Tao 2019 Collatz Blueprint

2 The support layer

Discrete-probability, generating-function, and Fourier scaffolding used everywhere below.

2.1 PMF calculus

Definition 2.1 Discrete-probability calculus
#

Total variation distance \(d_{\mathrm{TV}}(p,q) := \sum _a' |(p\, a).\mathrm{toReal} - (q\, a).\mathrm{toReal}|\), real and complex expectations \(\mathbb {E}_p[f] := \sum _a' (p\, a).\mathrm{toReal}\cdot f(a)\), and the \(n\)-fold i.i.d. product \(p^{\otimes n}\) on \(\mathrm{Fin}\, n\to \alpha \) (node D1). Together with the coupling inequality (paper (1.10)), stated in indicator-expectation form to avoid OuterMeasure friction: for any event \(E\subseteq \alpha \),

\[ \bigl|\mathbb {E}_p[\mathbf1_E] - \mathbb {E}_q[\mathbf1_E]\bigr| \le d_{\mathrm{TV}}(p,q). \]

Basic algebra of \(d_{\mathrm{TV}}\) (symmetry, non-negativity) and the support/apply lemmas of the i.i.d. product are part of this node.

Campaign estimate: 8–15 treadmill laps; risk low (90% confidence the node completes as stated).

2.2 The geometric family and local bounds

Definition 2.2 Geometric and Pascal laws; negative binomial
#

The lattice laws driving the Syracuse valuation: \(\mathrm{Geom}(2)\) with \(\mathbb {P}(a)=2^{-a}\) for \(a\ge 1\) (geomHalf); \(\mathrm{Geom}(4)\) (geomQuarter); \(\mathrm{Pascal}\), the law of \(a_1+a_2\) with \(a_i\sim \mathrm{Geom}(2)\), i.e. \(\mathbb {P}(b)=(b-1)2^{-b}\) for \(b\ge 2\) (pascal); and its conditioning off \(b=3\) (pascalNe3). The load-bearing computation is the exact negative-binomial point mass for the sum of \(n\) i.i.d. \(\mathrm{Geom}(2)\) variables,

\[ \mathbb {P}\! \Bigl(\textstyle \sum _{i{\lt}n} a_i = L\Bigr) = \binom {L-1}{n-1}\, 2^{-L}, \qquad L\ge n\ge 1 \]

(compositions of \(L\) into \(n\) positive parts), together with the moment generating functions and the one-dimensional Chernoff upper-tail bounds derived from them.

Campaign estimate: 6–12 treadmill laps; risk low (90% confidence the node completes as stated).

Risk kernel 1. Lemma 2.2(i)(ii): local-limit / Gaussian-type upper bounds on the point masses of sums of \(k\) i.i.d. copies of the relevant lattice distributions (\(\mathrm{Geom}(2)\), \(\mathrm{Geom}(4)\), \(\mathrm{Pascal}\), and the holding law \(\mathrm{Hold}\) of §7), uniform in the shift, of the shape \(\mathbb {P}(S_k = v)\ll k^{-d/2}\) with \(d\) the lattice dimension; together with the weight function \(G_n\) (paper (2.2)). Proved not by the paper’s contour/Morera route but by explicit generating functions plus real-variable exponential tilting (design decision D5): for the \(d=2\) holding sums this is the circle-method form \(\mathbb {P}(S_k=v)=(2\pi )^{-2}M(\lambda )^k e^{-\lambda \cdot v}\int _{[-\pi ,\pi ]^2} |\varphi _\lambda (t)|^k\, dt\) with \(\varphi _\lambda ,M\) explicit rational functions. Statements judge-ratified 2026-07-12 vs pp.14–15: per-instance forms in Prob/LocalInstances.lean (scalar) and Sec7/HoldLocal.lean (\(d=2\) Hold, with the correct \((n+1)^{-1}\) prefactor and product norm fed to the scalar \(G\)); the \(G_{1+n}\)-for-\(G_n\) index is constants-equivalent (in-paper precedent: Lemma 7.7).

Proof

Risk kernel 1 CLOSED (campaign laps 22–45; judge pass 5, 2026-07-12). All eight obligations — (i) and (ii) for \(\mathrm{Geom}(2)\), \(\mathrm{Geom}(4)\), \(\mathrm{Pascal}\), \(\mathrm{Hold}\) — machine-checked via the D5 route (circle-method inversion, character decay, exponential tilting, second-order MGF bounds with exact means, \(\lambda \)-clip Chernoff assembly), plus the generic engines iidSum_nat_local_of_quad / iidSum_nat_tail_of_quad. Judge-run #print axioms on all ten declarations (2026-07-12): [propext, Classical.choice, Quot.sound] — nothing extra.

2.3 Fourier analysis on \(\mathbb {Z}/3^n\mathbb {Z}\)

Definition 2.4 Additive characters and the oscillation functional
#

The additive character \(e(\theta ):=\exp (2\pi i\theta )\) on \(\mathbb {Q}\), the discrete Fourier transform and Parseval identity on \(\mathbb {Z}/3^n\mathbb {Z}\), and the oscillation functional \(\mathrm{Osc}_{m,n}\) (paper (1.24)) measuring how far a weight \(c:\mathbb {Z}/3^n\mathbb {Z}\to \mathbb {R}\) is from being constant on the fibres of the projection to \(\mathbb {Z}/3^m\mathbb {Z}\):

\[ \mathrm{Osc}_{m,n}(c)=\sum _{Y}\Bigl|\, c(Y)-3^{\, m-n}\! \! \sum _{Y'\mapsto \bar Y} c(Y')\, \Bigr|, \qquad m\le n. \]

Includes the Remark 1.18 triangle inequality that lets \(\mathrm{Osc}\) be controlled scale by scale.

Campaign estimate: 4–8 treadmill laps; risk low (85% confidence the node completes as stated).