The Anti-Correlation Engine: Variance Structure of ρ_E and the Local Smoothness of Integer Complexity

§1. Introduction

§1.1. Context

Paper 17 (DOI: 10.5281/zenodo.19016958) reduced the proof of $D(N) \to 1$ to four principal relay inputs plus Lemma II. The most leveraged input is $\sigma(G \mid \Omega = k) = O(1)$ — the formalization of Paper 13's Gaussian model. The present paper reports the M8 exploration results that resolve the variance structure of $\rho_E$ on $\Omega$-shells, identify three dead proof routes, and identify the most leveraged remaining target.

§1.2. Notation

As in Paper 17 §1.2. Additionally: $c^* = 1/\ln 2 \approx 1.4427$ (working normalization constant; not a proved asymptotic constant), $d_\rho(n) = \rho_E(n) - c^* \ln n$ (deficiency relative to $c^*$), $f(n) = \sum_{q^a \| n} \rho_E(q^a)$ (additive part), $r(n) = \rho_E(n) - f(n)$ (combinatorial remainder), $\Delta f(n) = f(n-1) - f(n)$, $\Delta r(n) = r(n-1) - r(n)$.

§1.3. Main Results

§2. Shell Deficiency Drift

§2.1. The Deficiency and Its Natural Scale

Define $d_\rho(n) = \rho_E(n) - c^* \ln n$ where $c^* = 1/\ln 2 \approx 1.4427$ is the numerical best-fit constant for $\rho_E(n)/\ln n$ (a working normalization constant; Paper 11 establishes $\rho_E(n) = \Theta(\ln n)$ but does not prove the existence of a precise asymptotic constant). On the $\Omega = k$ shell, the question is whether $\mathrm{Var}(d_\rho)$ is $O(1)$, $O(k)$, or $O(\ln\ln N)$.

§2.2. Data: $\mathrm{Var}(d_\rho)$ vs $k$ and $N$

$k$-direction ($N = 10^7$ fixed): $\mathrm{Var}(d_\rho)$ monotonically decreases from $11.6$ ($k=2$) to $6.6$ ($k=18$) across 17 values of $k$, zero exceptions.

$N$-direction ($k$ fixed): seven-point scan from $N = 10^4$ to $10^7$ shows positive growth for all tested $k$. Fit: $\mathrm{Var}(d_\rho) \approx a(k) + b(k) \cdot \ln\ln N$ with $b(k) > 0$ increasing significantly with $k$ (from $1.26$ at $k=4$ to $7.53$ at $k=10$).

§2.3. Resolution of the Apparent Contradiction

In the $k$-direction at fixed $N$, $\mathrm{Var}(d_\rho)$ decreases because higher-$k$ shells are more constrained (more prime factors → tighter factorization control → less room for $d_\rho$ to deviate). In the $N$-direction at fixed $k$, $\mathrm{Var}(d_\rho)$ increases because larger $N$ introduces numbers with larger primes, whose deficiency behavior has wider spread.

The two effects coexist: the $k$-dependent coefficient $b(k)$ grows faster than linearly in $k$ on the tested range, so on the Erdős-Kac central shell $k \sim \ln\ln N$, $\mathrm{Var}(d_\rho)$ is expected to grow at least as fast as $\Theta((\ln\ln N)^2)$, though the precise scaling remains open.

§2.4. Why Shell Concentration Fails

The additive part $f(n) = \sum_{q^a \| n} \rho_E(q^a)$ has $\mathrm{Var}(f(n) \mid \Omega = k) \approx 16$ — already large. The remainder $r = \rho_E - f$ anti-correlates with $f$ (forced by the DP min), keeping $\mathrm{Var}(\rho_E)$ below $\mathrm{Var}(f)$. But this compression is not strong enough to reach $O(1)$; it only reduces $\sim 16$ to $\sim 10$.

§3. The Anti-Correlation Engine

§3.1. Structural Origin

$r(n) = \rho_E(n) - f(n)$ by definition. Therefore $\mathrm{Cov}(f, r) = \mathrm{Cov}(f, \rho_E - f) = \mathrm{Cov}(f, \rho_E) - \mathrm{Var}(f)$.

If $\rho_E$ were unconstrained (i.e., $r$ independent of $f$), then $\mathrm{Cov}(f, r) = 0$. The negative covariance arises precisely because the DP min operation caps $\rho_E$: when $f(n)$ is large (unfavorable additive structure), $\rho_E(n)$ cannot fully follow $f$ upward (the min with $M_n$ or $S_n$ pulls it back), so $r = \rho_E - f$ becomes more negative. Conversely, when $f(n)$ is small, $\rho_E$ may exceed $f$ (via the successor route), making $r$ more positive.

§3.2. The Same Mechanism for Differences

$A(n) = \rho_E(n-1) - \rho_E(n) = \Delta f(n) + \Delta r(n)$.

