪ A , F x L − A for all | α d | ≤ τ ( d ) F , | β k | ≤ τ ( k ) F . The displayed covariance hypothesis is an assumption.
Proof. Cauchy in k bounds | B N | 2 by
∑ k ∈ K τ ( k ) 2 F ∑ d , e α d α e ¯ G N ( d , e ) .
The first factor is O F ( Y L h − 1 ) . The quadratic form is real and nonnegative, as follows from its original sum-of-squares form. The weighted diagonal is O F ( x L C F ) , since | w | ≪ L C 0 and there are O ( Y ) columns for each row. Its resulting contribution to | B N | is O F ( x 3 / 4 L C ′ F ) , smaller than x L − A . The assumed off-diagonal estimate pays the other term. At distinct rows the exact identity is G N = A A − A B − B A + B B , with the same interval and zero-support conventions. Thus neither Theorem D.2 nor the uncentered upper bound in Theorem D.4 proves the assumed cancellation. ▫

Appendix E. Selected Progressions and Exact Quadratic Identities

This section records reductions with complete centering terms. They are useful statements of the remaining arithmetic problem.
Definition E.1. For squarefree m ∣ P z with ( d m , N ) = 1 let
A ( d , m ) = { a   ( m o d   d m ) : ( a , d m ) = 1 ,   ( N − a , d m ) = d } .
Only distinct pairs ( q , a ) with q = d m are counted in selected norms.
Lemma E.2. Exact recovery and multiplicity. Every retained residue in (8.1) determines d = ( N − a , q ) and m = q / d . Conversely, an element of A ( d , m ) determines the other slope class
e ≡ N ( ( N − a ) / d ) − 1   ( m o d   m ) .
Moreover
| A ( d , m ) | = ∏ p ∣ m ,   p ∣ d ( p − 1 ) ∏ p ∣ m ,   p ∤ d ( p − 2 ) .
Proof. The chosen k 0 = N e − 1   ( m o d   m ) is a unit, so ( N − a , d m ) = ( d k 0 , d m ) = d . Conversely this gcd identity makes ( N − a ) / d a unit modulo m and permits the displayed inversion. The admissible slope classes are precisely the units e   ( m o d   m ) with ( e − d , m ) = 1 . At a prime dividing d only the zero class is forbidden; otherwise the two distinct classes 0 , d are forbidden. CRT proves the product, including m = 1 by empty-product conventions. ▫
Proposition E.3. A selected anchored variance is sufficient.
Let m ≤ M 0 = x o ( 1 ) range over a fixed-accuracy Bonferroni collection. Put
Δ 0 ( q , a ) = s u p x ≤ U ≤ 3 x / 2 θ < ( U ; q , a ) − θ < ( x ; q , a ) − U − x φ ( q ) .
For each fixed divisor envelope F , if
∑ s e l e c t e d   ( q , a ) Δ 0 ( q , a ) 2 ≪ B x 3 / 2 L − B   for   every   fixed   B > 0 ,
then the balanced weighted mixed estimate ∑ d ≠ e τ ( d ) F τ ( e ) F | A B d , e − H d , e C ∞ | ≪ A , F x 3 / 2 L − A follows for every fixed A . The variance estimate is not asserted here.
Proof. Recover d , m using Lemma E.2. The total weight of the other slope class at ( q , a ) is
W ( q , a ) = τ ( d ) F ∑ e ≡ e 0   ( m ) τ ( e ) F ,
with the actual row restrictions retained. Since m = x o ( 1 ) < Y , each class has O ( Y / m ) representatives. Cauchy within each class and the divisor moment bound imply, for some fixed C F ,
∑ s e l e c t e d W ( q , a ) 2 ≪ F Y 3 L C F ∑ m m − 1 ≪ F Y 3 L C F + 1 .
Here ∑ m 1 / m ≤ ∏ p ≤ z ( 1 + 1 / p ) ≤ V − 1 ≪ L . Every actual interval error is at most 2 Δ 0 ( q , a ) , including when its endpoints depend on e . Hence its aggregate is bounded by
2 ∑ W Δ 0 ≪ F Y 3 / 2 L C ′ F ∑ Δ 0 2 1 / 2 .
The assumed variance, the factor V − 1 , and the deterministic sieve completion error give the claim with sufficient fixed log reserves. ▫
Proposition E.4. Exact selected interval variance. Choose arbitrary class-dependent intervals B a = [ A a , B ′ a ) in [ x , 3 x / 2 ) , and write h a = B ′ a − A a , T a = ∑ p ∈ B a , p ≡ a   ( q ) l o g p , and c a = h a / φ ( q ) . Then
∑ a ( T a − c a ) 2 = ∑ a ∑ p ∈ B a , p ≡ a   ( q ) ( l o g p ) 2   + ∑ a ∑ p 1 ≠ p 2 ∈ B a p 1 ≡ p 2 ≡ a   ( q ) l o g p 1 l o g p 2 − 2 ∑ a c a T a + ∑ a c a 2 .
The first term is O ( x M 0 L 2 ) = x 1 + o ( 1 ) uniformly in all interval choices. If an anchored supremum is used, its near-maximizing interval must still be chosen separately for each class.
Proof. The identity is the finite expansion of each centered square. For fixed d , m , selected classes partition a subset of the primes with p ≡ N   ( m o d   d ) . There are O ( x / d + 1 ) = O ( Y ) integers in this progression in the ambient interval. There are O ( Y ) rows and at most M 0 moduli, and each squared log weight is O ( L 2 ) . This proves the diagonal estimate without any prime-distribution theorem. Changing the interval in each finite summand does not change that argument; it does not remove the off-diagonal term or either subtraction. ▫
Lemma E.5. Signed truncation with controlled prime weight. For each prescribed A > 0 , choose a sufficiently large fixed K and s ≍ K l . With the actual signed remainders of Appendix A, at distinct balanced rows,
E d , e , s = V ( A B d , e − H d , e C ∞ ) + O A ( Y L − A ) .
Proof. Let P s ( j ) = ∑ i = 0 s ( − 1 ) i j i . The binomial identity gives | P s ( j ) − 1 j = 0 | ≤ j s + 1 . Apply it with j = ω ( ( N − e k , P z ) ) and multiply by the nonnegative prime log weight, bounded by O ( L ) . Expanding the binomial as a sum of squarefree divisors with exactly s + 1 factors and counting each linear congruence yields an error at most
O L H ( ∑ p ≤ z 1 / p ) s + 1 ( s + 1 ) ! + L z s + 1 = O A ( Y L − A ) .
Stirling’s bound and sufficiently large fixed K pay the first term; z s + 1 = x o ( 1 ) pays the second. On actual prime support all inadmissible divisors have zero contribution. The finite main term completion in (A.4), with sufficient log reserve, then proves the equality. Thus renaming this signed remainder supplies no independent estimate for the balanced mixed error. ▫
Proposition E.6. Support obstruction for factorable weights. If a sequence is r -fold well-factorable at level Q , meaning it admits a supported convolution for every splitting Q = Q 1 … Q r , then it vanishes at every integer containing a prime p > Q 1 / r . The same holds for any finite linear combination of such weights at levels at most Q .
Proof. Choose the equal splitting Q i = Q 1 / r . In any factorization of an integer divisible by p , one of the r factors contains p , and therefore exceeds its support. Every convolution summand is zero. Taking a finite linear combination preserves a zero coordinate. ▫
This is a statement about the specified functionals, not a lower bound for actual prime errors or an exclusion of other dispersion arguments.

