The Insertion Identity, Variance Splitting, and the Relay Engine of Conjecture H'

1. Introduction

1.1 Context

Paper 15 (DOI: 10.5281/zenodo.19007312) established a proof architecture for D(N) → 1 under three assumptions: (A) monotone pointwise convergence of p_k(N), (A') p_∞(k) → 1, and (B) Sathe–Selberg. The present paper constructs the algebraic and analytic machinery needed to close Assumptions A and A'.

The key advance is the discovery that the relay mechanism — identified qualitatively in Paper 15 §6.4 — admits an exact algebraic formalization as the insertion identity, and that the variance control needed for Chebyshev application follows from a splitting lemma that exploits the one-sided bounds inherent in the DP definition of ρ_E.

1.2 Main Results

2. The Exact Insertion Identity

2.1 Statement and Proof

Theorem 1. For any composite n ≥ 4 with P⁻(n) = p and m = n/p, G_spf(n) = j(m) + K_p(m), where j(m) = max(G(m), 0) and K_p(m) = ρ_E(pm−1) − ρ_E(m−1) − ρ_E(p) − 2.

Proof. By definition, G_spf(n) = ρ_E(n−1) − ρ_E(p) − ρ_E(m) − 1. The DP recurrence gives ρ_E(m) = ρ_E(m−1) + 1 − j(m). Substituting:

Remark. When m is prime, G(m) ≤ 0, so j(m) = 0 and ρ_E(m) = ρ_E(m−1) + 1, consistent. Verified: N = 10⁷, k = 3 through 18, identity holds exactly for all composites.

2.2 The Insertion Recursion

The map (m, p) ↦ n = pm with Ω(m) = k, p prime, p ≤ P⁻(m), pm ≤ x gives a bijection onto {n ≤ x : Ω(n) = k+1, n composite}. Define the insertion measure I_k(x) by this bijection. Then:

This is exact. The first term transmits the previous-layer jump; the second is the local bridge cost.

2.3 Interpretation: the Relay Formalized

Paper 15 §6.4 identified the relay mechanism qualitatively: P(jump) drives at low k, E[j/ln | jump] drives at high k. The insertion identity reveals the precise algebraic mechanism: the (k+1)-th layer's mean gain inherits the k-th layer's jump sizes (via j(m)), augmented or diminished by the bridge cost K_p(m).

3. The Variance Splitting Lemma

3.1 The A/B Decomposition

Theorem 2. Write G_spf(n) = A(n) + B(n) + 1 where A(n) = ρ_E(n−1) − ρ_E(n) and B(n) = ρ_E(n) − ρ_E(P⁻(n)) − ρ_E(n/P⁻(n)) − 2.

Proof sketches. (a) ρ_E(n) ≤ ρ_E(n−1) + 1 (successor always available), so A(n) ≥ −1. (b) ρ_E(n) ≤ ρ_E(P⁻(n)) + ρ_E(n/P⁻(n)) + 2 (SPF split valid), so B(n) ≤ 0.

(c) Let p = P⁻(n), v = v_p(n). Write ρ_E(n) = f(n) + r(n). For v = 1 (squarefree-in-P⁻): ε_p = 0, so B(n) = r(n) − r(n/p) − 2. Lower bound: r(n) ≥ 0 and r(n/p) ≤ 2(k−2), giving B ≥ −2(k−1). Hence B ∈ [−2(k−1), 0] unconditionally. For v ≥ 2: ε_p = ρ_E(p^v) − ρ_E(p) − ρ_E(p^{v−1}); the available lower bound deteriorates with p, not yielding an N-independent constant — isolated into Numerical Hypothesis B-bound.

(d) Under Numerical Hypotheses A-tail and B-bound, Var(G_spf | Ω = k, n ≤ N) = O_k(1) uniformly in N. □

Remark (Why direct analysis fails). The "obvious" computation via G = X − Y − Z − 1 gives Var(X) ≈ 20.6, Var(Z) ≈ 23.1, Cov(X,Z) ≈ 19.85 — enormous covariance nearly cancelling the individual variances, a consequence of the shared factor ρ_E(n). The A/B decomposition algebraically removes this common factor, revealing the true variance is O(1) residuals.

