Zomeable 4-polytopes

Methodology limitations and open questions

This file collects open questions about the search methodology that should be revisited if any result becomes interesting enough to need formal proof.

Two-criteria distinction (zometool-aligned vs vZome-embeddable)

The 8-cell case revealed that the search engine’s notion of “zomeable” is strictly weaker than vZome-embeddability:

The vZome criterion is equivalent to: every length ratio |c_i|/|c_j| between projected edges lies in ℚ(φ). For most polytopes this is automatic. For the 8-cell, two of the 8 search hits at rng=1 fail it (magnitudes involve √2/√3 or √(2+φ)).

When emitting .vZome files we always run classify_direction on every strut as a final check: a None return means the model is not vZome-embeddable and we skip emission.

Saturation as evidence of completeness

Method used: for each polytope, run the kernel-direction search at range rng = max |a|, |b| over n components in ℤ[φ] = {a + bφ}, with rng = 2, 3, 4. If the count of distinct 3D shapes is the same at every range, we declare the search “saturated” and report that count.

Why this is probably correct (informal argument):

For each valid projection, the kernel direction n lies on a specific 1D ray in ℝ⁴ (modulo the icosahedral group acting on the 3D image). ℤ[φ]⁴ ∩ ray is a 1D lattice; all lattice points give the same 3D shape (just rescaled). As rng grows we accumulate more lattice points on the same finite collection of rays — hits grow linearly, shape count stays constant.

What could break this argument: a “missed” valid ray whose smallest ℤ[φ]⁴ lattice point lies at rng > 4. This requires the ray’s direction to have large numerator/denominator coefficients in (a + bφ) form, which would be unusual for a geometrically natural projection — but is mathematically possible.

Status: empirical, not formally proven.

Regular compound and Schläfli-Hess inventory checkpoint

The regular compound/star-polytope category uses the same bounded-saturation standard, but with a broader source inventory and a post-sweep certification step:

Thus the category has bounded numerical saturation through these completed bounds. This is stronger than a single gallery regression check, but still not a formal proof excluding all larger-coefficient kernel directions.

Open questions and future work (parked 2026-05-02)

These are documented for picking up later; none are blocking.

  1. Run rng=5 (or rng=6) saturation for the finite-classification cases. Currently saturated at rng ≤ 4 for 5/16/24/120/600-cell. rng=5 takes ~1 min; rng=6 ~4 min. If counts hold, evidence is near-formal. (python lib/run_search.py <name> 5.)

  2. Augment shape fingerprint with strut-color signature multiset. group_by_shape() currently uses sorted pairwise squared 3D distances. Two shapes with the same distance multiset but different strut colorings would be wrongly merged. Spot checks on current results suggest no collisions (the strut signatures already differ), but a formal audit would close the gap.

  3. Algebraic enumeration as a kernel-rng-free alternative. For each polytope, enumerate all valid color-assignments to edges (finite combinatorial set), solve for n ∈ ℝ⁴ in each case, classify resulting shapes. Removes the rng cutoff entirely. Combinatorially heavy but conceptually cleaner.

  4. Document the 2 non-vZome-embeddable support-3 8-cell shapes. The 8-cell search at rng=1 finds two zometool-strut shapes with sigs B:24+Y:8 and B:8+Y:8+R:16, requiring length ratios √2/√3 and √(2+φ) respectively. They could be built physically with default zome parts at incommensurate scaling but cannot be emitted to vZome. Worth a brief writeup with explicit ball coordinates in a (ℚ(φ), √2, √3)-extended field.

  5. vZome-embedding verification on the 12 non-hand-built .vZomes. The polytope4d-emitted files (5cell, 16cell, 120cell, 600cell — 12 files total) rely on vZome’s quaternion-based projection. Each should be opened in vZome and verified to render with default-color struts. Status of which have been confirmed by user vs. only classify_direction-checked at emit time should be tracked per-file. Spot-checked 5-cell files all confirmed as true orthographic via M·Mᵀ test (2026-05-02).

Strengthening options

If we ever need to upgrade saturation to proof:

  1. Push rng higher. rng=5 (≈250K dirs, ~1 min) and rng=6 (≈700K dirs, ~4 min) are cheap. rng=8 gets expensive (~10⁷ dirs).
  2. Algebraic enumeration. For each polytope, enumerate all valid color-assignments to the edges, solve for n ∈ ℝ⁴ in each case, then classify the resulting shapes. Finite, exhaustive, no rng cutoff — but combinatorially heavy.
  3. Random irrational sampling. Sample n from a continuous distribution outside ℤ[φ]⁴ and run try_align. If no new shapes appear over millions of samples, the discrete set really is what ℤ[φ]⁴ captures.

Verification: all emitted models are true orthographic projections

Each .vZome file we emitted can be verified to be a uniform-scale orthographic projection by the M·Mᵀ test:

  1. Parse the ball coordinates from <ShowPoint> elements (ZZ[φ]³).
  2. Find the vertex assignment (V_i ↔ ball_j, possibly merging V_i pairs for collapsed cases) that minimizes least-squares fit residual of a 3×4 linear map M with M·V₄ᵀ ≈ pts3ᵀ.
  3. Compute M·Mᵀ. For a true orthographic projection (with rotation + uniform scale), this must equal s²·I₃ (eigenvalues all equal).

Audit results (2026-05-02):

Remaining files (24-cell hand-built, 120/600-cell polytope4d-based) have not been audited yet via this test but should be cheap to add.

Per-polytope confidence

Polytope Result Saturation evidence Confidence
5-cell 4 shapes rng 2/3 both 4 high
8-cell 1 inf family + 3 sporadic rng 1/2/3 → 8/32/129 with per-support breakdown — only support-2 grows; sporadics are cube + rhombic dodec + phi-oblique high (theoretical + search)
16-cell 6 shapes rng 2/3/4 all 6 high
24-cell 3 shapes rng 2/3 both 3 high
120-cell 1 shape rng 2/3 both 1 high
600-cell 1 shape rng 2/3/4 all 1 high

8-cell special case (resolved)

The 8-cell is the only polytope where edge directions align with the 4D coordinate basis (e_1, e_2, e_3, e_4). Consequence: the projection formula c_i = e_i − n_i n makes c_i = e_i⁺ (a unit vector in n^⊥) whenever n_i = 0. So any 4D kernel with support 2 — i.e. n = (a, b, 0, 0) modulo permutation — automatically produces a zometool-aligned projection (3 of the 4 image vectors are unit B-axes, the 4th lies along the in-plane direction of (a, b)).

This gives the infinite family of split cuboids parametrized by the integer ratio (a:b), with vZome-embeddability further restricted to a² + b² ∈ ℤ[φ]² — i.e. either Pythagorean triples (giving integer c) or a² + b² = 5m² (giving c = m(2φ−1) = m√5). See 8cell/CLASSIFICATION.md.

The other 4-polytopes don’t have edges aligned with any 4D coordinate basis, so no analog of this argument applies — their shape sets are finite.

5-cell special case (resolved)

The 5-cell’s natural ring is ℤ[√5], but √5 = 2φ - 1 ∈ ℤ[φ], so it embeds in ℤ[φ]⁴ with an explicit aligned embedding (see 5cell/RESULTS.md). The default gen_dirs(permute_dedup=True) was incorrect for the 5-cell because the embedding singles out the 4th axis; using permute_dedup=False correctly enumerates all directions and finds 4 distinct shapes including the natural vertex-first “tetrahedron + center” projection.