最小可能的 nopert 多面体:Septopert
Smallest possible nopert polyhedra: the Septopert

原始链接: https://github.com/tom7/noperts/blob/main/septopert.md

Septopert 是一个具有 7 个顶点的凸多面体,它不是 Rupert 多面体(“nopert”);7 个顶点是最小可能值。其标准版本具有整数坐标、C3v 对称性、精确的立方体角顶点,以及精确的“Triforce”投影。一个 Lean 证明使用已检查的证书数据(镜像、弧、轴和 P 盖,以及视图覆盖表)验证其为 nopert。它还作为主定理中的 7 顶点见证。一个单独的极小性结果表明,在将整数族成员约化为 55,651 个标准元组和证书后,高度小于 7 的成员都不是 nopert。该证明仅依赖标准公理,大型证书数据被原生检查;从头复现大约需要 31 分钟。证书数据为机器生成,拟采用 CC0 协议。

对不起。
相关文章

原文

Tom 7, 7 Nov 2026

The Septopert is a convex polyhedron with 7 vertices that is not Rupert (a "nopert"). 7 vertices is the fewest possible for a nopert. The canonical Septopert, described on this page, is the best version of a family of similar noperts, including the holotype Septopert Classic, the SeptopertM18, and SeptopertM12. Its integer coordinates are the smallest possible for a nopert in this family (see "Least height" below). This page gives the exact statement, what Lean checked and how, and how to reproduce the result.

Lean entry point: proof/Septopert.lean; the headline theorem is in proof/Septopert/Integer.lean.

The Septopert, spinning around with confidence

Here is an exact STL file. On github, you can rotate it around in your browser by just clicking on that.

The Septopert is the convex hull of these points (its vertices):

(-7,  7,  7)
( 7, -7,  7)
( 7,  7, -7)
( 3, -4, -4)
(-4,  3, -4)
(-4, -4,  3)
( 7,  7,  7)

It is a very 7 solid: 7 vertices, coordinates ±7 and 3, 4 (3 + 4 = 7), apex edges of length 14, base edges 14√2 and inner triangle edges 7√2. Despite being discovered by Tom 7, the sevenness is really just a mathematical coincidence! Jackpot!

The Lean statement is about exactly these integer points (largest |v| = √147 ≈ 12.12). Any scale or rigid transformation of the solid is a Nopert too. (Internally, the proof actually reuses some older machinery and works in a different frame and scale, and then uses lemmas to prove the statement for the vertices above.)

  • Symmetry C3v, as for the other members: a 3-fold axis along (1, 1, 1) and three mirror planes. Vertices 0–2 are the permutations of (−7, 7, 7), vertices 3–5 those of (3, −4, −4), and vertex 6 is 7·(1, 1, 1).
  • 10 triangular faces, the same 10 as the rest of the family (vertex lists 015, 016, 024, 026, 045, 123, 126, 135, 234, 345); vertex degrees 5,5,5 (vertices 0–2), 4,4,4 (3–5) and 3 (vertex 6). Combinatorially a gyroelongated triangular pyramid.
  • Two special features (both exact), which is what makes the proof harder:
    • Cube corner. The apex, vertex 6, is a cube corner, as for SeptopertM18: the edges 6–0, 6–1, 6–2 are mutually orthogonal, so the three apex dihedral angles are exactly 90°.
  • Exact Triforce. In the orthographic projection along the 3-fold axis (onto the plane through the origin perpendicular to (1, 1, 1)), the image of vertex 3 is exactly −1/2 times the image of vertex 0 (and so on), so the inner triangle 3–4–5 projects onto the medial triangle of the base 0–1–2. Equivalently, vertex 4 projects onto the midpoint of 0–2, so the plane of face 0–2–4 is parallel to the axis, and face 0–2–4 is perpendicular to face 3–4–5 (whose normal is the axis). This is a second right angle, and it gives a second family of views at which the first order vanishes.

Exact arithmetic (integer face normals; lengths in the integer coordinates above).

Line drawing of two views of the Septopert, with colored edges

(Note: This is a schematic drawing used for any member of the Septopert family, not exact.)

color description edges dihedral angle edge length
blue apex of pyramid 6–0, 6–1, 6–2 90° (exactly) 14
yellow base of pyramid 0–1, 0–2, 1–2 144.735610° 14√2 ≈ 19.798990
green triforce exterior 0–4, 0–5, 1–3, 1–5, 2–3, 2–4 108.316731° √146 ≈ 12.083046
red triforce center 3–4, 3–5, 4–5 141.057559° 7√2 ≈ 9.899495

In proof/Septopert/Integer.lean:

theorem isNopert_intVerts (h : CertificatesValid) : IsNopert intVerts

intVerts is the set of the seven integer points above (septopert is another name for it).

IsNopert, IsRupert and the trust model are as for SeptopertM18. CertificatesValid says that the certificate data (four solution tables, the mirror cap certificate, and three special caps: the arc cap, the axis cap and the P cap) passes Lean's validity checks. The same IsNopert is used by the minimality result, and the Septopert is the 7-vertex witness in proof/ProofOfMainTheorem.lean.

The certificates are built in a congruent copy of the solid (verts, coordinates over 64: (−27,15,15), (27,6,6), (−13,−13,−13) and permutations). A short bridge lemma maps it to the integer coordinates with an exact similarity (scale 3/64, a reflection through the axis composed with the 180° rotation about it, and a translation), and similarities preserve Rupertness.