3.2 Numerical Verification

At N = 10⁷, Var(G_spf | Ω = k) ∈ [1.21, 1.83] across k = 2 to 20, no increasing trend. Component analysis at k = 4, N = 10⁶: Var(A) ≈ 0.81, Var(B) ≈ 0.54, Cov(A,B) ≈ −0.05.

k	Var(G_spf | Ω=k)	Var(A)	Var(B)	Cov(A,B)
2	1.21	—	—	—
4	1.43	0.81	0.54	−0.05
10	1.67	—	—	—
20	1.83	—	—	—

4. The Bridge Term

4.1 Algebraic Identity

Theorem 3. K_p(m) = A(pm) + B(pm) − A(m).

Proof. A(pm) = ρ_E(pm−1) − ρ_E(pm). B(pm) = ρ_E(pm) − ρ_E(p) − ρ_E(m) − 2. A(m) = ρ_E(m−1) − ρ_E(m). Therefore:

4.2 The Bridge Lower Bound: Status

The algebraic identity K_p = A(pm) + B(pm) − A(m) is unconditional. The one-sided bounds (A ≥ −1, B ≤ 0) constrain the range of K_p but do not directly yield a lower bound on its mean. A bounded-variance, non-positive random variable can still have arbitrarily negative mean. Therefore, the bridge lower bound is not a consequence of the hypotheses in this paper.

4.3 Numerical Structure

The insertion-weighted bridge term E_{I_k}[K_p] is negative for all k examined. The negative drift is driven by p = 2: its weight in I_k rises from 43% at k=3 to 88% at k=6.

k	E_{I_k}[K_p]	Weight of p=2
3	−0.19	43%
4	−0.31	62%
5	−0.46	74%
6	−0.57	88%
∞	≈ −0.7	88%

Per-prime structure: E[K_2] ≈ −0.65 (always negative), E[K_3] ≈ +0.30, E[K_5] ≈ +0.50, E[K_7] ≈ +0.02. The positive contributions from p = 3, 5 are overwhelmed by p = 2's dominance. The conjecture E_{I_k}[K_p] > 0 is false. The growth of μ_k cannot come from the bridge term; it must come entirely from jump transmission E_{I_k}[j(m)].

5. The Insertion Measure and Its Bias

5.1 Radon-Nikodym Formula

For m on the Ω = k shell with m ≤ x, define w_x(m) = π(min(P⁻(m), x/m)). The insertion measure has density

For any shell observable H: E^ins[H] − E^amb[H] = Cov^amb(H, w_x) / E^amb[w_x]. Lemma I is equivalent to Cov^amb(G_spf, w_x) ≥ −C_1 · E^amb[w_x]. The Radon-Nikodym derivative is not uniformly bounded.

5.2 Two-Parameter Shell Mean Identity

Theorem 4 reduces the insertion bias to a weighted average of two-parameter shell means μ_k(x/p, p) over primes p, with weights ω_{k,x}(p) = Φ_k(x/p, p) / N_{k+1}(x).

5.3 Covariance Decomposition

Decompose w_x(m) = π(P⁻(m)) + r_x(m) where r_x(m) = π(min(P⁻(m), x/m)) − π(P⁻(m)) ≤ 0. Then:

Numerically at N = 10⁶: for k ≥ 4, the first term is positive (larger P⁻ correlates with larger G_spf). For k = 3, the first term is negative, but the size cutoff overcompensates, making the total positive.

5.4 An Eliminated Approach: P⁻-Monotonicity

The route "prove E[G_spf | P⁻ = p, Ω = k] is non-decreasing in p, then apply Chebyshev rearrangement" is false in general. At k = 3, N = 10⁶: the shell mean is not monotone — it rises for p = 3, 5 then drops for p = 7, 11, …. Pure P⁻-monotonicity cannot establish Lemma I; the size cutoff is essential.

