Conjecture Unrefereed

Two mirrors (squares and doubled squares)

No dated manuscript catalogued; standalone verification archive linked above. On this site since · last revised · Verify & co-author
Kind
Conjecture; a computer-assisted check covers every number below the 1000th Pell level, about 8.9×107648.9 \times 10^{764}
Status
Holds up under agent review; not proved; unrefereed
Prior work
Rule and claim not found in the searches logged; nearest published relative Alwani and Vivaldi (2017)
On this page
  1. The statement
  2. Check it yourself
  3. Worked examples
  4. Status and who checked it
  5. Prior work
  6. Further reading

The statement

Cut the numbers 0,1,2,…0,\allowbreak 1,\allowbreak 2,\allowbreak \dots into blocks in two ways, and reverse each block (a mirror: inside a block from aa to bb, send nn to a+b−na + b - n).

  • A, the square mirror. Blocks start at the squares: [0], [1, 3], [4, 8], [9, 15], [16, 24], …. For nn in [s2,(s+1)2−1][s^2,\allowbreak (s+1)^2 - 1], A(n)=2s(s+1)−nA(n) = 2s(s+1) - n.
  • D, the doubled-square mirror. Blocks start at 2m22m^2: [0, 1], [2, 7], [8, 17], [18, 31], [32, 49], …. For nn in [2m2,2(m+1)2−1][2m^2,\allowbreak 2(m+1)^2 - 1], D(n)=(2m+1)2−nD(n) = (2m+1)^2 - n.

One step of ff is DD, then AA: f(n)=A(D(n))f(n) = A(D(n)).

Conjecture. For every whole number n≥0n \ge 0, applying ff again and again eventually gives nn back.

Each mirror undoes itself, so ff never merges two numbers or loses one: it is a reshuffle of the whole numbers. Every path is therefore either a cycle or runs off for ever in both directions; the question is whether any path runs off.

Check it yourself

  • Single steps. 2 lies in the D block [2, 7], so D(2)=7D(2) = 7; 7 lies in the A block [4, 8], so f(2)=4+8−7=5f(2) = 4 + 8 - 7 = 5. Likewise f(8)=23f(8) = 23: D(8)=8+17−8=17D(8) = 8 + 17 - 8 = 17, and A(17)=16+24−17A(17) = 16 + 24 - 17.
  • The cycle of 0: 0→3→6→1→00 \to 3 \to 6 \to 1 \to 0.
  • The cycle of 2 has 28 numbers and never goes above 47: 2 → 5 → 8 → 23 → 34 → 37 → 40 → 43 → 46 → 25 → 16 → 15 → 14 → 13 → 12 → 11 → 10 → 9 → 24 → 35 → 38 → 41 → 44 → 47 → 26 → 17 → 4 → 7 → 2.
  • A cycle of 4 at the level 8: 56 → 79 → 90 → 65 → 56. By hand: 56 lies in the D block [50, 71], so D(56)=65D(56) = 65; 65 lies in the A block [64, 80], so A(65)=79A(65) = 79; 79 lies in the D block [72, 97], so D(79)=90=9×10D(79) = 90 = 9 \times 10.
  • Every cycle length is a multiple of 4. f(n)−n=2s(s+1)−(2m+1)2f(n) - n = 2s(s+1) - (2m+1)^2: a multiple of 4 minus an odd square, so it always leaves 3 when divided by 4. After LL steps the path has moved by 3L3L modulo 4, which is 0 only when LL is a multiple of 4.

