Try
Suppose a finite group of spatial rotations has order 60, and a unit vector is fixed by a subgroup of order 5. Its orbit has 12 points. Does that alone make the points the vertices of an icosahedron? No: a vertex count is not a proof of edge lengths, faces, or uniqueness.
Here is the additional information we will extract. Form the convex hull of those orbit points. An edge cannot have more than two rotational symmetries preserving it. Its orbit therefore has at least 30 edges. A convex polyhedron with 12 vertices has at most 30 edges. Equality will force every face to be triangular and every edge to belong to the same group orbit. Only after proving those statements can we recover metric geometry.
Our final reconstruction theorem is conditional on exact coordinate models supplied in the coordinate-model lesson. This lesson proves the bridge and the uniqueness mechanism; it does not silently assume that the model coordinates form the claimed polyhedra.
Predict
Recall the pole count. The exceptional stabilizer triples are (2,3,3), (2,3,4), and (2,3,5), with group orders N=12,24,60. Choose an orbit O whose point stabilizer has order d=3,4,5, respectively. Its size is V=N/d, hence 4,6,12.
Predict which implications need proof: that the orbit spans space; that its points are actually vertices of its convex hull; that equal edge counts imply one edge orbit; and that triangular faces imply equal edge lengths. The last implication alone is false. Equality of lengths will come from the single edge orbit and preservation of distance.
We use a convex polyhedron to mean a three-dimensional convex hull of finitely many points. Its edges and faces are its one- and two-dimensional exposed boundary pieces. V, E and F will denote its numbers of vertices, edges and faces. The letter F here is a count, not a field.
Experiment
Complete this chain before opening the explanation: orbit size, largest possible edge count, smallest possible edge-orbit size, then average vertex degree. Each row concerns one of the exceptional triples, not an arbitrary collection of points.
| Stabilizer triple | N | V | 3V−6 | N/2 | Forced degree 2E/V |
|---|
| (2,3,3) | 12 | 4 | 6 | 6 | 3 |
| (2,3,4) | 24 | 6 | 12 | 12 | 4 |
| (2,3,5) | 60 | 12 | 30 | 30 | 5 |
Why matching bounds have geometric consequences
An edge exists. Its orbit already contains at least N/2 edges, while the whole hull has at most 3V−6=N/2. Thus this orbit contains every edge. The upper-bound equality also forces every face to have exactly three edges. Vertex transitivity makes every vertex degree equal to the displayed average. These are three separate deductions from the same equality, with different hypotheses.
Change just the information: suppose someone gives a 12-vertex polyhedron with 24 edges but no transitive rotation group. Euler’s relation allows 14 faces, so the numbers alone are consistent with nontriangular faces. The step forcing 30 edges would be missing. Do not import a group-theoretic lower bound into a problem that supplied no group.
Understand
Why the exceptional orbit has a genuine three-dimensional hull
First exclude a hidden invariant line. If a finite rotation group preserves an unoriented line, restriction to its perpendicular plane embeds it in O(2). The sign of its action on the line equals the determinant of its action on the plane. The proved finite O(2) classification therefore makes it cyclic or rotational-dihedral. In the latter case the poles on the main axis have stabilizer order n, while the horizontal half-turn poles have stabilizer order 2; for n≥2 its pole triple is (2,2,n). The case n=1 is a single half-turn and is cyclic. None of these has an exceptional triple. Thus our G preserves no line.
The average of O is fixed by G, because G permutes its summands. A nonzero fixed vector would supply an invariant line, so the average is zero. The linear span of O is invariant. If its dimension were one, it would be an invariant line; if two, its orthogonal complement would be one. Hence O spans ℝ³. Since zero is an affine combination of its points, its affine span is also ℝ³.