Appendix F. Balanced Prime Products with a Prime Remainder

This is a distinct almost-all result for N = p q + r ; its composite summand p q must not be identified with a binary Goldbach representation. Put I T = ( 2 T , 101 T / 50 ] ∩ 2 Z and L = l o g T .
Definition F.1. Let P = ( a T , b T ] and Q = ( c T , d T ] , where 0.59 ≤ a ≤ b ≤ 0.68 and 0.89 ≤ c ≤ d ≤ 1.02 . Define
J P = ∫ P d u l o g u ,   J Q = ∫ Q d v l o g v ,   R P , Q ( N ) = ∑ p ∈ P ,   q ∈ Q p , q   prime l o g ( N − p q ) 1 N − p q   prime .
For each application of a fixed-box estimate, these intervals are held fixed as N varies; their endpoints may depend on T .
Theorem F.2. Fixed-box prime-product mean square. For every fixed A > 0 , uniformly over the indicated boxes,
∑ N ∈ I T | R P , Q ( N ) − S ( N ) J P J Q | 2 ≪ A T 3 L − A .
Here S ( N ) = 2 C 2 ∏ p ∣ N , p > 2 ( p − 1 ) / ( p − 2 ) is the usual binary singular series. The assertion is not a supremum over boxes inside the sum.
Proof. The geometry gives 0.5251 T < p q ≤ 0.6936 T and 1.3064 T < N − p q < 1.4949 T . Thus the fixed remainder interval R = ( 1.3 T , 1.5 T ] contains every possible prime remainder. Set
A ( α ) = ∑ r ∈ R , r   prime l o g r   e ( α r ) ,   B ( α ) = ∑ p ∈ P , q ∈ Q   prime e ( α p q ) .
Their product has Fourier coefficient R P , Q ( N ) . Because the intervals P , Q are disjoint, unique factorization makes the map ( p , q ) ↦ p q injective. Orthogonality and the ordinary prime number theorem give ∫ 0 1 | B | 2 ≪ T / L 2 .
Choose D > 10 , H = ⌈ L D ⌉ , and major arcs | α − h / m | ≤ H / T with ( h , m ) = 1 , m ≤ H . They are disjoint for large T . On the minor arcs, Dirichlet approximation with K = ⌈ T / H ⌉ gives H < m ≤ K and | α − h / m | ≤ 1 / ( m K ) ≤ m − 2 : if m ≤ H , then 1 / K ≤ H / T puts α in a major arc. Vaughan’s bound [9, Theorem 17.1], weakened to log power four, applied to the two remainder prefixes gives
s u p m | A ( α ) | ≪ D T L 4 − D / 2 .
The proper powers cost O ( T L 2 ) and are absorbed. Parseval therefore bounds the minor-arc squared error by O D ( T 3 L 6 − D ) .
For m ≤ H define
V ( β ) = ∫ R e ( β r )   d r ,     W ( β ) = ∫ P ∫ Q e ( β u v ) l o g u l o g v   d u   d v .
Siegel–Walfisz and partial summation give, for | β | ≤ H / T ,
A ( h / m + β ) = μ ( m ) φ ( m ) V ( β ) + O D , E ( T L − E ) , B ( h / m + β ) = μ ( m ) φ ( m ) W ( β ) + O D , E ( T L − E )
for every fixed E . For the second formula first fix p , sum q in unit classes modulo m , and use ∑ b   ( m ) * e ( h p b / m ) = μ ( m ) , since p > m . Summing the error over p costs at most O ( T H 2 L − E ′ ) . Apply ordinary PNT and partial summation to the remaining p integral; its variation is O ( T ( 1 + | β | T ) / L ) . Choosing E ′ after D , E pays these costs uniformly in the boxes.
The total major-arc measure is O ( H 3 / T ) . Thus their contribution is
S H ( N ) I H ( N ) + O D , E 0 ( T L − E 0 ) ,   S H ( N ) = ∑ m ≤ H μ ( m ) 2 c m ( N ) φ ( m ) 2 ,
where I H integrates V W e ( − N β ) over [ − H / T , H / T ] . Integration by parts gives
| V ( β ) | ≪ T 1 + T | β | ,   | W ( β ) | ≪ T L 2 ( 1 + T | β | ) .
Fourier inversion for the integrable product, using the above geometric inclusion, gives I ∞ ( N ) = J P J Q , and its tail is O ( T / ( L 2 H ) ) . Also | S H ( N ) | ≪ l o g ( 2 H ) , by ∑ m ≤ H 1 / φ ( m ) ≪ l o g ( 2 H ) .
For clarity, the singular-series completion is averaged rather than pointwise. Rankin’s inequality bounds its absolute tail by H − 1 / 2 F ( N ) , where
F ( N ) = ∑ m ≥ 1 μ ( m ) 2 m   | c m ( N ) | φ ( m ) 2 ≤ C ∏ p ∣ N 1 + p p − 1 = C D ( N ) .
Indeed the unchanged good-prime Euler product has excess O ( p − 3 / 2 ) . Write D ( N ) = ∑ d ∣ N g ( d ) with squarefree support and g ( p ) = p / ( p − 1 ) . Positivity and ⌊ 3 T / [ d , e ] ⌋ ≤ 3 T / [ d , e ] give
∑ N ≤ 3 T D ( N ) 2 ≤ 3 T ∏ p 1 + 2 g ( p ) + g ( p ) 2 p ≪ T .
The product converges, so ∑ I T | S − S H | 2 ≪ T / H . As J P J Q ≪ T / L 2 , the total mean-square error is
O T 3 L 6 − D + L − 2 E 0 + l o g 2 ( 2 H ) L 4 H 2 + 1 L 4 H .
Take D = E 0 = A + 10 . Every source precision above was chosen after these fixed exponents, completing the proof. ▫
Corollary F.3. Moving balanced boxes. For x N = N / 2 let
P N = ( 3 x N / 5 , 2 x N / 3 ] ,     Q N = ( 9 x N / 10 , x N ] .
For each fixed B , C > 0 , outside O B , C ( T / L B ) targets in I T , one has
R P N , Q N ( N ) = S ( N ) J P N J Q N + O B , C ( T / L C ) .
In particular, outside such a set with suitable fixed C ,
# { ( p , q ) ∈ P N × Q N : p , q , N − p q   prime } = 2 75 S ( N ) x N ( l o g x N ) 3 1 + O B ( ( l o g x N ) − 1 ) .
Proof. Chebyshev applied to Theorem F.2 supplies any fixed logarithmic exceptional saving for each fixed pair of boxes. To retain moving endpoints, mesh t = N / ( 2 T ) ∈ [ 1 , 1.01 ] at spacing L − K , with fixed K > C + 2 . On each cell [ t j , t j + 1 ] , define inner intervals by taking the lower endpoints at t j + 1 and the upper endpoints at t j in the displayed formulas; use the reverse choices for outer intervals. They lie in the permitted compact box family and bracket the exact moving intervals, respecting all strict endpoints. There are O ( L K ) such tests. Apply the fixed-box consequence with exceptional exponent B + K + 2 and error T / L C + 1 , and take their finite union. Its size is O ( T / L B ) . The summands are nonnegative, so the inner and outer counts bracket the moving count. Each product J P J Q changes by O ( T L − K − 2 ) ; S ( N ) ≪ L makes the main-term displacement O ( T L − K − 1 ) . This proves the first assertion. Direct integration gives J P N J Q N = ( 2 / 75 ) x N ( l o g x N ) − 2 ( 1 + O ( 1 / l o g x N ) ) . Finally l o g ( N − p q ) = l o g x N + O ( 1 ) uniformly, so positivity permits removal of the prime log weight with relative error O ( 1 / l o g x N ) . Take, for example, C = 4 and use S ( N ) ≥ 2 C 2 > 0 . ▫
Corollary F.4. One actual prime-mask comparison. For the moving boxes in Corollary F.3 and each fixed B , F > 0 , outside O B , F ( T / L B ) targets,
∑ p ∈ P N , q ∈ Q N   prime ( a p q − b p q ) ≪ B , F x N ( l o g x N ) − F .
Proof. The prime products lie in ( x N / 2 , x N ] . Apply the inner comparison (R1) in the proof of Theorem B.4 with m = p and the actual interval of q ; it is a subinterval of ( Y / 2 , Y ] , Y = x N / p , and p > z . Summing its errors costs at most x N ∑ p ∈ P N p − 1 times an arbitrary negative log power. PNT for the two boxes and S z / S = 1 + O ( l 2 / z ) therefore give ∑ p , q b p q = S ( N ) J P N J Q N + O F ( x N ( l o g x N ) − F ) for each target. Subtract Corollary F.3 with sufficient precision. This tests one prime-mask family only; its exceptional set is retained. ▫