Both $\Delta f$ and $\Delta r$ individually have variance $\sim 7$–$10$ (growing with $k$), but their forced anti-correlation ($\mathrm{Cov} \approx -7.5$) compresses $\mathrm{Var}(A)$ to $\approx 1$.

The mechanism is the same: $\Delta r = \Delta \rho_E - \Delta f = A - \Delta f$. So $\mathrm{Cov}(\Delta f, \Delta r) = \mathrm{Cov}(\Delta f, A) - \mathrm{Var}(\Delta f)$.

If $\mathrm{Var}(A) = O(1)$ and $\mathrm{Var}(\Delta f) = \Theta(\ln\ln N)$, then $\mathrm{Cov}(\Delta f, A) = O(\sqrt{\ln\ln N})$ (by Cauchy-Schwarz), and:

$$\mathrm{Cov}(\Delta f, \Delta r) = \mathrm{Cov}(\Delta f, A) - \mathrm{Var}(\Delta f) \approx -\mathrm{Var}(\Delta f)$$

which explains the near-perfect cancellation.

§3.3. Remark: A Random-Walk Analogy

The numerical picture is suggestive of a random-walk structure: on the $\Omega = k$ shell, $d_\rho$ appears to have total variance growing as $\sim b(k) \cdot \ln\ln N$ while the increment variance $\mathrm{Var}(d_\rho(n-1) - d_\rho(n))$ remains $O(1)$ and the increments are approximately Gaussian (kurtosis $\approx 3$). This resembles a process with stationary increments. The analogy is heuristic; the proof of $\sigma(G) = O(1)$ requires formalizing the stationary-increment property, not the full random-walk structure.

§4. Eliminated Proof Routes

§4.1. Route E1: Turán-Kubilius on $\Delta f$

Target: $\mathrm{Var}(f(n-1) - f(n) \mid \Omega = k) = O(1)$.

Why it fails: $\mathrm{Var}(\Delta f) \approx 10$ at $N = 10^7$ and grows with $N$. Kátai-Subbarao's conditional TK gives $\mathrm{Var}(f(n) \mid \Omega = k) = O_k(1)$ (constrained side), but $f(n-1)$ is unconstrained ($\Omega(n-1)$ free), giving $\mathrm{Var}(f(n-1)) = O(\ln\ln N)$ by standard TK. With approximate independence of $n$ and $n-1$ (a heuristic supported by Elliott-type results on correlations of multiplicative functions), $\mathrm{Var}(\Delta f) \approx \mathrm{Var}(f(n-1)) + \mathrm{Var}(f(n)) - 2\mathrm{Cov} \approx O(\ln\ln N)$. The decisive evidence for E1 is the direct computation: $\mathrm{Var}(\Delta f) \approx 10$ at $N = 10^7$.

To the author's knowledge, no existing shifted-shell TK theorem resolves this; see Mangerel [6] for the current state of additive functions in short intervals and gaps.

§4.2. Route E2: Separate $f$ and $r$

Target: Prove $\mathrm{Var}(\Delta f) = O(1)$ and $\mathrm{Var}(\Delta r) = O(1)$ separately.

Why it fails: Both are $\gg 1$. The $O(1)$ variance of $A$ comes entirely from their cross-cancellation, not from individual boundedness.

§4.3. Route E3: Shell Concentration of $d_\rho$

Target: $\mathrm{Var}(d_\rho \mid \Omega = k) = O(1)$.

Why it fails: Seven-point $N$-scan confirms growth $\sim b(k) \cdot \ln\ln N$ with $b(k)$ increasing significantly in $k$.

§5. The Most Leveraged Proof Target: Local Lipschitz Property

§5.1. Statement

§5.2. Why This Is the Most Leveraged Route toward $\sigma(G) = O(1)$

By Proposition J (Paper 17): $A = j - 1 = \max(G, 0) - 1$, so $\mathrm{Var}(A) = \mathrm{Var}(\max(G, 0)) \leq \mathrm{Var}(G)$.

The cleaner statement: $\mathrm{Var}(G_{\mathrm{spf}}) = \mathrm{Var}(A) + \mathrm{Var}(B) + 2\mathrm{Cov}(A,B)$. With $\mathrm{Var}(A) = O(1)$ (Local Lipschitz), $\mathrm{Var}(B) = O_k(1)$ (B-bound, numerical), and $|\mathrm{Cov}(A,B)| \leq \sqrt{\mathrm{Var}(A) \cdot \mathrm{Var}(B)} = O_k(1)$: $\mathrm{Var}(G_{\mathrm{spf}}) = O_k(1)$.

