Zomeable 4-polytopes

Zomeable orthographic projections of the 24-cell

Status: complete enumeration — exactly 3 essentially distinct projections.

Question

Which orthographic projections of the 4-dimensional 24-cell into 3D produce a shadow whose 96 edges all lie along default zometool strut directions (blue / yellow / red / green of the icosahedral symmetry system)?

We do not require:

We do require:

Result

Up to W(F₄)-symmetry of the 24-cell and O(3)-symmetry of the target, exactly three projection families exist:

# Common name 3D shape Color signature (× 8 edges per unsigned 4D-direction class) Hull
1 Cell-first (a.k.a. short-root projection) cuboctahedron 6 blue + 6 green 18 vertices, 96 edges, no collapse
2 Vertex-first (a.k.a. long-root projection) rhombic dodecahedron 3 blue + 8 yellow + 1 collapsed 15 vertices, 88 visible edges + 8 zero-length
3 Triality (3-fold symmetric) a 3-fold-symmetric solid with 4 symmetric “wings” 3 blue + 3 yellow + 6 red (red splits 3 short + 3 long) 24 vertices, 96 edges

The classical names “cell-first” and “vertex-first” come from which feature of the 24-cell aligns with the projection direction:

Files in this folder

File Family Notes
24cell_short_root_cuboctahedron.vZome #1 hand-built; 18 balls + 60 struts (24 B0 + 36 G0)
24cell_long_root_rhombic_dodecahedron.vZome #2 hand-built; 15 balls + 44 struts (12 B2 + 32 Y2), φ⁴ scale
24_cell_triality.vZome #3 uses vZome’s Polytope4d with quaternion="0 0 1 3/2 3/2 5/2 5/2 4" (= (0, 1, φ, φ²)); renders 24 balls + 96 struts (24 b3 + 24 r2 + 24 r3 + 24 y3)
Nan-24-cell-F4*.vZome various exploratory early F4 / quaternion experiments

Methodology

Setup

We work in two equivalent embeddings of the 24-cell in ℝ⁴:

These two are congruent as polytopes (related by a 4D rotation + scale by √2), but projecting one along a fixed direction n in its coordinate frame does not give the same 3D shape as projecting the other along n in its frame — because the rotation between them does not commute with the projection operator. We therefore search both embeddings.

Default zometool axes

We construct the four families of unsigned 3D directions by acting with the icosahedral rotation group A₅ (60 elements) on the prototypes:

B = (1, 0, 0)         → 15 unsigned axes  (3 + cyclic + …)
Y = (1, 1, 1)/√3      → 10 unsigned axes
R = (0, 1, φ)/√(1+φ²) →  6 unsigned axes
G = (1, 1, 0)/√2      → 30 unsigned axes ←(originally 15 in icosahedral; 30 here covers icosi-dodecahedral)

(Yielding 61 unsigned default axes total in our search; 40 distinct values of |cos∠| occur between pairs.)

Per-direction test

For a candidate 4D kernel direction n:

  1. Project: form the 3×4 matrix Q whose rows span n⊥; let P = Q · E, where E is the 4×96 matrix of edge vectors.
  2. Cosine prefilter: compute |cos∠| between every pair of nonzero projected edges. If any value is not in the precomputed list of 40 allowed cosines among default-axis pairs, reject.
  3. Two-edge alignment: for the first two non-collinear projected edges, try every pair of default unsigned axes (with signs) whose mutual |cos| matches; this fixes a candidate rotation R ∈ O(3).
  4. Verify: under R, every projected edge must be parallel (within tol) to some default axis, and we record its color. Edges that project to the zero vector are tagged collapsed (_).

Search space

The kernel direction n only matters up to scale, so we enumerate 4-tuples with components in the lattice ℤ[φ] = {a + bφ : a,b ∈ ℤ}, bounded by |a|, |b| ≤ 3. After deduplicating by sign-flips and permutations of the four axes, this gives 20,474 candidate directions, exhausted in seconds with the cosine prefilter.

Shape deduplication

A given color signature can be produced by many directions (entire W(F₄) orbits, plus the long-↔short-root duality). To avoid overcounting, we fingerprint each hit’s 3D vertex set by:

  1. centering at the centroid,
  2. computing all pairwise squared distances,
  3. normalizing so the maximum is 1,
  4. sorting the list.

This fingerprint is invariant under translation, rotation, reflection, and uniform scaling — it depends only on the 3D shape.

Outcome

Across both 24-cell embeddings and all 20,474 directions:

Why exactly three?

A heuristic explanation: the W(F₄) action on S³ has a small number of “highly symmetric” orbits — the directions stabilized by large subgroups of W(F₄). The three families correspond to the three classes of mirror planes / fixed points of W(F₄) on the 3-sphere of unit kernel directions:

Other candidate special directions (e.g. through edge midpoints, through face-centers of cells, etc.) do not produce all-default-color shadows: the cosine spectrum of their projected edges contains values absent from the icosahedral system.

Reproducing the result

cd C:\Users\nanma\Documents\vZome\24-cell\search
python exhaustive_search.py 3   # ~16 seconds
python group_shapes.py          # shape fingerprinting

Key scripts:

Caveats / open ends


Last updated 2026-05-01.

3D Viewers

➡️ Open this page on GitHub Pages to interact with the 3D models below (the embeds only render when this file is served via GitHub Pages, not in github.com’s markdown preview).

24cell_long_root_rhombic_dodecahedron.vZome
24cell_short_root_cuboctahedron.vZome
24cell_triality.vZome