Appendix G. A Conditional Fixed Margin Implication for Real Zeros

This section states the hypothesis on actual Goldbach counts explicitly. It is not supplied by the preceding lower bounds outside exceptional sets.
Definition G.1. Let r 2 ( N ) = ∑ n = 1 N − 1 Λ ( n ) Λ ( N − n ) be the full ordered von Mangoldt convolution. For a conductor f and X > 0 let G f ( X ) = { N ∈ [ X / 2 , X ] ∩ 2 Z : f ∣ N } .
Theorem G.2. Conditional fixed margin exclusion. Fix A > 5 / 2 , 0 < δ < 1 , and ξ > 0 , and put X = f A . There is a sufficiently small fixed h δ > 0 with the following property for sufficiently large f . Suppose at least a proportion ξ of G f ( X ) satisfies the external band
δ S ( N ) N ≤ r 2 ( N ) ≤ ( 2 − δ ) S ( N ) N .     ( FM )
Then no primitive real nonprincipal character of conductor f has a real zero β with 0 < ( 1 − β ) l o g X ≤ h δ . For odd characters only the lower half of (FM) is needed; for even characters only the upper half is needed.
Proof. We first specify the cited analytic inputs, so the scope of the deduction can be assessed. Pintz’s original explicit formula [11, Theorem 1, Sections 4–5 and 9–11] uses a large fixed H , a source-selected P ∈ [ P 0 X − ε * , P 0 ] with P 0 ≤ X 4 / 9 − ε * , and the pole together with primitive zeros of conductor at most P in
R ρ ≥ 1 − H / l o g X ,     | I ρ | ≤ X .
Its major-arc expression has an error O ( S ( N ) X e − c H ) + O H ( X 1 − ε 0 ) after the fixed arc reserves are chosen, with ε 0 > 0 fixed. The first constant is the fixed-reserve constant in the source’s tail bound; no uncontrolled C ( H ) e − c H is used. Its original minor-arc estimate is
∑ N | R 2 ( N ) | 2 ≪ ( X 3 / P + X 13 5 ) ( l o g X ) 9 .    
These classical analytic theorems and zero-repulsion lemmas are cited inputs, not claims of a new proof of their full analytic machinery.
Choose P 0 = X 21 / 50 and fixed 0 < ε * < 1 / 100 . Then every allowed selected P exceeds X 2 / 5 > f . If u = ( 1 − β ) l o g X ≤ h , the source’s zero-free region, Landau uniqueness and Deuring–Heilbronn repulsion clear the retained set to the pole and β , for fixed large H followed by fixed small h . The repulsion is used over the full retained height range: with another zero 1 − ϵ + i γ and conductor q ≤ P , the source uses
Y 0 = ( f 2 q   c o n d ( χ χ q ¯ ) ( | γ | + 2 ) 2 ) 3 / 8 ≪ ( P 5 X ) 3 / 8 ≤ X 29 / 24 .
Also Y 0 ≥ f 3 / 4 . The fixed-parameter repulsion conclusion [11, Section 11, equations (11.3)–(11.5)] gives ϵ l o g X > ( 4 / 5 ) l o g ( 1 / ( 2 h ) ) for all other retained zeros, once h is sufficiently small. Choosing the right side greater than H , and meeting the fixed uniqueness threshold, removes them.
We record the aligned arithmetic exactly. Write τ q ( χ ) = ∑ a   ( q ) χ ( a ) e ( a / q ) for the Gauss sum of the induced character on modulus q . For the real characters used here, the generalized series is normalized as
S ( χ 1 , χ 2 , N ) = ∑ [ f 1 , f 2 ] ∣ q τ q ( χ 1 ) τ q ( χ 2 ) φ ( q ) 2 ∑ a   ( q ) * χ 1 ( a ) χ 2 ( a ) e ( − N a / q ) ,
with f i the primitive conductors; the principal primitive character has conductor one. In the aligned cases below the evaluated series converge absolutely, or vanish term by term. For f ∣ N , the mixed generalized singular series obey
S ( 1 , χ , N ) = S ( χ , 1 , N ) = 0 ,     S ( χ , χ , N ) = χ ( − 1 ) S ( N ) .
For the first identity the principal Gauss factor is μ ( q ) , which kills nonsquarefree q . For squarefree q = f s , ( f , s ) = 1 , CRT makes the conductor component an unweighted sum of the nonprincipal character, equal to zero. If f is nonsquarefree every multiple was already killed. For the second identity the induced Gauss formula restricts q to f s with squarefree s coprime to f , and τ ( χ ) 2 = χ ( − 1 ) f . The local term becomes
χ ( − 1 ) f φ ( f ) μ ( s ) 2 c s ( N ) φ ( s ) 2 .
The factors away from f form the principal series, and the missing factors at primes dividing f ∣ N multiply to f / φ ( f ) . Absolute convergence for fixed N justifies the Euler product and also covers primitive real conductors divisible by 4 or 8 .
Writing d = 1 − β , σ = χ ( − 1 ) and B ( s , t ) = Γ ( s ) Γ ( t ) / Γ ( s + t ) , the pole–zero formula and (Z2) give, on the conductor grid,
R 1 ( N ) = S ( N ) N { 1 + σ B ( 1 − d , 1 − d ) N − 2 d } + O ( S ( N ) X e − c H ) + O H ( X 1 − ε 0 ) .
Uniformly for N ∈ [ X / 2 , X ] and 0 < u ≤ h , B ( 1 − d , 1 − d ) N − 2 d = 1 + O ( u ) by differentiability at ( 1 , 1 ) and | N − 2 d − 1 | ≤ 2 d l o g N . At a fixed tolerance v δ X , (Z1) and P > X 2 / 5 leave at most O δ ( X 3 / 5 ( l o g X ) 9 ) bad targets.
The source truncates ordinary primes below X 1 = X 1 − ε c u t . Restoring these coordinates costs at most 2 ( X 1 + 1 ) ( l o g X ) 2 per target. Restoring proper powers costs O ( X ( l o g X ) 3 ) . Both are o ( X ) , so the full convolution is compared in both directions. Choose fixed H after δ , then fixed h after H , δ , then v δ , and finally large f . Since S ( N ) ≥ 2 C 2 and N ≥ X / 2 , the resulting normalized count outside the bad set satisfies
r 2 ( N ) S ( N ) N − ( 1 + σ ) < δ / 2 .
The number of grid points is
| G f ( X ) | = X 2 l c m ( 2 , f ) + O ( 1 ) ≍ X / f .
The bad-set proportion is O δ ( X 1 / A − 2 / 5 ( l o g X ) 9 ) = o ( 1 ) because A > 5 / 2 . Thus the positive proportion in (FM) contains a good point, contradicting its lower band when σ = − 1 and its upper band when σ = + 1 . This proves the implication with the stated parameter order. ▫
Remark G.3. The arithmetic band (FM) has not been established by this article. A positive lower floor tending to zero after normalization is insufficient. No parameters here are moved with H → ∞ or δ → 0 , and no new unconditional zero-free region is claimed.

