Research status · archive update · Yamba AI Math

Latest status of the Rogers–Ramanujan KR1 bijection construction Phase 2: filling the gap of residue subfamilies

We are looking for an explicit bijection for the KR1 partition identity (mod 9): a reversible, weight-preserving map between two families of partitions. This page reports where that construction stands. More of the map is built than before, and we now know precisely why the remaining part is hard. The full bijection is still open.

Full KR1 bijection: open Novelty: not yet checked Earlier check: 3974 / 5461 through weight 40 new all-a family G6, independently reviewed
Visual discussions · one page per iteration

Regroup, imagine, test, return

Explore the diagrams with us: Christopher's intuitions become concrete handoffs to separate Claude Code or Codex worktrees. We bring their results and failures back into the next visual discussion.

Latest checkpoint: the support obstruction is proved with an actual-column Lean certificate. The conditional accepted-preservation theorem and terminal-gap kernel have separate certificates; the positive auxiliary models and uniform C↔A endpoint have reviewed ordinary arguments. The full structural C↔U bridge remains open. The article records exact assurance and missing transfers. Earlier sector counts below retain their historical scope.

The problem

Four encodings and one gap

The plan is a chain of four reversible maps. The two outer links use established constructions: Andrews–Keith and Stockhofe between B and U, and Kurşungöz between C and A. The middle link, U to C, remains incomplete in our construction.

Bpartitions with parts ≡ 1, 3, 6, 8 (mod 9)
Uweighted columns F(h), height drops ≤ 4
Csectors (a, b, μ, ν), weight Q(a,b)+Σμ+3Σν
Agap-condition partitions (distance-two gaps ≥ 3)

Here Q(a,b) = a² + 3ab + 3b² and F(h) = 9⌊h/5⌋ + (0,1,3,6,8)[h mod 5]. The two sides are known to have equal counts (Mizuno, Theorem 1.1), but a count is not a bijection. The goal is an explicit object-by-object map with an explicit inverse.

Achieved · finite check through weight 40

How much of the U ↔ C bridge is built

H3, the reviewed baseline map fiber-splitting extensions: the y = 2, 3, 4 slabs of sector (4,0) no map yet

These counts are finite checks, not percentages of the infinite problem. Several pieces are proved for every weight: all of a ≤ 2, the sector a = 3, b = 0, the high-b tails, and the new all-a family G6.

Where the 1487 open states sit

These earlier sector counts add up to 1485; the notebook doesn't break down the other 2. The first open pair at that checkpoint was C=(3,1,(0,0,0),(0)) ↔ U=(9,9,3) at weight 21. The corrected boundary plane now covers it; see iteration 02.

Achieved · proved for all weights

Sector map: covered tails and an open core

Each column is one value of a. Height is b on a log scale. Green regions have a proved map. The amber core near the origin has none: the funded constructions only start once b is large.

proved, with an independent review partly covered (some slabs of b = 0) open ● has open states at weight ≤ 40

G6 thresholds: T(6)=119, T(7)=155, … T(15)=507, rising to T(100)=8197. G6's smallest weight is 44661, far above the open core.

Achieved · the extra-dimension idea, traced literally

What the extra axis remembers

The guiding intuition is Christopher Yamba's: when two inputs collide on one output, move the obstructed box into an extra direction so the collision splits. In the smallest example, two different tableaux collapse to the same output under the planar slide. Lifting the corner box into a third direction stores where the slide cut, and rotating it back into the plane gives two distinct outputs.

1 · Two inputs
1
2
3
4
5
T1
1
2
3
5
4
T2
2 · The planar slide collides
1
2
3
4
5
both T1 and T2
two inputs, one output: the cut position is lost
3 · Lift + rotate
1
2
3
4
5
T1 → slides flat
1
2
5
3
4
T2 → 5 lifted to z=1, rotated in
distinct outputs; each input can be recovered from its output alone

What does the extra direction remember? In this restricted ordinary-tableau family, the corner label and cut position can be recovered; one extra axis suffices, including for a traced ten-input fiber. Christopher's broader proposal allows a fresh dimension for every numbered-square assignment. Its general construction and original-target decoder remain open; explore the 2D, 3D, 4D and 7D illustrations in iteration 03.

In KR1, G6 uses a related multiplicity-coordinate encoding: one free axis per μ-gap quotient and one binary axis per residue bit (18 axes at a = 6, 67 at a = 15). It covers every a ≥ 6 once b meets its safe funding threshold T(a). This is a restricted encoding, not the complete per-square placement procedure.

How we got here

Timeline of the work so far

  1. 23 Oct 2026
    Problem set upSelected a structural KR1 bijection as the target. Fixed the B ↔ U ↔ C ↔ A plan, with the outer links taken from established constructions and the U ↔ C bridge as the new contribution.
  2. 24 Oct
    RR002: first construction roundsTwo author–critic rounds. They produced checked infrastructure and exact obstructions, but no bridge with an inverse.
  3. 24 Oct
    RR003: the bridge for a ≤ 2A U ↔ C bijection for all sectors with a ≤ 2, executable endpoint algorithms, and finite checks through weight 30.
  4. 24 Oct
    RR004: a reviewed obstructionProved that no fixed signed height-transition statistic can produce b. Fourteen objects of weight at most 10 already rule it out.
  5. 24–25 Oct
    Higher-dimensional Ferrers diagrams3D and d-dimensional Ferrers bijections with an article and Lean 4 proofs (PR #13, merged). A follow-up with the N = 40 octant recursion is in PR #15, which has two open review comments on the article.
  6. 25 Oct
    Phase 2: the reviewed H3 mapAll of a ≤ 2, the sector a = 3, b = 0, and the high-b tails for a = 3, 4, 5. That gives 3498 of 5461 states through weight 40.
  7. 25 Oct
    Fiber splittingChristopher's intuition recorded in his own words, then applied: the y = 2, 3, 4 slabs of sector (4,0) are added for all x, t, z (+476 states).
  8. 25 Oct
    Independent cube run (Claude)The literal cube trace, the no-single-linear-rule lemma K-1, the all-a family G6, the ping-pong lemma K-4, and eight recorded failures, F-1 to F-8.
Where it got hard