What is proved and what is verified

As for SeptopertM18: isNopert_intVerts is checked by Lean's kernel and depends only on the standard axioms (propext, Classical.choice, Quot.sound). The data (about 732,000 table rows and the cap certificates) is checked by the program constructSeptopert, which runs the same Lean validity checks as compiled code and only then constructs IsNopert intVerts, printing

instantiated: IsNopert Septopert.intVerts (CapOk, ArcOk, AxisOk, PCapOk checked natively; tube and chart tables checked)

This page describes the canonical Septopert, which has the smallest integer coordinates of the family. It sits on the edge (we think) of the family, in a special position with an exact cube corner and an exact Triforce.

We proved several other members of the family nopert on the way:

  • Septopert M12 is in the same double special position
  • Septopert M18 has the cube corner, but the triforce misses by about 3.7%.
  • Septopert Classic is in general position, although it is close (90.333818° dihedral angle, 3.5% off triforce).

The technique is the one in How the Nopert proofs work, with the arc cap of SeptopertM18. The view cover has 13 view codes and 16 view-code triangles (SeptopertM18: 15 and 22). The tube and the 5D search close everything except four small regions of views, each of which needs a blow-up certificate because the first-order margin vanishes there:

  1. The mirror cap at the mirror normal, as for every member of the family.
  2. The cube-corner arc, as for SeptopertM18: an arc of views ending at U, the view straight down an apex edge, where the face through that edge is edge-on and the right angle at the apex makes every first-order margin zero. It is certified by an arc cap (wedge charts along the arc plus a point cap at U).
  3. The axis view, the view along the 3-fold axis. Here the Triforce puts the inner triangle's shadow exactly on the base's medial triangle, and the first order again vanishes, at an isolated view. It is certified by a point cap around the axis (rotations within 1/46 in Cayley coordinates; tilt μ·(s₁, s₂), spin μ²·s₃).
  4. The face-024 strip. Every face is edge-on along a great circle of views. For face 0–2–4, which is perpendicular to face 3–4–5, the right angle makes the first order vanish along a stretch of that circle starting at a view P (as the apex's right angle does along the cube-corner arc), here with a 2-dimensional cone of first-order flexes. It is certified by the P cap: a point cap at the end view P together with a wedge of polar charts along the strip.

In each case the decision moves to second order, where the clearance is strictly negative, which is why a certificate exists (floating-point and symbolic study, not part of the proof).

Write a member of the family by its five coordinate values x = (x0, ..., x4): the points (x0, x1, x1), (x2, x3, x3) and their permutations, and (x4, x4, x4). Its height is half the range of those values, rounded up. The Septopert's values are −7, 7, 3, −4, 7 (range 14), so its height is 7. No integer member of lower height is a nopert. In proof/SeptopertFamily.lean:

theorem not_isNopert_of_familyHeight_lt_seven (hvalid : LeastHeightValid) :
    ∀ x : Fin 5 → ℤ, familyHeight x < 7 → ¬ IsNopert (familyVerts x)

The statement covers every integer tuple, with any signs. Translating, scaling and reflecting the tuple preserves being a nopert (proved in Lean), so it reduces to 55,651 canonical tuples with values in [0, 12], each with a small certificate checked by checkSeptopertFamily (about 4 seconds):

  • 161 are coplanar (a plane through all 7 points);
  • 25,932 have a point in the convex hull of the others (explicit convex weights), so at most 6 points span the hull and the minimality theorem applies;
  • 29,558 have an exact Rupert passage: rotations from integer quaternions q (as R(q)/|q|²), a rational offset, and for each inner vertex an outer triangle that strictly contains its shadow.

The table is proof/SeptopertFamily/data/leastheight.txt (2.7 MB). Like the other results, this is a kernel theorem (axioms propext, Classical.choice, Quot.sound) plus a native check of the data.

Verified on 2026-11-07: constructSeptopert (then constructM7) checked the mirror cap (52 charts), the arc cap (146 charts, 22,282 leaves), the axis cap (52 charts, 3,384 leaves), the P cap (78 charts, 43,869 leaves), the tube tables, and the four chart tables (406,726, 124,781, 108,449 and 92,514 rows), and printed the instantiated line above. On 12 cores of a busy machine it took about 23 minutes (stage times: mirror cap 196 s, arc cap 723 s, axis cap 16 s, P cap 300 s, tables the rest).

It was then reproduced from scratch on 2026-11-08, from a fresh clone of this repository (commit 3d8fce61), with 60 threads on an otherwise idle 64-core machine. All stages, including the three special cap certificates, were regenerated; the chart tables had the same row counts as above, the regenerated model files matched the committed ones, and constructSeptopert printed the same instantiated line:

./get-cc-lib.sh
(cd proof && lake exe cache get)    # Mathlib build cache
search/septopert/rebuild.sh --threads 60 --work work-fresh

The search is multithreaded and not deterministic, so other runs may differ in details. Stages: model, tube, idcap, tubepack, cap, arccert, axiscert, pcapcert, search, pack, verify (--from STAGE, --only STAGE, --work DIR; the default is to start at model). Measured stage times (31 minutes in all):

Stage Time
model 9 s
tube 35 s
idcap 781 s
cap 24 s
arccert 96 s
axiscert 2 s
pcapcert 93 s
search 69 s
pack 41 s
verify 727 s

The work directory was 274 MB.

The certificate data is machine output, intended to be CC0:

  • proof/Septopert/data/axis-cap.cert (148 KB) and proof/Septopert/data/pcap-wp.cert, pcap-u.cert (349 KB and 1.5 MB): the axis cap and the P cap, in the repository;
  • arc.cert: 916 KB, the arc cap certificate (146 charts, 22,282 leaves);
  • cap-1_16.cert: 408 KB, the mirror cap certificate;
  • tube_packs/: 944 KB, 16 identity-tube tables and manifest.txt;
  • packs/: 37 MB, 64 solution packs, one for each of the 4 charts and 16 view-code triangles (about 732,000 rows).

The four not in the repository compress to about 6.8 MB (tar czf).

联系我们 contact @ memedata.com