Appendix H. Paired Congruences and Local Ambiguity

These finite statements describe integer geometry. The target sum in this section is 2 N − ( d + e ) k , which need not equal the anchor N .
Definition H.1. Let d , e be positive integers, g = ( d , e ) , a = e / g , b = d / g , L = [ d , e ] = d e / g , and let K = [ A , B ] ∩ Z be a nonempty contiguous integer interval. Put p k = N − d k , q k = N − e k . An alternate at base k is a pair ( p k 1 , q k 2 ) with k 1 , k 2 ∈ K and p k 1 + q k 2 = p k + q k .
Lemma H.2. Generalized CRT and exact offsets. For any integer pair p + q = p k + q k , the conditions p ≡ N   ( m o d   d ) , q ≡ N   ( m o d   e ) are equivalent to p ≡ p k   ( m o d   L ) . Every alternate therefore has the form
p ′ = p k + L t ,   q ′ = q k − L t ,     k 1 = k − a t ,   k 2 = k + b t ,     t ∈ Z .
Proof. The two differences p − p k = − ( q − q k ) are divisible by d and e exactly when they are divisible by L . Substituting into p = N − d k 1 and q = N − e k 2 gives the displayed indices, and conversely. This least-common-multiple modulus concerns paired congruences; it does not replace the product modulus d m in the sieve-row transformation. ▫
Theorem H.3. Uniform paired uniqueness. Every base k ∈ K has only the zero-offset alternate if and only if B − A < ( d + e ) / g .
Proof. Let T 0 = a + b . If t ≠ 0 the two alternate indices are separated by T 0 | t | ≥ T 0 , impossible in an interval of diameter less than T 0 . Conversely, when B − A ≥ T 0 , take k 1 = A , k = A + a , k 2 = A + T 0 . All three belong to K , and Lemma H.2 with t = 1 gives a distinct alternate. ▫
Lemma H.4. Exact local filter. Assume in addition that N is even, ( d e , N ) = 1 , 3 ∤ N d e , and the base values p k , q k are units modulo 6 . Then both alternate values are units modulo 6 if and only if 6 ∣ t .
Proof. The slopes and L are odd. Preservation of parity forces t even. Modulo 3 , either d = e , in which case p k = q k , or d = − e . In the latter case, if k ≠ 0   ( m o d   3 ) , the two values N − d k , N + d k include zero; thus local admissibility forces k = 0 and again p k = q k ≠ 0   ( m o d   3 ) . Shifting two equal nonzero residues by opposite nonzero amounts modulo 3 makes one zero. Since L is a unit modulo 3 , preservation forces 3 ∣ t . Conversely 6 ∣ t preserves both base values modulo 6 . ▫
Proposition H.5. Exact phase-sensitive count. Under Lemma H.4, suppose d ≤ e and put
P = [ A + 6 a , B − 6 b ] ∩ Z ,   Q = [ A + 6 b , B − 6 a ] ∩ Z ,
R 6 = { r   ( m o d   6 ) : r   odd ,   r ≢ N d − 1 , N e − 1   ( m o d   3 ) } .
For an integer interval J = [ u , v ] ∩ Z let
C ( J ; R 6 ) = ∑ r ∈ R 6 m a x 0 , v − r 6 − u − r 6 + 1 ,
with empty intervals contributing zero. The exact number of locally admissible bases with a supported nonzero locally admissible alternate is C ( P ; R 6 ) + C ( Q ; R 6 ) − C ( P ∩ Q ; R 6 ) .
Proof. The base values are units modulo 6 precisely for indices in R 6 . By Lemma H.4, a positive alternate offset has t = 6 s , s ≥ 1 . If its two indices fit in K , the nearer indices for t = 6 also fit; their bases are exactly P . Negative offsets give Q by the same argument. Inclusion–exclusion and literal residue counting give the formula. None of these local tests asserts primality. ▫
Remark H.6. Finite checks and their scope. The integer instance N = 20000 , d = 21 , e = 31 has common-support K = { 239 , … , 322 } . At k = 270 , the pairs ( 14330 , 11630 ) and ( 14981 , 10979 ) have the same sum 25960 ; their indices exhibit the nonzero offset in Theorem H.3. They are integer witnesses, not claimed prime pairs. The local equivalence and phase count also have preserved bounded residue checks (24 and 5208 setups, respectively); their general validity follows from the proofs above, not from these finite checks.