Separating the inputs is easy. Landing in the original targets is the hard part.

Every lift we tried separates its collision. The hard part is putting the stored information inside a valid, ordinary target that no other sector already uses. Three obstructions explain why the open core resists:

K-1 · no single linear rule

The reviewed obstruction excludes the specified single global affine-rule ansatz. Residue branches are one proposed way around it; it does not exclude all higher-dimensional procedures.

restricted obstruction; independently reviewed
Funding

Each marker axis costs weight, which something must pay: the doubled ν-contour or a large μ-part. Near the apex of a sector nothing can pay. That is why the H3 tails and G6 only start at large b.

G6 starts at T(6) = 119
Collisions with H3

The open states have to land on outputs H3 doesn't already use. The natural rules for (3,1) all land on outputs from five other sectors.

sectors (2,1), (4,0), (2,2), (1,3), (0,3)

Candidates that failed (frozen before checking, kept on record)

F-1 · lift isn't a Ferrers diagram

Placing 5 at (2,0,1) leaves it unsupported along axis 0. The lift is valid only under a weaker “hanging” convention.

T2=((1,2),(3,5),(4))
F-2 · origin lift leaves target readout open

The reviewed corner lift to (0,0,1) retains the moved label and has an exact inverse on its image. It does not alone give the required general original-target decoder.

lossless restricted lift; general bridge unresolved
F-3 · fixed rotation fails at n=5

4 of the 20 colliding sources produce an invalid tableau: 4→5 is not a P-chain.

h=(1,2,4,5,5), ((1,4),(2,5),(3))
F-4 · printed Φ₂₂₁ changes inversions

The printed rotation formula changes inv_g on 2 inputs. The code avoids this by filtering, so the printed formula and the code differ.

h=(2,2,4,5,5): inv_g 1→2, 2→1
F-5 · first-eligible inverse is incomplete

12 singleton fibers at n=5 aren't recovered. Fixed by taking the least tail entry above v.

((1,4),(2),(3),(5))
F-6 · Y3 rule collides on its whole slab

Every checked point with x = 0 lands on an output H3 already uses.

(4,0,(0,t,t+3,t+3+z)), t,z ≤ 4
F-7 · naive (3,1) rule

Free 3s on the (9,9) core collide with sector (2,2) for every ν ≥ 3 (45 of 90 checked).

(9,9,3,3,3,3), weight 30
F-8 · one axis assignment for (3,1)

A single assignment for the whole sector is impossible: residue classes (0,0), (0,1) and (2,0) have no free head.

forced apex (9,9,3) vs retained (9,9,3⁴)
Approaches in progress

What we're trying, and how

1 · Extend sector by sector, keeping what's reviewed

How: a ping-pong (finite Schröder–Bernstein) walk. Design one explicit injection on the retained sources plus the new sector, then walk between it and the reviewed map until the walk leaves the old image. Every reviewed assignment stays as it is, and the new sector lands in the open part.

2 · Branch by residue class at the core

How: for sector (3,1), split by residues of the gaps and give each class its own head on a distinct marker core. Candidate families that avoid every used output are already recorded through weight 46, for example (15,8) + {6,9,3,1}.

3 · Fund markers from μ instead of ν

How: a large smallest μ-part can pay for marker axes when b is small. That would lower the funding thresholds toward the core, using the same axis vocabulary as G6.

4 · A signed identity plus the involution principle

How: a combinatorial proof of a signed KR1 identity would give an explicit bijection through the Garsia–Milne involution principle. The first step is checking whether any published proof of the mod-9 identities is already combinatorial.

What remains

Open work, in order

  1. Construct the uniform auxiliary branch mapsnext
    Begin at level 2→3 on the positive cut normal forms. Prove intrinsic branch recognition and literal inverses for the C-state split, then the second exchange, uniformly across every normalization threshold. See the current regroup and full article.
  2. Cover the rest of the open core
    Sectors (4,0) for y = 5…17 at small x, (5,0), (4,1), (6,0), (3,2), and then every sector with a ≥ 3 below its funding threshold. A trial rule for y = 5 has no counterexample through weight 46, but it isn't accepted yet.
  3. Fix the Phase 2 article before merging
    Fix or reject the printed rotation branch (F-4) and the determinant inverse that returns τ⁻¹. Then state exactly what the Lean proofs certify.
  4. Check novelty
    Compare against Mizuno's counting proof and the closest published explicit constructions.
Source

Code, proofs and research records

The public Yamba AI Math repository contains definitions, checks and earlier research records. The regrouping archive also summarizes locally recorded follow-up work, with checkpoint identifiers and explicit assurance boundaries. It does not publish the private journal.