Worked examples

Figure 1. The numbers 0 to 63. Above the line, the square blocks and, as nested arcs, which number A swaps with which; below it, the doubled-square blocks and D. The dark arrows trace two steps of ff from the cycle of 2: D takes 8 to 17, A takes 17 to 23, D takes 23 to 26, A takes 26 to 34. At 8 and 9, and at 49 and 50, a doubled-square block and a square block start one apart: the Pell levels 8 and 49.Full size
Figure 2. The whole cycle of 2. Each row gives nn, its doubled-square block and D(n)D(n), then the square block of D(n)D(n) and f(n)f(n), which is the next row's nn. On the right, the same step on the line 0 to 47.Full size

Status and who checked it

  • Proved (short arguments by agents, each re-derived by a second agent):
  • ff is a reshuffle, and every cycle length is a multiple of 4.
  • The gate lemma. A square and a doubled square are never equal, but they come one apart at the Pell levels g=1,8,49,288,1681,9800,…g = 1,\allowbreak 8,\allowbreak 49,\allowbreak 288,\allowbreak 1681,\allowbreak 9800,\allowbreak \dots (OEIS A001108). At each level exactly one number below gg crosses to gg or above, and exactly one crosses back, both on the path through gg. So at most one path can run off, and the conjecture is the same as "the path through every Pell level comes back".
  • A barrier. A made-up pair of mirrors with the same parities, the same pattern modulo 4 and the same one-door gates has a path that runs for ever, so no argument from those features alone can prove the claim.
  • Checked by computer. The paths through levels 1 to 25 were followed step by step: level 25 is g = 3,441,313,796,169,221,281, and its path takes 13,652,206,968 steps to come back. A method that cancels whole blocks against each other (the way Euclid's algorithm cancels one number against another) then closed every level up to 241, and levels 400 and 1000. So every path that meets a number below the 1000th level, a 765-digit number, is a cycle. Three programs agree at level 1000; the method was checked against the direct walks at levels 1 to 25. The programs are not formally verified, and the agents grade this range medium-low.
  • Open: the claim itself, above the 1000th level. Level 2000 was started and stopped unfinished; it is not claimed.
  • The neighbours are settled. Pair the square mirror with blocks at the triangular numbers instead, and every path comes back (proved: the square triangular numbers 1, 36, 1,225, … are exact walls). Pair it with blocks at the pronic numbers s(s+1)s(s+1), and no path ever comes back (proved). With the doubled squares there is no exact wall, only the one-door near-walls.
  • Who checked it: a falsifier agent and a reviewer agent, each with its own code; later an agent that tried to prove it, and a checker agent that tried to refute that work and reported "No claim was refuted". No human referee has read it.

Prior work

Searches run. The falsifier agent ran 32 OEIS searches, 8 arXiv searches that returned a record, 16 web searches, and searched Lagarias's two annotated 3x+1 bibliographies; the reviewer agent ran 23 OEIS, 5 arXiv and 6 web searches. On 2026-11-19 a keyed CORE search made 7 requests and got 11 records, 1 of them a relative. On 2026-11-20 two more agents searched again with their own logs (OEIS by computed terms, arXiv, zbMATH, Math Stack Exchange and MathOverflow, Zenodo, DataCite, HAL, GitHub, and the formal-conjectures collection of Lean statements). The map, the claim and its pairing form were not found in the searches logged.

Nearest relatives, credited to their authors.

  • Alwani and Vivaldi, "Nonlinear Rotations on a Lattice" (2017, published 2018), the nearest published relative found. Their return map is a product of two block mirrors on block systems of the same growth rate, with a conjecture of the same kind (all orbits periodic, or all escaping), proved for a set of parameters of full density. The pair here has two different rates and is not in their family.
  • Eppstein, "Reflections in an octagonal mirror maze" (2022): a general method for following a path among mirrors to its end, a precedent for the block-cancelling step.
  • Finite meanders (Coll, Giaquinto and Magnant; Coll, Magnant and Wang): ff is their "meander permutation" for two finite block lists. Lu (October 2026) defines the infinite meander of two infinite sets and studies random sets only; the random "infinite noodle" of Curien, Kozma, Sidoravicius and Tournier proves at most one infinite path.
  • OEIS A020703 (the square mirror alone) and A061579 and A038722 (the triangular mirror alone): single mirrors, no composition.

Still owed: Google Scholar; Guy's Unsolved Problems in Number Theory; the Lowenstein–Vivaldi papers beyond their abstracts; the seqfan mailing-list archive; Fiedler (2014) beyond its summary.

Further reading