5.5 Numerical Evidence

At N = 10⁷, the insertion bias on j(m) — that is, E^ins[j(m)] − E^amb[j(m)] — is positive for all tested k:

k	j-bias (E^ins − E^amb)
3	+0.302
4	+0.415
5	+0.432
8	+0.300
10	+0.271
15	+0.226
18	+0.421

Mean j-bias ≈ +0.31, stable across k. Direct computation of the G_spf-bias at N = 10⁶ gives values in +0.03 to +0.10 (all positive), providing stronger but still numerical support for Lemma I.

6. The Toy Model

6.1 Finite Prime Alphabet

Theorem 6. Let P = {q_1 < ··· < q_r} be a finite set of primes. On the restricted shell, partition by P⁻(n) = q_i with class weights α_i and class means h_i = E[H | P⁻ = q_i]. The toy insertion weight is w_i = i (number of primes ≤ q_i in P). If h_1 ≤ h_2 ≤ ··· ≤ h_r, then E^ins[H] ≥ E^amb[H].

Proof. E^ins − E^amb = Cov_α(w_i, h_i) / Σ_i w_i α_i. Since w_i = i is increasing and h_i is assumed increasing, Chebyshev's sum inequality gives Cov_α(w, h) ≥ 0. □

6.2 Significance and Limitations

The toy model proves that insertion bias is automatically non-negative when the shell observable is P⁻-monotone. In the unrestricted setting, G_spf is not globally P⁻-monotone (§5.4), so the toy proof does not directly apply. However, it identifies the structural mechanism: insertion reweighting favors larger-P⁻ classes, which tend to have higher gains. The gap between the toy model and the full result is precisely the role of the size cutoff.

7. The Conditional Growth Criterion

7.1 Statement and Bridge Corollary

Theorem 5 (Finite-Window Recursion). Under Lemma I and Lemma II, define the truncation gain δ_k(x) = E_{I_k}[max(−G_spf(m), 0)]. Then:

Proof. By Theorem 1, μ_{k+1} = E_{I_k}[j(m)] + E_{I_k}[K_p]. Since G(m) ≥ G_spf(m), j(m) = max(G(m),0) ≥ max(G_spf(m),0) = G_spf(m) + max(−G_spf(m),0). Taking expectations: E_{I_k}[j] ≥ E_{I_k}[G_spf] + δ_k ≥ (μ_k − C_1) + δ_k. With E_{I_k}[K_p] ≥ −C_2(k):

Bridge Corollary. Define μ_spf,∞(k) := liminf_{x→∞} μ_k(x) and δ_{k,∞} := liminf_{x→∞} δ_k(x). If Lemmas I and II hold uniformly in x, then μ_spf,∞(k+1) ≥ μ_spf,∞(k) + δ_{k,∞} − C_1 − C_2(k). In particular, μ_spf,∞(k) → ∞ whenever Σ_{j=k_0}^{K} (δ_{j,∞} − C_1 − C_2(j)) → +∞ as K → ∞.

7.2 The Truncation Gain

The truncation gain δ_k is the expected value of |G_spf(m)| on the negative part of the G_spf distribution, under insertion measure. At N = 10⁷, δ_k > 0 for all tested k: δ_3 ≈ 0.45, δ_5 ≈ 0.35, δ_{10} ≈ 0.20. At k ≥ 17, p_k(N) = 1.000, so δ_k becomes very small in the current window.

7.3 The Growth Mechanism

Combining the three ingredients:

• Truncation gain δ_k > 0 (relay engine)
• Insertion bias ≈ +0.31 (measure favorability)
• Bridge drag ≈ −0.7 (bounded cost)

The net per-step increment s_k := μ_{k+1} − μ_k ≈ 0.26, stable across k = 2 to 20 at N = 10⁷. The mechanism: even as μ_k grows, the insertion bias provides a persistent uplift that compensates the shrinking truncation gain and the bridge drag.

