The last-digit Collatz family across all bases

Edition 3 · On this site since · last revised · Verify & co-author
Kind
Theorem with exact computations
Status
Proved (mixed): Lean-verified for base 2, base 10 descent, base 17 runaway, and even-base drift; written proof for bases 4 and 6 descent; checked by certificates for finite screens; unrefereed
Prior work
Companion to the base-10 Ouroboros paper; builds on Terras (1976), Lagarias (1985), Matthews and Watts (1984, 1985), Krasikov and Lagarias (2003), and Simons and de Weger (2005)
On this page
  1. The statement
  2. Check it yourself
  3. Status and who checked it
  4. Prior work
  5. Further reading

The statement

For any integer base b≥2b \ge 2, write db(n)=n mod bd_b(n) = n \bmod b for the last base-bb digit of nn, and define the last-digit Collatz map fbf_b on positive integers by

fb(n)={n/2,n even,db(n) n+1,n odd.f_b(n) = \begin{cases} n/2, & n \text{ even}, \\ d_b(n)\,n + 1, & n \text{ odd}. \end{cases}

In base 44 this map is conjugate on odd integers to the classical 3x+13x+1 Collatz map; base 1010 is the Ouroboros map. This companion paper develops the theory of the family {fb}b≥2\{f_b\}_{b \ge 2} across all bases and sharpens the base-1010 bounds across four parts:

  1. All even bases. The 22-adic odd-step Markov chain PbP_b admits a closed recurrent class of non-positive logarithmic drift if and only if b∈{2,4,6,10}b \in \{2,\allowbreak 4,\allowbreak 6,\allowbreak 10\}. The near-zero band [−14,14][-\frac{1}{4},\allowbreak \frac{1}{4}] contains precisely {4,8,10,14}\{4,\allowbreak 8,\allowbreak 10,\allowbreak 14\}, and [−14,0)[-\frac{1}{4},\allowbreak 0) contains only {4,10}\{4,\allowbreak 10\}. The least drift satisfies Δ(b)≥12ln⁡(b+2)−72ln⁡2\Delta(b) \ge \frac{1}{2}\ln(b+2) - \frac{7}{2}\ln 2, and if pk(b)p_k(b) is the natural density of starts still at or above their initial value after every shortcut step through kk, then p∞(b)=lim⁡k→∞pk(b)p_\infty(b)=\lim_{k\to\infty}p_k(b) satisfies p∞(b)=0p_\infty(b) = 0 if and only if b∈{2,4,6,10}b \in \{2,\allowbreak 4,\allowbreak 6,\allowbreak 10\}. A positive limit here is a statement about finite horizons. Individual orbits in these even bases require a separate argument.

  2. Odd bases and zero-digit trapdoors. For odd b≥3b \ge 3, odd multiples of bb carry last digit 00 and drop immediately to 0⋅n+1=10 \cdot n + 1 = 1. Entry into this trapdoor is classified by x2+1≡0(modb)x^2 + 1 \equiv 0 \pmod b: rootless bases (divisible by a prime p≡3(mod4)p \equiv 3 \pmod 4) have trapping density 1/b1/b, while rooted bases have 2ω(b)2^{\omega(b)} entry roots. Exact rational trapping and runaway densities (τ(b),ρ(b))(\tau(b),\allowbreak \rho(b)) are established for b=3b = 3 through 1919.

  3. How many decimal starts reach 1? If N1(X)N_1(X) counts odd starts up to XX that reach 1, then N1(X)≥0.0145X0.49875N_1(X)\ge0.0145X^{0.49875}. Some constant CC gives N1(X)≤CX0.6245N_1(X)\le CX^{0.6245}. Whether a single growth exponent exists remains open.
  4. Short decimal cycles. An exact computation through eight odd members finds only (1)(1) and (5,13)(5,\allowbreak 13) in the odd-to-odd map. A separate Lean proof covers up to three. Combined with the recorded sweep through 101510^{15}, the large-cycle bounds require at least 34,905 halvings, 15,496 odd steps and 53,185 total steps for any other cycle.

Check it yourself

Even-base Markov drifts (b=2b = 2 through 1818).

Even base bb 2 4 6 8 10 12 14 16 18
Least drift Δ(b)\Delta(b) −1.386294-1.386294 −0.143841-0.143841 −0.849815-0.849815 +0.111572+0.111572 −0.134161-0.134161 +0.279808+0.279808 +0.067947+0.067947 +0.405465+0.405465 +0.376180+0.376180

Only b=2,4,6,10b = 2,\allowbreak 4,\allowbreak 6,\allowbreak 10 have Δ(b)≤0\Delta(b) \le 0, and only b=4b = 4 (Collatz) and b=10b = 10 (Ouroboros) lie in [−14,0)[-\frac{1}{4},\allowbreak 0).