Numerically, $\mathrm{Var}(G)$ and $\mathrm{Var}(G_{\mathrm{spf}})$ remain close on the tested range (within $\pm 0.1$ for $k \geq 5$), with no evidence of variance blow-up in the passage from $G_{\mathrm{spf}}$ to $G$. So $\mathrm{Var}(G) = O_k(1)$.

§5.3. Connection to the DP Recursion

$\rho_E(n) = \min(\rho_E(n-1) + 1, M_n)$, so $A(n) = \rho_E(n-1) - \rho_E(n) = \max(\rho_E(n-1) + 1 - M_n, 0) - 1 = j(n) - 1$ (Proposition J).

The DP guarantees $A \geq -1$ (successor bound). The upper tail of $A$ is controlled by how much $M_n$ can undershoot $S_n$. On the $\Omega = k$ shell, $M_n$ is determined by $n$'s factorization, and $S_n = \rho_E(n-1) + 1$ is determined by the predecessor. Their near-cancellation to $O(1)$ is the local Lipschitz property.

§5.4. Possible Proof Directions for Local Lipschitz

• Direction 1: DP-induced smoothing. The recursion $\rho_E(n) = \min(S_n, M_n)$ creates a feedback loop: $S_{n+1} = \rho_E(n) + 1 = \min(S_n, M_n) + 1$. When $M_n \ll S_n$ (large jump), the next successor $S_{n+1}$ resets to $M_n + 1$, reducing the "memory" of previous values. This smoothing effect may bound the variance of consecutive differences.
• Direction 2: Quantitative smoothness of $\rho_E$. Numerically, $|\rho_E(n) - c^* \ln n|$ appears sublinear in $\ln n$ on the tested range. Since $c^*(\ln n - \ln(n-1)) = c^*/n + O(1/n^2) \to 0$, we have $|d_\rho(n) - d_\rho(n-1)| \leq |A(n)| + c^*/n$. This reduces to bounding $|A(n)|$, which is circular.
• Direction 3: Conditional second-moment method. Prove directly that $E[A^2 \mid \Omega = k] = O(1)$ using the insertion identity (Paper 16 Theorem 1) and the A/B decomposition.

§6. Complete Variance Anatomy

§6.1. $\mathrm{Var}(A)$ and $\mathrm{Var}(B)$ Across All $k$

Observations: $\mathrm{Var}(B)$ converges to $1/4$ at high $k$, consistent with a Bernoulli-like distribution of $B$ values concentrated near $\{0, -1\}$. $\mathrm{Var}(A)$ saturates at $\approx 1.3$. $\mathrm{Cov}(A, B)$ is small and negative ($\approx -0.10$), indicating weak coupling.

§6.2. $G$ vs $G_{\mathrm{spf}}$

$G(n) \geq G_{\mathrm{spf}}(n)$ always (Paper 15, Theorem A). Non-SPF splits can be strictly better, making $G > G_{\mathrm{spf}}$:

At low $k$, non-SPF optimal splits are common and $\mathrm{Var}(G) < \mathrm{Var}(G_{\mathrm{spf}})$. At high $k$, $\mathrm{Var}(G) \approx \mathrm{Var}(G_{\mathrm{spf}})$, with $\mathrm{Var}(G)$ slightly exceeding $\mathrm{Var}(G_{\mathrm{spf}})$ by up to $0.1$.

§7. The Revised Proof Landscape

§7.1. What M8 Establishes

Discoveries (numerical, strong evidence on tested range $N \leq 10^7$, $k \leq 18$):

• Shell deficiency drift: $\mathrm{Var}(d_\rho)$ consistent with $a(k) + b(k) \cdot \ln\ln N$, $b(k) > 0$ increasing in $k$
• Local smoothness: $\mathrm{Var}(A)$ in $[0.3, 1.35]$ ($[0.8, 1.35]$ for $k \geq 5$; seven-point $N$-scan, zero growth trend)
• Anti-correlation engine: $\mathrm{Cov}(\Delta f, \Delta r) \approx -\mathrm{Var}(\Delta f)$, appears forced by DP min
• Gaussian character: kurtosis $\in [2.62, 2.92]$, skewness $\to 0$
• Variance anatomy: $\mathrm{Var}(B)$ converges toward $1/4$, $\mathrm{Var}(A)$ saturates near $1.3$

Eliminated routes: E1 (TK on $\Delta f$), E2 ($f/r$ separate), E3 (shell concentration)

§7.2. Updated Open Inputs for $D(N) \to 1$

§7.3. Proof Dependency Graph (Updated)

$$\text{Local Lipschitz (N)} \longrightarrow \mathrm{Var}(A) = O(1) \xrightarrow{\text{+ B-bound (N)}} \mathrm{Var}(G_{\mathrm{spf}}) = O_k(1)$$

$$\mathrm{Var}(G_{\mathrm{spf}}) = O_k(1) + \text{Prop C (T)} + \text{Scale mono (N)} \longrightarrow \text{I-b}$$

$$\text{I-b} + \text{I-a (T high-}k\text{ / N fixed-}k\text{)} \longrightarrow \text{Lemma I}$$

$$\text{Lemma I} + \text{Lemma II (N)} + \text{B-bound (N)} \longrightarrow \text{Thm 5 (Paper 16, T)}$$

$$\text{Thm 5} + \text{SPF-positivity (N)} + \text{Sathe-Selberg (T)} \longrightarrow D(N) \to 1$$

Methodological Note: Four-AI Parallel Exploration and the Thermodynamic Interface

All mathematical claims, citations, and computations in this paper were independently checked and are the author's responsibility. AI systems were used for literature search, exploratory computation, and drafting support.