The strictly positive average places zero in the interior, not merely in the hull. Here is an explicit justification. Start with coefficients 1/V at every point. Choose four affinely independent orbit points v₀,v₁,v₂,v₃. Every sufficiently small vector z can be written Σⱼ₌₁³ cⱼ(vⱼ−v₀) with all cⱼ arbitrarily small, by the inverse of this fixed basis matrix. Add cⱼ to the coefficient of vⱼ and subtract their sum from the coefficient of v₀. The coefficients still sum to 1 and remain positive for sufficiently small z. Thus a whole neighborhood of zero consists of convex combinations.
Finally every orbit point v is a vertex: the functional x↦v·x takes value 1 at v and strictly less than 1 at every other unit orbit point. It therefore uniquely exposes v. There are exactly V vertices, with none lost inside a face.
The planar bound, including its topological bookkeeping
The boundary of a convex polyhedron is a sphere: radial projection from an interior point meets the boundary once on each ray, continuously in either direction. Its edges give a connected graph embedded on that sphere, with polygonal disk faces. Connectivity can also be seen directly by cutting a vertex with a small plane: incident faces form a connected polygonal fan; face paths across the boundary connect all these fans. Removing a point inside one face and projecting the remaining sphere to a plane gives a plane graph with that face as its outer face.
For clarity, Euler’s relation here is not an assumed count of a familiar solid. Choose a spanning tree of this connected graph. Delete edges outside the tree one at a time. Each deleted edge still belongs to a cycle, because the tree supplies another path between its endpoints. A simple cycle separates the plane into two sides, so that edge has distinct faces on its sides; deleting it merges those faces. Thus E and F each decrease by one, V stays fixed, and V−E+F is unchanged. A plane tree has V−1 edges and one complementary face: remove pendant edges successively to see that they create no enclosed region. Its value is 2, proving V−E+F=2 for the original embedding. This uses the elementary separation property of a simple polygonal closed curve, not a classification of regular solids.
Every face has at least three edges; every edge is incident to two faces. Hence 3F≤2E. Substituting F=2−V+E gives E≤3V−6. Equality holds exactly when each face has three edges.
The edge orbit supplies the opposite inequality
An edge cannot have antipodal endpoints: their segment passes through zero, an interior point. Distinct unit endpoints of an edge are therefore linearly independent. A rotation fixing both endpoints is the identity by the axis theorem. Consequently a rotation preserving the edge as a set is determined by its permutation of the two endpoints, and the setwise edge stabilizer has at most two elements.
Orbit–stabilizer gives at least N/2 edges in each edge orbit. The table shows N/2=3V−6. The planar bound therefore forces E=N/2, exactly one edge orbit and triangular faces. Since rotations preserve distances, all edges have equal length. Transitivity on the V vertices gives their common degree 2E/V=d.
A convex vertex fan determines the angle, not a star polygon
Fix a vertex v. Its stabilizer has order d. No nonidentity element of this stabilizer can fix another vertex adjacent to v: that would fix two independent vectors. Thus its action on the d neighbors is free and transitive. All neighbors have the same dot product t with v, and lie on a circle in the plane v·x=t.
Why are consecutive angular neighbors joined by a face? Cut off v by a plane v·x=1−ε, with ε positive and sufficiently small that no other vertex lies above the plane. The section is a convex polygon. Its vertices lie on the incident edges, and its sides lie in the incident faces. This follows by intersecting the local supporting half-spaces at v with the cutting plane; each polygon corner is an extreme ray of the local cone, hence an incident edge. The equal neighbor heights mean the intersection points occur at the same fraction along each edge from v. Consequently this polygon has the same cyclic angular order as the neighbors. Its sides correspond to consecutive neighbors, and the triangular faces join those pairs. Convexity rules out replacing this order by a crossing star polygon.
The finite cyclic stabilizer contains its least positive turn 2π/d. Write a neighbor as t v+√(1−t²)w with w a unit vector in v⊥. Consecutive neighbors have perpendicular components separated by 2π/d. They are themselves joined by an edge, so their dot product is t as well. Therefore, putting c=cos(2π/d),
t = t²+(1−t²)c.