Appendix I. A Sharp Prime Boundary and Its Exact Unit Mean

We now isolate a sharply truncated bilinear component. The parameters in this and the following three appendices are local to these appendices; in particular, their D is different from a major-arc denominator cutoff. The statements below concern the explicitly defined component. No equivalence between a bound for this component alone and the binary Goldbach conjecture is asserted.
Definition I.1. Boundary coefficients. Let N = ∏ p ≤ y p , with y → ∞ , and put
D = N / 3 ,   K = N / 2 ,   L = l o g ( 4 D ) ,     C = { q ∈ N : 9 D / 10 < q ≤ D ,   ( q , N ) = 1 } .
Thus K and all q ∈ C are odd. Fix 0 < κ < 1 / 448 and set
a 0 = 4 κ ,   b 0 = 1 / 56 − 4 κ ,   c 0 = 1 / 6 − 4 κ ,     z 0 = D a 0 ,   z 1 = D c 0 .
Let G 1 and G 2 consist of the primes in [ D a 0 , D b 0 ] and [ D a 0 , D c 0 ] , respectively. For a finite prime set G , define A G ( n ) = 1 p ∤ n   for   every   p ∈ G . On odd positive integers put
a ( r ) = μ ( r ) A G 1 ( r ) ,     η ( u ) = A G 2 ( u ) 1 z 1 / 2 < u ≤ z 1 − 1 z 0 / 2 ≤ u < z 0 , δ ( s ) = ∑ p u = s p > D 2 / 5   prime l o g p   η ( u ) .
Fix intervals I a ⊂ [ D / 2 , 4 D ] and J b ⊂ [ D / 4 , 2 D ] , and define
I ( q ) = [ K − 11 D q / 20 , K − D q / 2 ) ,     T q ( t ) = 1 t ≡ K   ( q ) − 1 ( t , q ) = 1 φ ( q ) .
For a right coefficient f , let S f ( q ) be the sum of a ( r ) f ( s ) over odd r ∈ I a , s ∈ J b with r s ∈ I ( q ) and r s ≡ K   ( m o d   q ) . Let E f ( q ) be 1 / φ ( q ) times the same box sum with the congruence replaced by ( r s , q ) = 1 . Write Q δ = S δ − E δ and L 0 = S 1 − E 1 .
All assertions in these appendices hold for sufficiently large D , with thresholds allowed to depend on the fixed κ . The two cofactor slabs are then disjoint. The lower slab has A G 2 ( u ) = 1 . The relevant finite harmonic sums are
l q = ∑ z 0 / 2 ≤ u < z 0 u   odd ,   ( u , q ) = 1 1 u , v q = ∑ z 1 / 2 < u ≤ z 1 u   odd ,   ( u , q ) = 1 A G 2 ( u ) u , H q = v q − l q , A q = v q + l q , B q = ∑ u   odd ( u , q ) = 1 η ( u ) 2 l o g u u , ρ q = φ ( q ) 2 q ,     c ( q ) = H q ρ q .
Lemma I.2. Rough moduli and unique prime labels. Uniformly for q ∈ C , one has q / φ ( q ) = 1 + o ( 1 ) , l q = ( l o g 2 ) / 2 + o ( 1 ) , 0 ≤ v q ≤ ( l o g 2 ) / 2 + o ( 1 ) , and c ( q ) = O ( 1 ) . For s ∈ [ D / 4 , 2 D ] , at most one pair ( p , u ) contributes to δ ( s ) . Every such prime satisfies
D / ( 4 z 1 ) ≤ p ≤ 4 D / z 0 ,     p > 2 D ,   p < q ,   p > u ,   p ∉ G 1 .
Proof. Every prime divisor of q exceeds y . The prime number theorem gives y ≍ l o g D , whence
ω ( q ) ≤ l o g q l o g y ,     ∑ p ∣ q 1 p ≤ l o g q y l o g y = O ( 1 / l o g y ) .
The Euler product proves the totient assertion. In a slab of length comparable to Z , the reciprocal mass removed by multiples of a prime p is O ( 1 / p + 1 / Z ) . Summing over p ∣ q removes o ( 1 ) , since Z ≥ z 0 / 2 is a fixed positive power of D . Odd harmonic sums on either slab are ( l o g 2 ) / 2 + O ( 1 / Z ) . This proves the mass assertions and boundedness of c ( q ) . The displayed prime bounds follow from s = p u and z 0 / 2 ≤ u ≤ z 1 . All further inequalities follow from 0 < a 0 < b 0 < 1 − c 0 and c 0 < 1 / 6 . Two distinct eligible primes would have product greater than 2 D , so cannot both divide s . For each prime the cofactor is determined. This proves uniqueness. ▫
Theorem I.3. Marginal moments and the optimal constant subtraction.
For every fixed T > 0 and every interval J ⊂ [ D / 4 , 2 D ] , uniformly in its endpoints and in q ∈ C ,
∑ s ∈ J ,   s   odd s , q = 1 δ s = J H q + O T D L T .
If J = ( x , z ] , the second moment is
∑ s ∈ J ,   s   odd s , q = 1 δ ( s ) 2 = [ z l o g z − x l o g x − | J | ] A q − | J | B q + O T ( D / L T ) .  
For endpoints in this compact range with | J | ≥ ϵ D , where ϵ > 0 is fixed, these imply
∑ s ∈ J ,   s   odd s , q = 1 δ ( s ) 2 = | J | l o g D { ( 1 − a 0 ) l q + ( 1 − c 0 ) v q } + O ϵ ( D ) ,  
# { s ∈ J : s   odd ,   ( s , q ) = 1 ,   δ ( s ) ≠ 0 } = | J | l o g D l q 1 − a 0 + v q 1 − c 0 + O ϵ ( D / l o g 2 D ) .  
In particular, the least possible squared norm after subtracting a constant is of order D l o g D :
m i n c ∈ C ∑ s ∈ J ,   s   odd s , q = 1 | δ s − c | 2 ≍ κ , ϵ D l o g D .
Also ∑ s ∈ J ,   s   odd , ( s , q ) = 1 | δ ( s ) | ≪ D .
Proof. Use the classical prime number theorem with an arbitrary fixed logarithmic precision. Partial summation gives
θ ( t ) = t + O B ( t / l o g B t ) ,     ∑ p ≤ t ( l o g p ) 2 = t l o g t − t + O B ( t / l o g B t ) .
The usual estimate for π ( t ) − l i ( t ) follows as well. All endpoints x / u , z / u are between fixed positive powers of D . Expanding δ and applying the two prefix estimates costs O B ( D / L B ) after summation, since ∑ | η ( u ) | / u = O ( 1 ) . The same argument with | η | proves the absolute first-moment bound. Primes dividing q must be removed: at most O ( z 1 ) cofactors occur for each such prime, so their first- and second-moment costs are O ( z 1 L 2 ) and O ( z 1 L 3 ) , respectively. Both are smaller than D / L T for each fixed T .
Lemma I.2 makes the expansion of the square exactly diagonal in ( p , u ) . Summing the second prefix formula gives (I.3); the first gives (I.2). Changing either endpoint convention costs O ( L 2 ) at most. These are absolute estimates at scale D , even when J is very short. They are not relative prime asymptotics on singleton intervals.
On the lower and upper slabs, respectively, l o g u = a 0 l o g D + O ( 1 ) and l o g u = c 0 l o g D + O ( 1 ) . Thus B q = ( a 0 l q + c 0 v q ) l o g D + O ( 1 ) , which proves (I.4). For the support count use the corresponding prime-counting prefix formula. Uniformly for x ≤ t ≤ z on a slab u ≍ D α ,
1 l o g ( t / u ) = 1 ( 1 − α ) l o g D + O ( 1 / l o g 2 D ) .
Integration in t , followed by summation over u , proves (I.5); removing p ∣ q is again harmless. The positive lower bound for l q proves the asserted orders.
Finally inclusion–exclusion on the odd lattice gives the exact-unit count M q ( J ) = ρ q | J | + O ( τ ( q ) ) ≍ D . If S 1 , S 2 denote the first two moments, completion of the square gives
∑ | δ ( s ) − c | 2 = S 2 − | S 1 | 2 M q ( J ) + M q ( J ) c − S 1 M q ( J ) 2 .
Here S 2 ≍ D l o g D and | S 1 | 2 / M q ( J ) = O ( D ) . This proves (I.6), including complex constants. ▫
Theorem I.4. Exact recentering of the principal average.
Put Z c = S δ − c ( q ) S 1 and E c = E δ − c ( q ) E 1 . For every fixed T > 0 ,
E c ( q ) ≪ T D / L T ,     Q δ ( q ) − Z c ( q ) = c ( q ) L 0 ( q ) − E c ( q ) .
If the additional estimate | L 0 ( q ) | ≪ D / L P holds uniformly on E ⊂ C , with fixed P > 0 , then
∑ q ∈ E | Q δ ( q ) | 2 − | Z c ( q ) | 2 ≪ D 3 L P − 1 .
The L 0 estimate in this last assertion is an explicit hypothesis.
Proof. Subtract c ( q ) times M q ( J ) = ρ q | J | + O ( τ ( q ) ) from (I.2). Since τ ( q ) = D o ( 1 ) , the result is O T ( D / L T ) for every interval J ⊂ [ D / 4 , 2 D ] . For each odd unit row r use J = J b ∩ ( I ( q ) / r ) , retaining its endpoints. There are O ( D ) rows, | a ( r ) | ≤ 1 and φ ( q ) ≍ D , proving the estimate for E c without cancellation in the rows. The displayed identity is exact algebra. The congruence selects only O ( 1 ) partners per row in the bounded boxes, whereas the principal box has O ( D 2 ) pairs. Thus | Q δ | + | Z c | ≪ D L . Under the stated hypothesis their difference is O ( D / L P ) . The difference-of-squares inequality, summed over at most D moduli, proves (I.7). ▫
The raw E δ is not asserted to be small: the exact formula is E δ = c ( q ) E 1 + E c . In particular, the terms − 2 c ( q ) S δ S 1 + c ( q ) 2 S 1 2 remain in | Z c | 2 . Replacing H q by an unrestricted harmonic mean with only an o ( 1 ) error would not preserve arbitrary logarithmic precision.

