Work

The parsimony elbow of the primorial tower: marginal efficiency of the sieve of Eratosthenes

Year: 2026
Type: preprint
Abstract: We study the marginal efficiency of the sieve of Eratosthenes along the primorial tower $m_k = \prod_{i=1}^k p_i$, defined for the transition $m_k \to m_{k+1}$ by $$E(k) := \frac{\rho(k) - \rho(k+1)}{\phi_{k+1} - \phi_k} = \frac{\rho(k)}{\phi_k \, p_{k+1} \, (p_{k+1} - 2)} = \frac{1}{m_k \, p_{k+1} \, (p_{k+1} - 2)},$$ where $\rho(k) = \varphi(m_k)/m_k$ is the survivor density and $\phi_k = \varphi(m_k)$ is the totient product. We prove that the transition $2 \to 6$ ($m_2 = 6$) is the unique global maximizer (the parsimony elbow) with $E(1) = 1/6$, $E(k) \le 1/90$ for all $k \ge 2$, and $E(k)$ strictly decreasing—accounting for $93.22\%$ of the convergent cumulative efficiency ($98.01\%$ in the quadratic interaction model $E_{\mathrm{int}}(k)$). We further characterize general increasing cost functions $C(\phi)$, establishing universal first-step dominance for all convex, power-law ($\phi^\alpha$, $\alpha > 0$), and logarithmic ($\log(1 + \phi)$) costs, as well as the exact bifurcation threshold $\lambda_* \approx 1.2523734$ under saturating costs $C_\lambda(\phi) = 1 - e^{-\lambda\phi}$. Repository Contents 1. Manuscript Parsimony_Elbow.pdf & Parsimony_Elbow.tex: Complete manuscript accompanied by vector figures. 2. Lean 4 Formal Verification Suite Formalization_pack.zip Lean_Parsimony_Elbow.ipynb & ParsimonyElbow_Lean4_Reproducibility_Suite.zip: Self-contained formal verification suite in Lean 4 with Mathlib4 at Level N1 ($215$ certified theorems/lemmas and $27$ verified definitions across six independent modules, with zero sorry placeholders and zero custom axioms beyond Lean's foundational axioms propext, Classical.choice, and Quot.sound), together with the per-module cryptographic manifest (SHA256SUMS): ParsimonyElbow.lean (63 theorems/lemmas, 6 definitions): Canonical prime sequence, primorial tower, totient identities, recurrences (Lemma 2.1), closed and simplified forms of $E(k)$ and $E_{\mathrm{int}}(k)$, linear and quadratic ratio laws, prime-gap correction $\delta(k) = g_k/(p_{k+2} - 2)$, product formula, and the linear and quadratic parsimony elbows at $m_2 = 6$ (Theorems 4.1 and 4.2). CumulativeBounds.lean (59 theorems/lemmas, 9 definitions): Exponential growth bounds $\phi_k \ge 8 \cdot 4^{k-3}$ ($k \ge 3$), summability and explicit rational upper bounds $\sum E(k) \le 122/675$ and $\sum E_{\mathrm{int}}(k) \le 1/18 + 1/900 + 1/16875$ (Proposition 4.5), bilateral tail enclosure (Corollary 4.6), and the exact Chebyshev representation $E(k) = e^{-\theta(p_k)} / \bigl(p_{k+1}(p_{k+1} - 2)\bigr)$ (Remark 3.10). GeneralCostElbow.lean (52 theorems/lemmas, 8 definitions): Generalized marginal efficiency $E_C(k)$, consecutive ratio identity, sharp threshold criterion $\Delta C_k / \Delta C_1 > B(k) = 6\rho(k)/p_{k+1}$ (Proposition 4.8), and the universal convex cost bound $E_C(k) \le \frac{1}{15} E_C(1)$ (Theorem 4.9(i)). WheelUnitStructure.lean (17 theorems/lemmas, 2 definitions): Exact unit sets $(\mathbb{Z}/m_k\mathbb{Z})^\times$ for $m_1 = 2$, $m_2 = 6$, and $m_3 = 30$, modulo-$6$ involution, nontrivial