A.1. Four-AI Contributions to M8

The M8 results were developed through structured parallel exploration with four AI systems, following the protocol documented in Paper 16 Appendix.

ChatGPT conducted the literature search that proved decisive: identifying Kátai-Subbarao's conditional Turán-Kubilius [4], Goudout's shifted Erdős-Kac [5], and Mangerel's gap theory [6] as the relevant boundary of existing results. ChatGPT's prediction that $\mathrm{Var}(\Delta f) = O(\ln\ln N)$ (not $O(1)$) was confirmed by Gemini's data, and ChatGPT designed the $k$-vs-$N$ scaling test (§2.2) that resolved the $O(1)$ vs $O(\ln\ln N)$ question for $\mathrm{Var}(d_\rho)$.

Gemini computed the $f + r$ variance decomposition (§3) that revealed the anti-correlation engine: $\mathrm{Var}(\Delta f) \approx 10$, $\mathrm{Var}(\Delta r) \approx 7$, $\mathrm{Cov}(\Delta f, \Delta r) \approx -7.5$. This was the single most important computation of M8.

Grok computed $\mathrm{Var}(G)$ vs $\mathrm{Var}(G_{\mathrm{spf}})$ across all $k$ (§6), the kurtosis data confirming Gaussian character, the 17-point $k$-scan and 7-point $N$-scan of $\mathrm{Var}(d_\rho)$ (§2.2), and the $\mathrm{Var}(A)$ $N$-stability scan that confirmed local smoothness is independent of shell drift.

Claude proposed the initial $\mathrm{Var}(d_\rho) = O(1)$ conjecture (refuted), identified the tautological nature of the $S - M$ and $j - \Delta$ decompositions (§4, eliminating pure-DP routes), and formulated the Local Lipschitz target (§5.1).

A.2. Role of the Thermodynamic Interface

A parallel thread on the thermodynamic interpretation of ZFCρ contributed to M8 in three specific ways:

• (i) Priority setting. The thermodynamic analysis identified $\sigma(G) = O(1)$ as the single most leveraged target ("one stone, three birds": A-concentration + Prop C input + fluctuation-dissipation bridge), directly determining M8's attack choice over alternatives (scale monotonicity, I-a).
• (ii) Fluctuation-dissipation framing. The thermodynamic interpretation of the anti-correlation engine as a discrete fluctuation-dissipation relation — where the DP min operation plays the role of a dynamical constraint linking fluctuations ($\Delta f$) to dissipation ($\Delta r$) — provided the conceptual framework that guided the search for the mechanism behind $\mathrm{Var}(A) = O(1)$.
• (iii) Direction correction. The thermodynamic perspective's emphasis on local dynamics ("the system's constraint limits each step, not the global position") directly suggested the shift from shell concentration to local Lipschitz as the proof target — a suggestion subsequently confirmed by the $N$-scaling data.

These cross-domain contributions illustrate how physical intuition can guide the formulation (though not the proof) of mathematical conjectures.

References

1. H. Qin, "Monotonicity, roughness stability, and the narrowing of the relay architecture" (Paper 17), Zenodo, DOI: 10.5281/zenodo.19016958.
2. H. Qin, "The insertion identity, variance splitting, and the relay engine of Conjecture H'" (Paper 16), Zenodo, DOI: 10.5281/zenodo.19013602.
3. H. Qin, "Zero-inflated lattice normal model" (Paper 13), Zenodo, DOI: 10.5281/zenodo.18991986.
4. I. Kátai and M. V. Subbarao, "Some remarks on a paper of Ramachandra," Lithuanian Math. J. 48 (2008), 170–183.
5. É. Goudout, "Lois de répartition des diviseurs," doctoral thesis, 2018–2021.
6. O. Mangerel, "Additive functions in short intervals, gaps and a conjecture of Erdős," Int. Math. Res. Not. (2022).
7. O. Mangerel, "On the bivariate Erdős-Kac theorem," Proc. London Math. Soc., 2021.
8. G. Tenenbaum, Introduction to Analytic and Probabilistic Number Theory, 3rd ed., AMS, 2015.