Appendix J. Combining a Prime Label with the Möbius Variable

Retain Definition I.1. Let R S ( q ) and R E ( q ) be the parts of S δ ( q ) and E δ ( q ) in which the prime p in s = p u divides r , and put R Q = R S − R E .
Theorem J.1. A negligible repeated-prime contribution. For every fixed ε > 0 and every E ⊂ C ,
∑ q ∈ E ( | R S ( q ) | + | R Q ( q ) | ) ≪ ε D 1 + c 0 + ε , s u p q ∈ C ( | R S ( q ) | + | R Q ( q ) | ) ≪ ε D 2 c 0 + ε .
Consequently, their squared sums are O ε ( D 1 + 3 c 0 + ε ) . These assertions hold also for nonsquarefree physical moduli.
Proof. Write r = p t . With P * = D / ( 4 z 1 ) , enlarge the positive representation sum by discarding the guards and primality. Its mass is at most
L ∑ p ≥ P * 2 D p 4 D p ≪ D 1 + c 0 L .
For a selected term, q divides the nonzero integer K − p 2 t u , and the product window forces this integer to lie between 9 D 2 / 20 and 11 D 2 / 20 . The divisor bound gives O ε ( D ε ) possible moduli for each triple. For the principal sum the factor 1 / φ ( q ) ≪ 1 / D gives | R E ( q ) | ≪ D c 0 L . These estimates prove the first assertion.
For the pointwise bound, u ≤ z 1 and t ≤ 16 z 1 . If ( t u , q ) > 1 there is no selected solution. Otherwise p 2 ≡ K ( t u ) − 1   ( m o d   q ) . Each odd prime power has at most two square roots of a unit: roots modulo its prime lift uniquely because 2 p is a unit at a root. CRT gives at most 2 ω ( q ) roots, and p < q gives at most one eligible representative per root. The result is O ( L z 1 2 2 ω ( q ) ) = O ε ( D 2 c 0 + ε ) . Multiplying the supremum and first-norm bounds, with smaller preliminary values of ε , gives the squared-sum assertion. ▫
Definition J.2. Restricted prime-divisor weight. For odd u , w > 0 put
W u ( w ) = ∑ p ∣ w ,   p > D 2 / 5   prime w / p ∈ I a ,   p u ∈ J b l o g p ,
and define
Q * ( q ) = − ∑ u , w   odd u w ∈ I ( q ) η ( u ) μ ( w ) A G 1 ( w ) W u ( w ) T q ( u w ) .
Let S * be its selected part and Z * = S * − c ( q ) S 1 .
Theorem J.3. Exact restricted product identity. One has Q δ = Q * + R Q and Z c = Z * + R S . Furthermore,
∑ q ∈ E | Q δ | 2 − | Q * | 2 + | Z c | 2 − | Z * | 2 ≪ ε D 2 + c 0 + ε .  
At a fixed u , w , all retained prime representations have the same sign − η ( u ) μ ( w ) . There are at most two distinct eligible primes in W u ( w ) .
Proof. Eligible primes are outside G 1 . If p 2 ∤ w , then μ ( w / p ) = − μ ( w ) and A G 1 ( w / p ) = A G 1 ( w ) , including the case when both Möbius values vanish. If p 2 ∣ w , the collapsed coefficient vanishes and the original term belongs to R . Summing these identities over the eligible prime divisors proves both exact formulas, with all box restrictions retained. Each eligible prime is at least D / ( 4 u ) and w ≤ 8 D 2 / u . Three such divisors would imply D ≤ 512 u 2 ≤ 512 D 2 c 0 , which is impossible eventually. Thus 0 ≤ W u ( w ) ≤ 2 L .
The absolute selected mass is O ( D L ) , by uniqueness of the right prime label and O ( 1 ) partners per right index; the principal mass has the same bound. These bounds also apply to the collapsed expressions and to their centered versions. Apply the difference-of-squares inequality and the first-norm estimate in Theorem J.1 to obtain (J.1). This step includes the cross terms, not just the squared discarded mass. ▫
Proposition J.4. The exact remaining progression. For ( u , q ) = 1 choose v 0 ∈ { 0 , … , u − 1 } with 2 q v 0 ≡ K   ( m o d   u ) , and put K u ( q ) = ( K − 2 q v 0 ) / u ∈ Z . Then
S * ( q ) = − ∑ u   odd ( u , q ) = 1 η ( u ) ∑ D / 4 − v 0 u < k ≤ 11 D / 40 − v 0 u μ ( w k ) A G 1 ( w k ) W u ( w k ) , w k = K u ( q ) − 2 q k .
Its ambient size is X u ≍ D 2 / u and q / X u ≍ u .
Proof. Since K , u w , q are odd, the conditions u w ∈ I ( q ) and u w ≡ K   ( m o d   q ) are equivalent to u w = K − 2 q v with D / 4 < v ≤ 11 D / 40 . The required divisibility by u is exactly v = v 0 + u k . Substitution proves the identity. The product window gives u w ≍ D 2 , and the claimed scale follows. ▫
For I a = [ A a , B a ) and J b = [ A b , B b ) the prime window in W u ( w ) is [ A b / u , B b / u ) ∩ ( w / B a , w / A a ] , with the additional prime cutoff. It moves with w . A bound for an unweighted Möbius progression does not automatically survive this arithmetic multiplier.