8. Numerical Evidence at N = 10⁷

8.1 Summary Tables

k	E[G_spf | Ω=k]	Var(G_spf | Ω=k)	p_k(N)	E_{I_k}[j(m)]	E_{I_k}[K_p]
2	−0.47	1.21	0.229	0.30	−0.23
5	1.24	1.49	0.752	1.71	−0.46
10	2.87	1.67	0.970	3.57	−0.81
15	4.08	1.75	0.998	4.72	−0.89
20	4.52	1.83	1.000	4.93	−0.97

E[G_spf | Ω = k]: linear growth from −0.47 (k=2) to 4.52 (k=20), slope ≈ 0.26, R² = 0.995. p_k(N): strictly increasing from 0.229 to 1.000.

8.2 Eliminated Conjectures

E_{I_k}[K_p] > 0: false. The bridge term is persistently negative, driven by p=2 dominance.

P⁻-monotonicity of shell means: false at k = 3. Size cutoff is essential for positive insertion bias.

9. Toward Convergence of SPF-Positivity Rates

9.1 The (A, B) Reformulation

The SPF-positivity condition G_spf(n) > 0 becomes A(n) + B(n) > −1. Define p_k^spf(N) := P(G_spf(n) > 0 | Ω(n) = k, n ≤ N). Since G(n) ≥ G_spf(n), p_k^spf provides a lower bound for the actual jump rate p_k(N). Convergence of p_k^spf(N) would support (but not directly establish) Assumption A.

9.2 Why Convergence is Expected

Convergence reduces to the convergence of the joint distribution of (A, B) on Ω-shells. B(n) takes O_k(1) distinct values in [−2(k−1), 0] in the squarefree-dominant case (Sathe-Selberg), and their frequencies are controlled by Sathe-Selberg asymptotics. A(n) = ρ_E(n−1) − ρ_E(n) involves the predecessor n−1, connecting to the bivariate Erdős-Kac framework: Goudout proved that ω(n−1) retains Erdős-Kac behavior conditional on ω(n) = k; Mangerel gave quantitative bivariate results for (ω(n), ω(n+a)). Extending such results to ρ_E (which is near-additive by the f+r decomposition) is a well-defined problem at the current frontier.

9.3 The Remaining Gap

Full proof requires: (i) convergence of r-value frequencies on Ω-shells (from Sathe-Selberg + DP smoothness), and (ii) predecessor stability — the distribution of ρ_E(n−1) − ρ_E(n) conditional on n's factorization type converges, connecting to the bivariate Erdős-Kac literature.

10. Conclusion

10.1 Established Results

This paper proves unconditionally: (a) Theorem 1 (Insertion Identity): G_spf(pm) = j(m) + K_p(m); (b) Theorem 2(a)(b)(c-squarefree) (Variance Splitting): G_spf = A + B + 1 with A ≥ −1, B ≤ 0, B ∈ [−2(k−1), 0] when v_{P⁻}(n) = 1; (c) Theorem 3 (Bridge Identity): K_p = A(pm) + B(pm) − A(m); (d) Theorem 4 (Two-Parameter Reduction): insertion mean is a weighted average of μ_k(x/p, p); (e) Theorem 6 (Toy Model): Chebyshev rearrangement proves bias ≥ 0 under P⁻-monotonicity.

Conditionally (under Hypotheses A-tail and B-bound): (f) Theorem 2(d): Var(G_spf | Ω = k) = O_k(1).

Numerically observed but not proved: (g) Bridge Lower Bound (Lemma II), open input for Theorem 5; (h) Theorem 5 (Finite-Window Recursion) and Bridge Corollary under Lemma I + II.