Odd-base trapdoors and runaways by hand.

  • Base 5 (certain trapping, τ=1\tau = 1): 55 ends in base-55 digit 00, so f5(5)=0⋅5+1=1f_5(5) = 0 \cdot 5 + 1 = 1. Every odd residue class enters the trapdoor with density 11.
  • Base 13 (rooted split): τ(13)=13,210/16,991\tau(13) = 13,\allowbreak 210/16,\allowbreak 991 traps while ρ(13)=3,781/16,991\rho(13) = 3,\allowbreak 781/16,\allowbreak 991 escapes.
  • Base 17 (rooted split with runaway orbit): τ(17)=148,789/417,911\tau(17) = 148,\allowbreak 789/417,\allowbreak 911 and ρ(17)=269,122/417,911\rho(17) = 269,\allowbreak 122/417,\allowbreak 911. Starting at 3333 (33≡16(mod17)33 \equiv 16 \pmod{17}) gives 33→529→1,059→5,296→2,648→1,324→662→331→2,649→37,087→370,87133 \to 529 \to 1,\allowbreak 059 \to 5,\allowbreak 296 \to 2,\allowbreak 648 \to 1,\allowbreak 324 \to 662 \to 331 \to 2,\allowbreak 649 \to 37,\allowbreak 087 \to 370,\allowbreak 871. Every positive nn in this residue class satisfies f1710(n)=11200n+1271f_{17}^{10}(n)=11200n+1271, which returns to the same class. Iterating gives divergence.

Replay the checks. Run python3 directly on the downloaded verification archive (python3 last-digit-family-*-verification.zip). Its README also gives commands for the separate C++ period enumeration and the Lean build.

Figure 1. Left: least closed-class drift Δ(b)\Delta(b) for even bases b=2b = 2 to 4040, showing b=4b = 4 (Collatz) and b=10b = 10 (Ouroboros) as the unique even bases in the near-critical negative band [−14,0)[-\frac{1}{4},\allowbreak 0). Right: exact rational trapping, runaway, and wandering densities for odd bases b=3b = 3 to 1919.Full size
Figure 2. Left: bounds on the growth of the odd basin of 1, tightened from [0.464958,0.626383][0.464958,\allowbreak 0.626383] to [0.49875,0.6245][0.49875,\allowbreak 0.6245]. Right: all 19 surviving 4D-LLL cycle candidates up to K≤64,000K \le 64,\allowbreak 000 under the 101510^{15} sweep, binding at K≥34,905K \ge 34,\allowbreak 905, m≥15,496m \ge 15,\allowbreak 496, and K+m≥53,185K + m \ge 53,\allowbreak 185.Full size

Status and who checked it

  • Formal proofs. The archived Lean library records 424 declarations: Lean-verified for base 2 convergence, base 10 density-one descent, base 17 divergence modulo 34, the even-base stationary drift classification, parity coding, decimal backward moves, and decimal cycles through three odd members. The finite-state density-one descent step from negative drift in bases 4 and 6 rests on the manuscript's written proof.
  • Exact computations. The checker verifies the even-base drift screen through 208, odd-base certificates through 1,999, three decimal inverse-tree weight vectors and the short-period enumeration through eight odd members.
  • Open. Global boundedness in bases 4 and 10 and the exact growth of the decimal basin of 1 remain open.

  • Unrefereed. Checked by Lean 4, exact-arithmetic scripts, and agent reviewers; no human referee has read the manuscript.

Prior work

Searches run. Sweeps of arXiv, zbMATH Open, OpenAlex, and OEIS through 2026-11-21 found no prior work defining {fb}b≥2\{f_b\}_{b \ge 2}, proving the all-even-base drift classification, or establishing the odd-base zero-digit trapdoor criterion. Twenty-nine of the manuscript's 3939 references were read in full.

Context and credit. Parity coding and stopping-time density build on Terras (1976, 1979) and Lagarias (1985); Markov-chain methods for generalized Syracuse maps follow Matthews and Watts (1984, 1985), Leigh (1986), and Matthews (2010); predecessor-tree lower bounds build on Applegate and Lagarias (1995) and Krasikov and Lagarias (2003), matching OEIS A003269 on digit-11 chains; cycle lower bounds adapt Simons and de Weger (2005, 2010), Simons (2008), Hercher (2023), and Barina (2025).

Still owed. Ten paywalled or print-only references marked unchecked in the manuscript (Belaga and Mignotte 1998, Eliahou 1993, Everett 1977, Guy 2004, Krasikov 1989, Lagarias and Weiss 1992, Steiner 1978, Venturini 1992 and 1997, and Wirsching 1998).

Further reading