Inclusion-Exclusion Prime Layers
Inclusion-Exclusion Prime Layers
Core Principle
Most ONM analysis treats “prime vs composite” as a binary distinction. The ω-decomposition reveals it is layered: every integer sits in an ω-layer defined by its number of distinct prime factors.
| ω | Name | Example | Character |
|---|---|---|---|
| 0 | Unit | 1 | Source |
| 1 | Primes | 2, 3, 5, 7 | Atomic — derivative = 1 |
| 2 | Semiprimes | 6, 10, 15 | First composition — derivative = sum of factors |
| 3 | 3-factor | 30, 42, 66 | Cross-terms dominate |
| k | k-factor | products of k primes | eₖ₋₁ complexity |
The Alternating Structure
When weighted by μ(n), each layer contributes with sign (−1)^k. This is inclusion-exclusion — the same combinatorial principle behind:
- Euler’s totient φ(n) = n ∏(1 − 1/p)
- The Möbius inversion formula
- Sieve methods in prime counting
The derivative-weighted Mertens function M’(N) makes this vivid: at N=10,000, the layers contribute magnitudes summing to ~10 million, yet the net is only ~50 thousand (99.5% cancellation).
ONM Interpretation
Entropy as Layer Cancellation
Each ω-layer represents a tier of compositional complexity. The alternating signs mean:
- ω=1 (primes): negative contribution — primes subtract from composite smoothing
- ω=2 (semiprimes): positive — first level of composition adds structure
- ω=3: negative — deeper composition reverses again
- Net effect: near-perfect cancellation → entropy (smooth composite average)
This IS entropy as composite smoothing, made precise: the Möbius function ensures that prime structure and composite structure nearly cancel, leaving only a sublinear residual.
Membrane Selectivity Is Layer-Dependent
Nagaπ’s Law says “only primes are membrane-permeable by default.” The ω-decomposition refines this:
- ω=1: freely permeable (Goldman ions K⁺=19, Na⁺=11, Cl⁻=17)
- ω=2: mimicry required, but possible (heavy metals mimicking essential elements)
- ω=3+: increasingly opaque — too much compositional structure to pass through
Model Collapse Is Layer Collapse
Each generation of LLM self-training corresponds to moving up one ω-layer:
- Gen 0 (human data): contains ω=1 prime information
- Gen 1 (model output): recombination → ω=2 composite
- Gen 2+: further recombination → higher ω, but alternating signs mean each generation partially undoes the previous
- The net signal (M’(N)) shrinks sublinearly → model collapse
The Felin & Holweg insight (“Theory Is All You Need”): human theorising operates at ω=1 (prime injection from outside the dataset). LLM prediction operates at ω≥2 (composite recombination of existing data).
Mathematical Backing
- Growth rate per layer: |layer_k(N)| ~ N^(0.84 + 0.5k) approximately
- Layer counts follow Poisson distribution peaking at ω ~ log log N
- Cancellation stable at 97-99.5% across N=200 to N=10,000
- Corrected Dirichlet series (no Euler product): F(s) = Σ_p p^{−s} ∏_{q≠p}(1 − q^{1−s})
Status
Established as formal tool. The mathematical content is classical (Möbius inversion, elementary symmetric polynomials). The ONM contribution is the interpretation: connecting ω-layers to compositional complexity tiers, membrane permeability gradients, and information-theoretic entropy.
Validated in “Stress-Testing a Metaphor” (pwt.life, Sep 2026).