4x4x4 upper bounds: 53 OBTM and 51 BT

I would like to report two new upper bounds for the 4x4x4 cube, both computer-assisted, with positions taken modulo whole-cube rotation:

  • Outer block turn metric (OBTM): every position can be solved in at most 53 moves. The previous published bound was 55 (Chen's comment on node 541), with 54 claimed this year.
  • Block turn metric (BT, inner and middle blocks allowed): every position can be solved in at most 51 moves. The previous bound was 53 (Tom Rokicki, node 541).

Both follow Chen's three-stage reduction. For OBTM the bound is 17 + 16 + 20 = 53.

  • Stages 1+2 in at most 17 moves. We use an exact phase-1 distance table on the 6,902,491,710 symmetry classes (diameter 12). Every class at distance up to 10 is finished in stage 2 by a common depth-7 family. For the 860,329,718 classes at depth 11 and the 17,470 classes at depth 12, we checked row certificates class by class: every pairing has a stage-2 finish within 17 total moves (11 + 6 or 12 + 5). The certificates are transported to every position of a class through a 16-element residual symmetry group.
  • Stage 3 in at most 16 moves. We ran our own exhaustive search over all 440,974,716,000 cells of a 16-fold symmetry-reduced encoding of the 7,041,323,520,000 stage-3 states. The per-depth counts match Chen's reduced distribution and Rokicki's unreduced distribution at all 17 depths, and the depth-16 layer has 119,610,148 cells (1,909,413,996 states).
  • The final 20 come from the 3x3x3 diameter theorem.

For BT the bound is 17 + 14 + 20 = 51.

  • The BT stage-3 group (the 20 stage-3 moves plus the nine inner and middle half turns) has diameter 15. All 16 depths match Rokicki's BT column, and exactly 56 states (5 cells) are at depth 15.
  • All 56 are strict maxima: every neighbour is at depth 14. As in Tom's 2015 post, this saves one move at the stage-2/stage-3 junction.
  • We prove the saving for every possible final syllable of the stage-1+2 word, on every axis, not only for a final quarter turn.

We also looked at whether leaving the stage-3 subgroup ("excursions") could bring OBTM stage 3 down to 15 for 52 moves. Sampling says rarely: about 0.3-1.7% of depth-16 cells. So 52 OBTM is not obtained this way.

Separately, for the n x n x n cube in OBTM we show that the counting lower bound is not tight up to an additive constant: G(n) - LB(n) >= f(n) with f(n) = Theta(n / log^2 n). Explicitly, G(n) - LB(n) >= 1 for n >= 101,465 and >= 11 for n >= 10^7. The same paper describes the shape of G(n): the normalised growth LB(n) ln(7.4n)/n^2 -> ln(24!/(4!)^6)/4 = 8.929, a parity sawtooth with two separate sources, a heuristic model in which G(n) - LB(n) is 2 or 3 for all n <= 100, and the unbounded growth of the gap as n -> infinity.

The work rests on three ideas of mine:

  1. Quotient reduction by symmetry. We compute on classes of positions under the spatial symmetries and whole-cube rotation instead of on raw positions. Examples are the 6,902,491,710 phase-1 classes and the 16-fold reduced stage-3 cells.
  2. Search and partitioning with no intersections between paths. The search space is split into disjoint parts (shards and layer-separated searches), so no state is expanded twice and each part can be checked on its own.
  3. Indexing a whole cube state by a single integer. Every coordinate is ranked, so a state or a class becomes one integer (for example the stage-3 cell index k·29,400 + c), and the tables become flat arrays addressed by that number.

I developed these ideas into concrete algorithms and proofs together with Claude Opus 5.5 (Anthropic) and ChatGPT Astra (OpenAI). The computations were run on external computing resources (Google Colab GPU hosts and Kaggle notebooks), and the conclusions above were drawn from them.

Manuscripts, code, certificates and logs, with SHA-256 checksums for every file:

No independent party has rerun the computations yet, so I would very much welcome checks, questions or criticism. The phase-1 distance table is 3.45 GB and is not in the archives. It comes from the public r54 release, and its hash is in the paper.

I hope this work will be useful to others working on God's number problems and on cube group computations.

JiwonByun