10.2 The Remaining Frontier

The proof of μ_spf(k) → ∞ reduces to four inputs:

• Numerical Hypothesis A-tail: P(A(n) > t | Ω = k) ≤ Ce^{−αt}. Verified to N = 10⁷. Proving requires quantitative control of "smoothness jumps" between adjacent integers (Barban-Davenport-Halberstam).
• Numerical Hypothesis B-bound: Var(B | Ω = k, n ≤ N) = O_k(1) uniformly. Verified to N = 10⁷. Proving requires Sathe-Selberg concentration on squarefree factorizations dominating the v ≥ 2 tail.
• Lemma I (Insertion non-compression): E_{I_k}[G_spf(m)] ≥ μ_k(x) − C_1. Positive j-bias (+0.31) and G_spf-bias (+0.03 to +0.10) provide numerical support.
• SPF-positivity convergence: Reformulated via (A, B) decomposition as a shifted-integer joint distribution problem at the frontier of analytic number theory.

10.3 Proof Dependency Graph

10.4 Eliminated Approaches

"E_{I_k}[K_p] > 0": false (p = 2 dominates at 88%). "P⁻-monotonicity of shell means": false at k = 3. "Bounded Radon-Nikodym derivative": false. "p_∞(k) = 1/2 via Gaussian noise": false (Cov(ρ_E(n−1), ρ_E(n/P⁻)) ≈ 20 invalidates independence assumption). Each eliminated approach sharpened the correct formulation.

11. Methodological Note: Four-AI Parallel Exploration

11.1 Protocol

The results in this paper were developed through a structured parallel exploration involving four AI systems — Claude (Anthropic), ChatGPT (OpenAI), Gemini (Google), and Grok (xAI) — coordinated by the author. The protocol is documented here as a reproducible methodology for AI-assisted mathematical research on open problems.

The exploration proceeded in two rounds. In Round 1, the three open sub-problems from Paper 15 were assigned based on capability profiles:

• ChatGPT: creative multi-route exploration → Unbounded Mean Gain formalization
• Gemini: numerical computation and distribution analysis → variance control
• Grok: stateful REPL with high-precision computation → ρ_E to N = 10⁷ and cross-line numerical verification
• Claude: long-chain deductive reasoning → Assumption A (deepest theoretical gap)

Each system received a tailored prompt containing the shared mathematical context, the specific sub-problem, suggested attack vectors, and explicit instructions to report obstructions and failures.

11.2 Contributions by System

ChatGPT discovered the exact insertion identity (Theorem 1): G_spf(pm) = j(m) + K_p(m), which became the paper's central algebraic result. It also developed the two-parameter shell mean identity (Theorem 4), the covariance decomposition (§5.3), the finite prime alphabet toy model (Theorem 6), and the conditional growth criterion (Theorem 5). In the review phase, ChatGPT identified: the f-additivity bug in Theorem 2(c) (the non-coprime case p² | n), the fixed-x quantifier error in the Bridge Corollary, the positive-part sum fallacy, the variance-does-not-control-mean gap, and the lower-bound-deterioration distinction in ε_p. Each correction sharpened the paper's logical hygiene after eight revision rounds.

Gemini discovered the A/B variance splitting decomposition (Theorem 2) and the bridge term identity K_p = A(pm) + B(pm) − A(m) (Theorem 3). This was the key structural insight that resolved the variance control problem, bypassing the enormous covariance (~20) between ρ_E(n−1) and ρ_E(n/P⁻). Gemini also computed the per-prime K_p tables revealing p = 2 dominance at 88% weight and refuted the conjecture E_{I_k}[K_p] > 0.

Grok computed ρ_E(n) to N = 10⁷ (529 seconds, 40 MB), providing the numerical backbone for all quantitative claims. It verified the insertion identity exactly for k = 3 through 18, confirmed E[G_spf] linear growth (slope 0.26, R² = 0.995), established Var(G_spf) ∈ [1.21, 1.83] across k = 2 to 20, and measured the insertion bias at +0.31. Grok's data refuted Claude's p_∞(k) = 1/2 conjecture by confirming p_2(10⁷) = 0.2288, inconsistent with convergence to 1/2.