Appendix K. Two Obstructions to Losing the Joint Signs

The following lower bound is for a positive majorant of the actual component, not for its signed value.
Theorem K.1. Failure of a rowwise absolute character bound.
Take I a = [ 2 D , 4 D ) and J b = [ D / 4 , D / 2 ) in Definition I.1. Put f q ( s ) = δ ( s ) − c ( q ) , J r ( q ) = J b ∩ ( I ( q ) / r ) and
B χ ( q , r ) = ∑ s ∈ J r ( q ) s   odd f q ( s ) χ ( s ) ,   R ( q ) = 1 φ ( q ) ∑ r ∈ I a ,   r   odd ( r , q ) = 1 | a ( r ) | ∑ χ ≠ χ 0 | B χ ( q , r ) | .
Then | Q δ ( q ) − c ( q ) L 0 ( q ) | ≤ R ( q ) , but
∑ q ∈ C R ( q ) 2 ≫ κ D 3 / L .
The lower bound already comes from prime moduli and characters of conductor q . It does not give a lower bound for | Z c | 2 .
Proof. Character orthogonality gives exactly
Q δ − c ( q ) L 0 = 1 φ ( q ) ∑ χ ≠ χ 0 χ ( K ) ¯ ∑ r ∈ I a r   odd a ( r ) χ ( r ) B χ ( q , r ) .
Two triangle inequalities yield the majorant. Restrict its nonnegative sum to primes q ∈ ( 9 D / 10 , D ] and primes r ∈ [ 21 D / 10 , 11 D / 5 ] . Here a ( r ) = − 1 . The full interval I ( q ) / r lies in [ 19 D / 44 , D / 2 ) , with length D q / ( 20 r ) ≥ 9 D / 440 . Thus it equals J r ( q ) , and its integers are distinct nonzero residues modulo q . Theorem I.3 and boundedness of c ( q ) give
M 2 : = ∑ s ∈ J r ( q ) s   odd | f q ( s ) | 2 ≫ D L ,     M 1 : = ∑ s ∈ J r ( q ) s   odd | f q ( s ) | ≪ D .
The principal transform is O T ( D / L T ) by the proof of Theorem I.4. Orthogonality therefore yields
∑ χ ≠ χ 0 | B χ ( q , r ) | 2 = ( q − 1 ) M 2 − | B χ 0 ( q , r ) | 2 ≫ D 2 L .
Since every | B χ | ≤ M 1 ≪ D , the sum of their absolute values is ≫ D L . There are ≍ D / l o g D such prime rows, so R ( q ) ≫ D . There are also ≍ D / l o g D such prime moduli. This proves (K.1) by the ordinary prime number theorem in fixed-ratio intervals. Every nonprincipal character modulo a prime has conductor that prime. All restrictions made in this argument were in a positive sum; making them in the signed identity would not be valid. ▫
Proposition K.2. Multiplicative marking of a fixed prime window.
Let g ( n ) = 1 n   odd μ ( n ) A G 1 ( n ) and ω ≤ t ( n ) = # { p ∣ n : p ≤ t } . For | z | ≤ 1 put F z , t ( n ) = g ( n ) ( 1 + z ) ω ≤ t ( n ) . This is multiplicative, | F z , t ( n ) | ≤ τ ( n ) , and
G t ( n ) : = g ( n ) ω ≤ t ( n ) = 1 2 π ∫ 0 2 π e − i θ F e i θ , t ( n )   d θ .
For 0 < A < B one has exactly
g ( n ) ∑ A < p ≤ B p ∣ n l o g p = l o g B   G B ( n ) − l o g A   G A ( n ) − ∫ A B G t ( n ) d t t .
Any uniform norm bound for a finite linear functional on F z , t therefore passes to G t without a degree factor, and to the fixed logarithmic window with cost at most 2 l o g B when A ≥ 1 .
Proof. The prime factors of coprime integers are disjoint, which proves multiplicativity; nonsquarefree values vanish. On squarefree integers 2 ω ( n ) = τ ( n ) . The integral extracts the coefficient of z in a finite polynomial. A divisor prime p ≤ A has zero coefficient on the right of (K.2); a prime A < p ≤ B has coefficient l o g B − l o g ( B / p ) = l o g p . This proves the identity. Apply the triangle inequality in the finite coefficient integral and then in (K.2). The total coefficient variation is l o g B + l o g A + l o g ( B / A ) = 2 l o g B . ▫
The construction preserves the Möbius sign and the avoidance factor. Its prime values are piecewise constant, so it is compatible with the type of class considered by Fouvry and Tenenbaum [2]. This observation does not replace the hypotheses on the modulus norm, the target residue or the moving window. In particular, their Theorem 1.5 controls a signed inner modulus average with a short outer absolute average, while their Theorem 1.8 with individual modulus absolute values has a range below the square-root scale.
Proposition K.3. Small representatives are rare. Let u ≤ D c 0 be a positive integer and T ≥ 1 satisfy u T ≤ K / 2 . For every fixed ε > 0 ,
# { q ∈ ( 9 D / 10 , D ] : q   prime ,   ∃ a ∈ Z ,   | a | ≤ T ,   u a ≡ K   ( m o d   q ) } ≪ ε ( T + 1 ) D ε .
Thus almost every prime modulus in this interval has no such representative for T any fixed power of l o g D , or for T = ( D 2 / u ) δ with fixed 0 < δ < 1 / 2 .
Proof. For each integer a , the modulus divides the nonzero integer K − u a in [ K / 2 , 3 K / 2 ] . Apply the divisor bound and sum over the at most 2 ⌊ T ⌋ + 1 choices. For the last assertion, T ≤ D 2 δ and u T = o ( D 2 ) ; choose ε < 1 − 2 δ and compare with the prime number theorem count ≍ D / l o g D . ▫
There is also a purely norm-theoretic distinction. List these prime moduli as q 1 < … < q M and put d ( q j ) = ( − 1 ) j D , with d = 0 elsewhere. Every signed prefix has modulus at most D , whereas ∑ j | d ( q j ) | 2 = M D 2 ≍ D 3 / l o g D . For X u = D 2 / u and u ≤ D c 0 , one has D = o ( X u / l o g T X u ) for every fixed T . Hence even such arbitrary-logarithmic prefix bounds do not imply the required squared norm. This is a counterexample between abstract norms, not an arithmetic counterexample for prime discrepancies.

