dm³ 102 · Week 11 · Open problem

The Terras/Everett Argument — What Closing the Gaps Takes

Reading a real, self-correcting account of two open Lean sorries
dm³ 102 · Week 11 · Open problem
The Terras/Everett Argument — What Closing the Gaps Takes
Course: dm³ 102  ·  Open problem  ·  Source: CollatzDescent.lean, §7

The file’s own \(\S\)7 names the two open gaps precisely. Sorry #A (orbit unrolling, \(n\equiv1\bmod4\) branch): needs lemma orbit_halving (m k : ℕ) (hdvd : 2^k ∣ m) : orbit m k = m / 2^k — an elementary induction on \(k\), described in the file as “fully provable in Lean with existing Mathlib,” omitted only because wiring it into the surrounding proof needs roughly 15 more lines of bookkeeping.

Sorry #B (\(n\equiv3\bmod4\) branch): genuinely harder. Strong induction on \(n\)’s value fails here, because one odd step from \(n\equiv3\bmod4\) produces \((3n+1)/2>n\) — the induction hypothesis doesn’t apply to a larger number. The file’s documented fix: induct on \(\text{{Nat.log}}_2 n\) (bit-length) instead of on \(n\) directly. A “double burst” (one odd step, then repeated bursts) multiplies by less than \((3/2)^2=9/4<4\), so the bit-length grows by at most 1 per double-step, and the number of times the orbit can stay above \(n\) is bounded — this is the standard Terras/Everett argument for Collatz, made constructive.

The file corrects its own earlier estimate — read this passage directly
Verbatim from \(\S\)7: “This requires approximately 80–100 additional lines of Lean and is NOT ‘~40 lines.’ The earlier claim that sorry #1 was ‘closeable in ~40 lines’ was incorrect. It conflated the provable \(n\equiv1\pmod4\) sub-lemma with the general odd case. Both sorries here are mathematically closeable; neither is trivial.” This is exactly the discipline this course has asked you to practice since 101 Week 1 — not just distinguishing proved from open, but catching and correcting your own prior overconfidence about how hard an open gap actually is. A course built on real formalization should show you what that correction looks like when it happens inside the actual research, not just when an AI-generated course description gets it wrong.
This week’s content is grounded directly in the AXLE Lean sources cited above — no material in this page depends on the external, unverified source removed from earlier drafts of this course.
-- dm³ 102 · Week 11 · The two open gaps, precisely

-- Sorry #A (n ≡ 1 mod 4): needs
--   lemma orbit_halving (m k) (hdvd : 2^k ∣ m) : orbit m k = m / 2^k
-- "Fully provable with existing Mathlib" — elementary induction on k,
-- omitted only for ~15 lines of wiring/bookkeeping.

-- Sorry #B (n ≡ 3 mod 4): genuinely harder.
-- Strong induction on VALUE fails: one odd step gives (3n+1)/2 > n.
-- Fix: induct on Nat.log 2 n (bit-length) instead — the standard
-- Terras/Everett argument, made constructive. "Double burst"
-- multiplies by < 9/4 < 4, bounding bit-length growth.

-- File's own correction: "~80-100 lines... NOT ~40 lines... the
-- earlier claim... was incorrect... Both sorries are mathematically
-- closeable; neither is trivial." — an honest re-scoping, in the
-- actual source, not a hypothetical example.
example : True := trivial