ordered-pair lift identity $(p - 1)^2 - 1 = p(p - 2)$ (Proposition 3.11(iv)), and open intermediate wheel-slot counts $N_{m_k}(a, g)$ modulo $6$ and $30$. OptimalityM6.lean (12 theorems/lemmas, 1 definition): Exact evaluations of $\varphi(m)$ and $\rho(m) = \varphi(m)/m$ for $m \in \{1, 2, 3, 4, 6\}$, certifying that $m_2 = 6$ uniquely minimizes survivor density among moduli with $\varphi(m) \le 2$ at $\rho(6) = 1/3$. EratosthenesTransition.lean (12 theorems/lemmas, 1 definition): Degeneracy of the parity step $m_0 = 1 \to m_1 = 2$ ($\phi_1 - \phi_0 = 0$, Remark 3.2), activation of the first channel at $2 \to 6$, and the exact $15\times$ and $50\times$ efficiency drop factors at $m_2 = 6$. 3. Computational Replication & Experimental Suite Experimental_pack.zip EXP_Parsimony_Elbow.ipynb & parsimony_elbow_replication_bundle.zip: Deterministic Python 3 replication suite (SEED = 42) generating all numerical tables, JSON manifests, SHA-256 checksums (replication_manifest_sha256.json), and publication-ready PDF/PNG figures: Exact Rational & Arbitrary-Precision Arithmetic (Tables 1–3): Symbolic verification via fractions.Fraction and $100$-digit mpmath evaluation of cumulative sums ($S_\infty = 0.178781422926204408449354583771\ldots$) and the saturation threshold $\lambda_* = -\log u_* \approx 1.2523734$. Asymptotic Profile & Series Tail up to $10^8$ (Tables 4–5): Sieve evaluation across all $5{,}761{,}455$ primes $\le 10^8$ comparing $p_k - \theta(p_k)$ against $L_k = \log E(k) + p_k + 2\log p_k$, correlating oscillations with the first $10$ non-trivial zeros of $\zeta(s)$ ($r = 0.751$), and validating the two-term tail approximation $1 + r_1 + r_1 r_2$. Multi-Scale Gap Deficit & Wheel Conditioning up to $10^9$: Odd-only Sieve of Eratosthenes across all $50{,}847{,}534$ primes $\le 10^9$ quantifying the $O(1/\log X)$ normalized gap tail deficit $D_t(X) \approx c_t / \log X$, testing mod-$6$ vs. mod-$30$ diagonal homogeneity, and validating $P_{\mathrm{wheel}}(a \mid g) \propto \exp\bigl(-N_{30}(a, g) / (\rho(3)\,\overline{\log p})\bigr)$ ($r = 0.9968$, mean $\mathrm{TVD} = 0.01012$). Hardy–Littlewood Singular-Series Verification: Exact inclusion-exclusion evaluation $P_{\mathrm{HL}}(a \mid g)$ over subsets of open wheel channels $\mathcal{C}_{a,g}$ weighted by the truncated Euler product $\mathfrak{S}_{>5}$ ($5 < p \le 5000$) across scales $X \in [10^6, 10^{20}]$. Reproducibility Instructions Both notebooks are self-contained and designed to run sequentially in Google Colaboratory or any standard local environment: Lean 4 Notebook: Automatically installs elan, synchronizes the exact Lean 4 toolchain pinned by Mathlib4, downloads precompiled .olean caches (lake exe cache get!), compiles all six modules with lake build, and runs an independent file-by-file kernel audit. Python Notebook: Requires Python 3 with numpy, scipy, mpmath, and matplotlib. Full execution through $X = 10^9$ completes in under two minutes on a standard CPU runtime. Status: Date - 17 Nov 2026 submitted to Revista Matemática Iberoamericana (RMI) - 261117-PeinadorSala-2.
Language: English
Open Access status: green
OpenAlex ID: W7235066516
Cites:
Cited by:
Related to: 10