7

Suppose I have a regular polytope $P$ which I'm representing as a graph $G_P$ with vertices and edges. I can already put this data into a computer program to find a list of symmetries of $P$---they're just the graph isomorphisms from $G_P$ to itself.

But as I understand it, half(?) of these isomorphisms will be orientation-preserving and the other half will be orientation-flipping.

How can I make a list of the graph isomorphisms that are orientation-preserving? Does the adjacency data from the graph (which is abstract and not embedded in any geometric space) already determine which isomorphisms preserve/flip orientation, or is it necessary to supply some additional structure such as an embedding of the graph, a directed traversal of the edges, or a list of the vertices of a face in clockwise order? Given whatever additional data, what should the algorithm be?

Edit: I'm wondering if I can find the orientation-preserving symmetries using algebra: I believe the square of any symmetry is a rotation. So there's a subgroup generated by the squares, and they're all rotations. Is this all the rotations (I'd generally doubt it), or how might we get the rest?

Maybe there's only a few groups that satisfy the properties that (1) they contain exactly half the members of the symmetry group (2) they contain all the perfect squares. In that case, it might be that any one of those groups could be the rotation group given an appropriate embedding --- i.e. each one is isomorphic to the rotation group. In that case I'd wonder if we need any additional constraints to guarantee that all the subgroups that satisfy these constraints are (isomorphic to) the group of rotations.


Edit #2: I can prove that the sign of a group automorphism is independent of choice of embedding, but I am still searching for a constructive method to compute it from the graph without using coordinates as an intermediate.

The proof that the sign is independent of the choice of embedding is as follows:

  1. Let $P$ be a regular polytope, fix $S\subset \mathbb{R}^n$ as points for $P$ in $\mathbb{R}^n$, let $G$ be the abstract graph of $P$ and let $V$ be its vertices.

  2. A graph automorphism is any bijection $\rho:V\rightarrow V$ that preserves and reflects the adjacency relationships of the graph $G$.

  3. A choice of coordinates for $G$ is a map $\chi: V\cong S\subset \mathbb{R}^n$. Let's fix one such coordinate choice as our canonical coordinates: $\chi :V\cong S$.

  4. Every other choice of coordinates factors uniquely through $\chi$: if $\chi^\prime$ is any other choice of coordinates, then $\chi^\prime = \sigma\circ \chi$ for some unique isometry $\sigma:\mathbb{R}^n\rightarrow \mathbb{R}^n$.

  5. Given a choice of coordinates $\chi$, every graph automorphism $\rho:V\rightarrow V$ corresponds to a unique isometry of Euclidean space $\sigma_\rho :\mathbb{R}^n\rightarrow \mathbb{R}^n$, such that $\sigma_\rho \circ \chi = \chi \circ \rho$ ("the isometry after embedding is equivalent to the automorphism before embedding")

  6. Let $\rho:V\rightarrow V$ be any graph automorphism, and let $\chi, \chi^\prime : V\cong S \subset \mathbb{R}^n$ be two embeddings of the graph. We'll show that, under the two embeddings $\chi,\chi^\prime$, the automorphism $\rho$ induces two isometries $\sigma,\sigma^\prime$ with the same sign. This shows that the isometry induced by $\rho$ has the same sign under any choice of coordinates, which means that $\rho$'s sign is determined independent of any embedding.

    By (5), we can find the unique isometries $\sigma$ and $\sigma^\prime$ such that $$\sigma\circ \chi = \chi \circ \rho$$ and $$\sigma^\prime \circ \chi^\prime = \chi^\prime \circ \rho.$$

    By (4), we can factor $\chi^\prime = \tau\circ \chi$ for some unique isometry $\tau$. Factoring $\tau$ out of the second equation above gives: $$\sigma^\prime \circ \tau \circ \chi = \tau \circ \chi \circ \rho$$

    Composing $\tau$ on the left of the first equation gives: $$\tau \circ \sigma \circ \chi = \tau \circ \chi \circ \rho $$

    The right hand sides of these two equations are equal; hence their left hand sides must be equal as well:

    $$\sigma^\prime \circ \tau \circ \chi = \tau \circ \sigma \circ \chi$$

    The choice of coordinates is injective, so we can drop it, leaving just a commutative square: $\sigma^\prime \circ \tau = \tau \circ \sigma$.