Factoring gives (t−1)((1−c)t−c)=0. The root t=1 would identify two vertices and is forbidden. The remaining root t=c/(1−c) is −1/3, 0 or 1/√5 for d=3,4,5. For d=5 use cos(2π/5)=(√5−1)/4. These values determine the common chord length √(2−2t).
Face propagation proves conditional uniqueness
Suppose the coordinate-model lesson supplies a model hull with the same V, unit vertices, triangular faces, and the same edge dot product t. Choose one face in each hull. Its three vertex vectors have Gram matrix with diagonal 1 and off-diagonal t. Its determinant is (1−t)²(1+2t)>0 for our three values, so the vectors form a basis. The linear map matching the ordered triples preserves this Gram matrix and is orthogonal. Align the first hull with the model by this map.
Across a fixed edge with endpoints a,b, every possible third vertex x of a triangular face satisfies
x·a=t, x·b=t, x·x=1.
The two independent linear equations describe a line perpendicular to span(a,b). Its intersection with the sphere has exactly two points: one is the existing third vertex, which has a nonzero perpendicular component by the positive Gram determinant; the other is its reflection across span(a,b). The adjacent face must use that other point. There are exactly two incident faces at an edge, and two distinct triangular faces cannot have the same three vertices. Thus the neighboring face is forced and agrees with the model’s neighboring face.
This local argument reaches every face. To justify that assertion, radially project the boundary to the sphere. Join interior points of two faces by a path avoiding the finitely many vertices and meeting edges transversely at interior points. Such a path is obtained with finitely many great-circle segments after small changes of intermediate points; avoid the finitely many forbidden vertices, coincidences with edge arcs, and tangencies. Its successive face crossings give a path in the face-adjacency graph. Hence that graph is connected.
Propagating from the aligned face puts every vertex of the first hull among model vertices. The counts are equal, so their vertex sets and convex hulls are identical. This proves uniqueness conditional on the model certificate; it does not prove that unverified coordinates have the required faces.
There is also no room for extra rotational symmetries. A rotation preserving this hull is determined by the images of an ordered adjacent pair, since two independent vectors determine an orientation-preserving orthogonal map. There are 2E=N such pairs, so the full rotation group has at most N elements. It already contains our conjugated G of order N, hence equals it. If the aligning matrix Q has determinant −1, then −Q has determinant 1 and induces exactly the same conjugation on matrices. Thus the classification is up to conjugacy within SO(3), even if the first geometric alignment reversed orientation.
Apply
A group with pole triple (2,3,4) acts on its orbit with stabilizer order 4. Derive the edge count, face count and edge dot product without drawing an octahedron. Then diagnose this proposed shortcut: “Two neighboring vertices determine the next face uniquely.”
Check the counts and repair the uniqueness claim
Here N=24 and V=6. The lower and upper bounds both give E=12; triangularity gives F=8. Degree is 4, so c=cos(π/2)=0 and t=0. The common chord length is √2. Two edge vertices alone permit two third vertices, not one. Once one incident face is fixed, the other solution is forced for the adjacent face. Global propagation additionally needs connectedness of the face-adjacency graph and a verified model with matching triangular faces.
As a changed boundary check, apply E≤3V−6 to a cube’s eight vertices. Its 12 edges satisfy a strict inequality, since 12 is less than 18. Its square faces do not satisfy the equality argument. This does not contradict the octahedral rotation group of order 24: the cube vertex orbit has stabilizer order 3, whereas the exceptional reconstruction deliberately selected the orbit with stabilizer order 4 and size 6.
Continue
Use exact polyhedral coordinate models. That lesson must verify unit coordinates, supporting triangular faces and their common edge products before invoking this conditional uniqueness proof. Keep construction, uniqueness and identification with an abstract permutation group as separate proof obligations.
Continue to polyhedral permutation identifications to prove the abstract names through faithful actions, rather than identifying groups only by their orders.
