Conjecture Unrefereed

Collatz with memory has a loop of every even length

No dated manuscript catalogued; standalone verification archive linked above. On this site since · last revised · Verify & co-author
Kind
Conjecture; a loop found and checked exactly at every even length from 12 to 36 (333 loops)
Status
Holds up under agent review; two parts proved; open from length 38 on; unrefereed
Prior work
Not found in the searches logged; second-order Collatz rules are in print
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

The odd part of a number is what is left after dividing out every factor 2 (the odd part of 24 is 3). Keep two numbers, the last one and the one before it. The next number is

xn+1=odd part of (2xn+xn−1+1).x_{n+1} = \text{odd part of } (2x_n + x_{n-1} + 1).

The state is the pair of the two newest numbers, so a loop closes when a pair repeats. Its length is the number of steps in one lap.

Conjecture. For every even L≥12L \ge 12 the rule has a loop of exactly LL steps. Hence it has infinitely many loops.

Small starts never see these loops. Every pair with both numbers at most 16,000 ends in one of four small loops (of lengths 1, 2, 54 and 29), and so does every pair of odd numbers up to 200,001. The loops of this page all live above 760,000.

Check it yourself

The smallest loop of length 12. Start from the pair (768,247, 3,121,313) and apply the rule 12 times:

sum 2xn+xn−1+12x_n + x_{n-1} + 1 divided by next
2 × 3,121,313 + 768,247 + 1 = 7,010,874 2 3,505,437
2 × 3,505,437 + 3,121,313 + 1 = 10,132,188 4 2,533,047
2 × 2,533,047 + 3,505,437 + 1 = 8,571,532 4 2,142,883
2 × 2,142,883 + 2,533,047 + 1 = 6,818,814 2 3,409,407
2 × 3,409,407 + 2,142,883 + 1 = 8,961,698 2 4,480,849
2 × 4,480,849 + 3,409,407 + 1 = 12,371,106 2 6,185,553
2 × 6,185,553 + 4,480,849 + 1 = 16,851,956 4 4,212,989
2 × 4,212,989 + 6,185,553 + 1 = 14,611,532 4 3,652,883
2 × 3,652,883 + 4,212,989 + 1 = 11,518,756 4 2,879,689
2 × 2,879,689 + 3,652,883 + 1 = 9,412,262 2 4,706,131
2 × 4,706,131 + 2,879,689 + 1 = 12,291,952 16 768,247
2 × 768,247 + 4,706,131 + 1 = 6,242,626 2 3,121,313

The pair (768,247, 3,121,313) is back. One lap halves 20 times in all.

The other lengths. 333 loops of lengths 12 to 36 have been solved and walked in exact integers. The smallest number in the loop of length 36 is 1,460,685,166,063,904,419.

Worked examples

How a loop is found without trying starts. Fix how many times each step halves (for the loop above: 1, 2, 2, 1, 1, 1, 2, 2, 2, 1, 4, 1). Then every number of the lap is a linear formula in the unknown starting pair, and "the pair comes back" is two linear equations. Their answer is usually a fraction or negative; when it is a pair of positive odd whole numbers, it is a loop. For the pattern above the answer is exactly (768,247, 3,121,313).

The proved criterion. Whether the answer is whole comes down to one number made from the pattern: a sum of powers of 2, one for each way of choosing places on a ring of LL places with no two neighbours chosen. When that sum is exactly 2S2^S, where SS is the total number of halvings, the pattern gives a real loop. For the loop above the ring of 12 places has 322 such choices, and their powers of 2 add up to 2202^{20} = 1,048,576. So the conjecture is a counting puzzle about adding powers of 2; no arithmetic of the rule is left in it.

Figure 1. The smallest loop of length 12, clockwise from its least term 768,247. Node colour says how many halvings made the number (light ÷2, mid ÷4, dark ÷16).Full size
Figure 2. The same 12 steps, each sum also written in binary. The shaded trailing 0s are the halvings, one per 0; they add up to 20.Full size

Known loops at each even length: 2, 6, 30, 52, 46 and 172 at lengths 12 to 22 (exact counts of every loop whose lap halves at most a stated number of times), and at least 4, 3, 2, 1, 10, 4 and 1 at lengths 24 to 36 (searches that stopped early).

Status and who checked it

  • Checked by computer: a loop of every even length from 12 to 36, each walked in exact integers.
  • Proved (by agents, checked line by line by a falsifier agent and a reviewer agent): every pattern of even length whose sum of powers of 2 equals 2S2^S, and that does not merely repeat a shorter pattern, gives a real loop with positive odd numbers; the rule has only finitely many loops of each length.
  • Open: a loop of every even length from 38 on (the conjecture), and so "infinitely many loops". Also open, and as hard in kind as Collatz: that no orbit grows forever.
  • Caution, from the falsifier agent: the counts do not rise steadily (52 at length 18, then 46 at 20), and a related rule (with −1-1 in place of +1+1) has a length with no loop after its counts start.
  • Who checked it: one falsifier agent and one adversarial reviewer agent, which reran the counts with a different algorithm. An earlier claim that the rule has exactly four loops was overturned by a reviewer agent solving these equations; this conjecture grew from that break. No human referee has read it.

Prior work

Searches run. OEIS by terms (the loop counts 2, 6, 30, 52, 46, 172; the numbers of the 12- and 14-step loops; the small loops) and by words; web searches for second-order and "with memory" Collatz rules, odd-part recurrences, cycle counting and trace criteria; arXiv abstracts. On 2026-11-18, 8 CORE searches followed (no Google Scholar search: the budget was used up). The rule and the claim were not found in the searches logged. On 2026-11-22, a Google Scholar sweep (4 queries) found no prior work stating this claim.

Nearest relatives, credited to their authors: OEIS A248797 (the odd part of the sum of the two previous terms, with an easy proof that numbers never grow); OEIS A048739 (the same recurrence without the odd part); second-order Collatz analogues of Clark and Lewis (1995) and of Amleh, Grove, Kent and Ladas (1998), as described by Zeilberger; Clark (2006), whose different rule has periodic orbits of arbitrarily large period; Simons (2022), closest in coefficients but a different rule; Kent (2020), a modified second-order Collatz equation, found by CORE.

Still owed: Guy's Unsolved Problems in Number Theory; Matthews' generalized 3x+1 pages; the full texts of Clark and Lewis and of Amleh, Grove, Kent and Ladas. No expert has been asked.

Further reading