    Now every term in this equation is an isometry $\mathbb{R}^n\rightarrow \mathbb{R}^n$, meaning it has a determinant whose value is $\pm 1$.

    $$\det(\sigma^\prime \circ \tau) = \det(\tau \circ \sigma)$$ $$\det(\sigma^\prime)\cdot \det(\tau) = \det(\tau) \cdot \det(\sigma)$$ $$\det(\sigma^\prime) = \det(\sigma)$$ Hence the two isometries induced by the two choices of coordinates for the graph automorphism $\rho$ have the same sign, QED.

user326210
  • 19,274
  • 1
    Is the polytope $3$-dimensional or higher? The solution for the $3$ dimensional case has a simple description: a non-trivial isomorphism preservers orientation if the total number of preserved edges + the total number of preserved faces + the total number of preserved vertices is $2$. – user3257842 Nov 06 '22 at 16:02
  • Wikipedia uses the Burnside lemma (iirc) to compute the case for a cube. I think this can be done in general – FShrike Nov 06 '22 at 16:47
  • If you have a representation of the ($n$-dimensional) polytope's symmetry group in $O(n)$, then each symmetry will have determinant $\pm1$, indicating whether it preserves orientation. I don't think the results depend on which embedding you use. – Karl Nov 07 '22 at 21:00
  • @Karl I have a list of the symmetries, each one represented as a bijection from vertices to vertices. I'm not sure how to get a determinant from that data, but your approach would definitely work if I could. – user326210 Nov 07 '22 at 22:18
  • Hmm, how are you representing/describing the polytope in order to get the graph data? – Karl Nov 07 '22 at 22:34
  • @Karl Symbolically, depending on the polytope. For a 4-cube, e.g., the set of vertices is ${a,b}^4$ and neighbors are vertices that differ in exactly one coordinate. I could pick an embedding --- but I wonder if topology could be sufficient. – user326210 Nov 07 '22 at 23:38
  • 2
    For dimensions $n$ higher than $4$, you only have the $n$-simplex, $n$-cube and $n$-orthoplex as regular polytopes. It's feasible to construct a criterion for each case. – user3257842 Nov 08 '22 at 15:04
  • @user3257842 That's useful. I know a graph-based criterion for cubes in 3D, which is to check how the isomorphism flips the diagonals. I don't know criteria for other shapes/dimensions. Could you suggest a reference? – user326210 Nov 08 '22 at 21:57
  • 1
    For an $n$-simplex you simply check the parity of the vertex permutation induced by the isomorphism.

    Eg. for a $4$-simplex you have vertices labelled as $(1,2,3,4,5)$ An isomorphism will give you a permutation, eg $(2,1,4,5,3)$, and you check its parity.

    For $n$-cubes and $n$-orthoplex, the embedding is the simplest way to go. The cube vertices will have coordinates $(\pm 1, \pm 1, \dots \pm 1)$ . The orthoplex vertices will have coordinates $(0,0,\dots, 0,\pm1, 0, \dots, 0)$.

    You select $n$ vertices of non-zero determinant, and see if the isomorphism preserves det sign.

    – user3257842 Nov 09 '22 at 08:01
  • @user3257842 Thanks. That solves the simplex, and I'll think about the others to see if I can find a purely adjacency-based way to characterize whether the determinant changes sign. I wonder if it depends on embedding. – user326210 Nov 09 '22 at 08:49
  • The rules for $n$-cube embedding are: Two vertices are connected if their coordinate values (of shape $(\pm 1, \pm 1, \dots \pm 1)$ ) differ at exactly one place. The rules for $n$-orthoplex embedding are: Two vertices are connected if their coordinate values (of shape $(0,0,\dots, 0,\pm1, 0, \dots, 0)$) are not negatives of one-another.

    You could use those rules to associate (in a way unique up to isomorphism) embedding coordinates to your vertices, then do the determinant trick. The only question remaining is if this could be done without the intermediate step.

    – user3257842 Nov 09 '22 at 09:49
  • For the $n$-cube you could start by associating $(1, 1, \dots 1)$ to a corner and $(1, 1, \dots -1, \dots 1)$ each of it's $n$ neighbors, then the rules will give you a unique full embedding. (compute coordinates to vertices of distance $2$ from the start, then distance $3$, etc. ).

    For the $n$-orthoplex, the $2n$ vertices should group into $n$ pairs of unconnected ones. Associate $(0,0,…,0,±1,0,…,0)$ for each of those pairs, and you have an embedding.

    – user3257842 Nov 09 '22 at 09:59
  • For both the $n$-cube and the orthoplex, there are exactly $n! * 2^n $ isomorphisms. This is exactly the number of ways you could assign rules-respecting coordinates to the graph. – user3257842 Nov 09 '22 at 10:05

1 Answers1

0

If you are considering planar graphs, you could think about indexing the vertices in any arbitrary way. Then list any polygon by means of an (oriented!) circuit of vertex indices, best done (in order to become unique) by breaking it up so that it gets the smallest lexicographical sequence.

Now apply your group action onto the graph, and look where the vertices will be mapped to. Thus you again could list the polygonal sequences in the same way again. Whenever those keep their order you where using a "rotation", if they would instead reverse their respective sequenceing order, then you had used a "reflection".

--- rk

  • In dimensions above $3$, isn't it possible to rotate a polygon of the polytope $180$ degrees around an axis of symmetry? – David K Nov 07 '22 at 12:20