Appendix L. A Rational-Target CRT Kernel

An elementary reciprocity identity permits a restricted large-target estimate without replacing the rational target by a small integer. The resulting estimate is a corollary of the Kloosterman-fraction bound of Bettin and Chandee [1]; it is not an estimate for the whole sum in Proposition J.4.
Lemma L.1. Reciprocity with the cofactor retained. Let u be odd, let l ≠ l ′ be primes coprime to u , and let q ≠ q ′ be primes exceeding u , l , l ′ . Let R modulo q q ′ be specified by
u l R ≡ K   ( m o d   q ) ,     u l ′ R ≡ K   ( m o d   q ′ ) .
Put m = l q ′ , n = u l ′ q , and let m ¯ denote the inverse of m modulo n . Then for every integer h ,
h R q q ′ ≡ K h l ′ − l m ¯ u l ′ q + K h u l ′ q q ′ − K h   l q q ′ ¯ u   m o d   1 .
The last term is zero modulo one when u = 1 .
Proof. CRT expresses the left side as the sum of K h u l q ′ ¯ / q and K h u l ′ q ¯ / q ′ modulo one. Write v = m ¯ and choose j ≡ − l q q ′ ¯   ( m o d   u ) (take j = 0 for u = 1 ). Then v + j q is divisible by u , and ( v + j q ) / u is an inverse of u l q ′ modulo q . The first CRT term is consequently K h l ′ v / n + K h j / u . The second is K h l n ¯ / m . For coprime m , n , ordinary reciprocity is n ¯ / m + m ¯ / n ≡ 1 / ( m n )   ( m o d   1 ) ; it follows directly by multiplying by m n . Combining gives K h ( l ′ − l ) v / n + K h l / ( m n ) + K h j / u , which is (L.1). All numerators are integers, so the modular substitutions are legitimate even when K / u is not integral. ▫
Proposition L.2. A bound for separate physical-modulus weights.
Suppose 1 ≤ u ≤ D c 0 is odd, l , l ′ ≤ Z ≤ D 1 / 4 , l ≠ l ′ , and ( u , l l ′ ) = 1 . For arbitrary complex | α q | , | β q ′ | ≤ 1 supported on primes in ( 9 D / 10 , D ] , let
B h = ∑ q ≠ q ′ α q β q ′   e h R ( q , q ′ ) q q ′ ,     h ≠ 0 .
For every fixed ε > 0 ,
| B h | ≪ ε D ε ( 1 + | h | / u ) 5 / 2 ( D 39 / 20 u 11 / 10 Z 19 / 20 + D 15 / 8 u Z 7 / 8 ) .
The implied constant is uniform in these weights; no arbitrary joint weight depending on both q , q ′ is permitted.
Proof. On ( Z / u Z ) * expand F ( t ) = e ( − K h l t ¯ / u ) in multiplicative characters: F ( t ) = ∑ χ c χ χ ( t ) . Parseval gives ∑ | c χ | 2 = 1 and ∑ | c χ | ≤ φ ( u ) ≤ u . For u = 1 use the single constant character. Thus the last factor in (L.1) separates at a cost at most u .
For each character apply [1, Theorem 1] with its third variable supported at one, and with
M ≍ l D ,   N 1 ≍ u l ′ D ,   ϑ = K h ( l ′ − l ) ≠ 0 .
Here N 1 denotes the source scale, not our primorial. The two coefficient sequences are supported on m = l q ′ and n = u l ′ q ; their Euclidean norm product is O ( D ) . Coprimality of m , n is exactly q ≠ q ′ . The cited bound is
≪ ε D ( 1 + | ϑ | / ( M N 1 ) ) 1 / 2 ( ( M N 1 ) 7 / 20 + ε ( M + N 1 ) 1 / 4 + ( M N 1 ) 3 / 8 + ε ( M + N 1 ) 1 / 8 ) .
It also holds on every rectangular prefix, by restricting the coefficients. The smooth middle factor in (L.1) is retained by two-dimensional partial summation. On the rectangle q , q ′ ≍ D , the sum of its corner, edge-variation and mixed-variation norms is O ( ( 1 + | h | / u ) 2 ) : differentiating e ( K h / ( u l ′ q q ′ ) ) once in each variable gives, after integration, at most a constant times T + T 2 , where T = | K h | / ( u l ′ D 2 ) ≪ | h | / u . Also | ϑ | / ( M N 1 ) ≪ | h | / u .
Finally use M N 1 ≪ u Z 2 D 2 , M + N 1 ≪ u Z D and multiply by u . The first term becomes D 39 / 20 u 1 / 2 + 7 / 20 + 1 / 4 Z 7 / 10 + 1 / 4 , and the second becomes D 15 / 8 u 1 / 2 + 3 / 8 + 1 / 8 Z 3 / 4 + 1 / 8 . Since all source scales are bounded powers of D , a smaller preliminary epsilon absorbs their epsilon factors into D ε . This proves (L.2). ▫
Corollary L.3. A lower-cofactor frequency budget. Under Proposition L.2, suppose Z ≤ D δ , 1 ≤ H ≤ u D δ and 0 < δ ≤ 1 / 1000 . Then
D − 2 ∑ 0 < | h | ≤ H | B h | ≪ ε D 5 δ + ε D − 1 / 20 u 21 / 10 + D − 1 / 8 u 2 .
For u ≤ D a 0 this is O ( D − 1 / 40 ) .
Proof. Use O ( u D δ ) frequencies and ( 1 + | h | / u ) 5 / 2 ≪ D 5 δ / 2 . The two total delta exponents are 89 δ / 20 and 35 δ / 8 , both at most 5 δ . Since a 0 < 1 / 112 , 1 / 20 − ( 21 / 10 ) a 0 > 1 / 32 and 1 / 8 − 2 a 0 > 3 / 28 . Taking the displayed epsilon to be 1 / 1000 leaves a saving strictly larger than 1 / 40 in both terms. ▫
This result treats prime physical moduli and separate weights for a specified kernel. It does not freeze the moving weight W u ( w ) , justify a full Poisson frequency range, sum over the dilation primes, or control the upper cofactor slab u ≍ D c 0 . In that slab the first exponent in the corollary gives no saving. Green’s method for multiplicative functions in large progressions [4] motivates this kernel, but its small common integer target hypothesis cannot be replaced by the present rational target without these additional arguments. The elementary identity above keeps the archimedean reciprocity factor precisely because it is not negligible for a numerator of size comparable to D 2 .

Concluding Remarks

The proof of Theorem 1.1 isolates three uniformity costs: the correction for a possible exceptional character, the loss ( T / Y ) 2 from location-scale prime prefixes, and the finite union over all prescribed residue pairs and moduli. The arithmetic kernel norm allows the first cost to retain both suppressed factors, while the deliberately small fixed parameters pay the other two. Improving the admissible exponents would require sharper analytic input or a more efficient treatment of these costs.
The classical polylogarithmic asymptotic and the fixed-power lower bound have different strengths and should be kept separate. The proper-power estimate is an elementary normalization device; it does not replace a prime-distribution theorem. All results retain their stated exceptional sets or hypotheses, and the binary Goldbach conjecture is not resolved by them. The appendices also show the extent of the comparison method: its Type I and convex regularity properties can be established without the missing Type II input. An uncentered upper sieve bounds balanced two-prime mass, but does not yield a centered covariance asymptotic or a positive lower bound. The fixed-margin zero implication remains conditional on its explicitly stated arithmetic band. Thus none of the additional statements changes the every-target limitation.
The sharp-boundary analysis separates three further issues. Its exact unit mean makes the principal average small, but subtraction of even the best constant leaves coefficient energy of order D l o g D . A specific rowwise absolute character majorant has squared norm at least D 3 / l o g D , so that majorant cannot prove the desired arbitrary-logarithmic saving. After the repeated-prime contribution is removed, the remaining product weight preserves both original boxes and the common Möbius sign. The rational-target kernel estimate supplies a restricted lower-cofactor bound; it does not control the upper slab or the moving prime window. These results therefore leave a concrete signed correlation problem, as well as the other components needed for an every-target conclusion.

References

  1. Bettin, S.; Chandee, V. Trilinear forms with Kloosterman fractions. Adv. Math. 2018, 328, 1234–1262. [Google Scholar] [CrossRef]
  2. Fouvry, É.; Tenenbaum, G. Multiplicative functions in large arithmetic progressions and applications. Trans. Am. Math. Soc. 2022, 375, 245–299. [Google Scholar] [CrossRef]
  3. Gallagher, P.X. A large sieve density estimate near 4. Invent. Math. 1970, 11, 329–339. [Google Scholar] [CrossRef]
  4. Green, B. A note on multiplicative functions on progressions to large moduli, preprint. arXiv 2024, arXiv:1604.04481v4. [Google Scholar]
  5. Halupczok, K. Goldbach’s problem with primes in arithmetic progressions and in short intervals. J. Théor. Nr. Bordx. 2013, 25, 331–351. [Google Scholar] [CrossRef]
  6. Kawada, K. The prime 𝑘‐tuplets in arithmetic progressions. Tsukuba J. Math. 1993, 17, 43–57. Available online: https://tsukuba.repo.nii.ac.jp/record/15690/files/3.pdf. [CrossRef]
  7. Li, R. Primes in arithmetic progressions to large moduli and refinements of Harman’s sieve. arXiv 2026, arXiv:2602.20917v6. [Google Scholar]
  8. Montgomery, H.L.; Vaughan, R.C. The exceptional set in Goldbach’s problem. Acta Arith. 1975, 27, 353–370. Available online: https://matwbn.icm.edu.pl/ksiazki/aa/aa27/aa27126.pdf. [CrossRef]
  9. Montgomery, H.L.; Vaughan, R.C. Multiplicative Number Theory II: Primes and Sieves. author-hosted draft, Chapter 17, Theorem 17.1. pp. 65–66. Available online: https://personal.science.psu.edu/rcv4/571s25/montgomery-vaughanII.pdf.
  10. Pascadi. On the exponents of distribution of primes and smooth numbers. arXiv 2025, arXiv:2505.00653v2. https://arxiv.org/abs/2505.00653v2. [Google Scholar]
  11. Pintz, J. A new explicit formula in the additive theory of primes with applications I. The explicit formula for the Goldbach and Generalized Twin Prime Problems. arXiv 2018, arXiv:1804.05561v1. https://arxiv.org/abs/1804.05561v1. [Google Scholar]
  12. Vaughan, R.C. The Bombieri–Vinogradov Theorem. author-hosted lecture note, pp. 1, 3. Available online: https://personal.science.psu.edu/rcv4/Bombieri.pdf.
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.