Claude proposed the C1/C2/C3 sub-conjecture framework for Assumption A and the natural density reformulation (§9). Claude also proposed — and then retracted — the p_∞(k) = 1/2 conjecture (based on a Gaussian noise argument that incorrectly assumed ρ_E(n−1) and ρ_E(n/P⁻) fluctuate independently). The retraction, forced by Gemini's Cov ≈ 20 data and Grok's numerics, clarified the essential role of the DP-induced covariance structure.

11.3 Cross-Validation and Error Correction

The protocol's central feature is cross-validation: each system's conjectures and claims were tested against the others' data and analysis. Four conjectures were refuted during the process:

1. E_{I_k}[K_p] > 0 (ChatGPT heuristic → refuted by Gemini's per-prime tables)
2. p_∞(k) = 1/2 (Claude's theoretical prediction → refuted by Gemini's covariance data and Grok's 10⁷ numerics)
3. P⁻-monotonicity of shell means (implicit in early toy model → refuted by ChatGPT's k=3 computation)
4. Bounded Radon-Nikodym derivative (Claude's Approach 3 → refuted by ChatGPT's analysis of w_x(m))

Each refutation produced positive information: (1) identified truncation gain as the sole growth engine; (2) revealed the DP-covariance structure; (3) established the essential role of the size cutoff; (4) motivated the two-parameter shell mean reduction. The f-additivity bug (Theorem 2(c)) — caught by ChatGPT's systematic verification against n = 8 — would have been a serious error in the published paper.

11.4 The Human Role

The author's contributions were: (i) problem selection and direction — choosing which gaps to attack, when to parallelize, and when to converge; (ii) cross-line integration — synthesizing results from four independent explorations into a unified narrative; (iii) arbitration of contradictions — when Gemini's data contradicted Claude's prediction, deciding which to trust and why; (iv) termination judgment — determining when the paper had reached a stable, publishable state. No AI system proposed the four-way parallel structure, decided the round boundaries, or judged when to stop exploring and start writing.

11.5 Reproducibility

The protocol is fully reproducible: prompts, round summaries, and cross-validation steps are documented. Any researcher with access to comparable AI systems can replicate the workflow for other open problems. Essential ingredients: (a) a well-defined problem with identifiable sub-components; (b) AI systems with complementary capability profiles; (c) a human coordinator who provides direction, arbitrates, and terminates.

References

1. H. Qin, "Sieve structure, compositeness discount, and the architecture of Conjecture H'" (Paper 15), Zenodo, DOI: 10.5281/zenodo.19007312.
2. H. Qin, "The formal framework of ZFCρ" (Paper 1), Zenodo, DOI: 10.5281/zenodo.18914682.
3. H. Qin, "Concentration and growth of ρ_E" (Paper 11), Zenodo, 2025.
4. H. Qin, "Telescoping identity and jump statistics" (Paper 12), Zenodo, 2025.
5. H. Qin, "Zero-inflated lattice normal model" (Paper 13), Zenodo, DOI: 10.5281/zenodo.18991986.
6. H. Qin, "Three-component decomposition and variance bounds" (Paper 14), Zenodo, 2026.
7. A. Selberg, "Note on a paper by L. G. Sathe," J. Indian Math. Soc. 18 (1954), 83–87.
8. G. Tenenbaum, Introduction to Analytic and Probabilistic Number Theory, 3rd ed., AMS, 2015.
9. P. Erdős and M. Kac, "The Gaussian law of errors in the theory of additive number theoretic functions," Amer. J. Math. 62 (1940), 738–742.
10. É. Goudout, "Lois de répartition des diviseurs," doctoral thesis and related publications, 2018–2021.
11. O. Mangerel, "On the bivariate Erdős–Kac theorem," Proc. London Math. Soc., 2021.