Arxiv:2108.05893V1 [Math.CO] 12 Aug 2021

arXiv:2108.05893v1 [math.CO] 12 Aug 2021 httedrc rdc Aut product direct the that neetdi ae hr hssbru sproper: is 1.2 subgroup Definition this where cases in interested ento 1.1 Definition undirected.) and rhs“wn etcs(e ento . eo) then below), 2.5 Definition X (see vertices “twin” has or ento 1.3 Definition examples: trivial these out rp ob ntbe SeDfiiin23frtedfiiinof definition the for 2.3 Definition Cay( (See notation graph” unstable. be to graph then unstable, and iatt graph bipartite Letting Let .Wlo on h olwn neetn odtosta f that conditions interesting following the found Wilson S. ti ayt e adwl nw)ta if that known) well (and see to easy is It Date stetiilgahwt nyoevre) h olwn defi following The vertex). one only with graph trivial the is DMRHUJDUROVI ADEMIR ( v, NATMRHSSO H OBECOVER DOUBLE THE OF AUTOMORPHISMS ON )i daett ( to adjacent is 0) uut1,2021. 13, August : X atgah htwr o rvosykon)W lofidalo all find also We 2 order known.) of un graphs previously circulant nontrivially not unstable of were nontrivially families that infinite graphs yield lant neighbors (These same the unstable. exactly have X of vertices distinct two no unstable nontrivially n icln rp forder of graph circulant number. K Abstract. htd o oefo uoopim fisfactors its of automorphisms from come not do that 2 2 = eacruatgah Algah nti ae r nt,sim finite, are paper this in graphs (All graph. circulant a be efidtrenwcniin htec ml icln graph circulant a imply each that conditions new three find We u eut ml htteede o xs otiilyun nontrivially a exist not does there that imply results Our S as aldtecnncldul oe of cover double canonical the called (also 2 p o oepienumber prime some for , etesmercgopo h -lmn set 2-element the on group symmetric the be BX [1 .160]) p. ([11, c.[0 .360]) p. [20, (cf. ([17]) graph A X FACRUATGRAPH CIRCULANT A OF with ,S G, is . C, ´ w, The otiilyunstable nontrivially fi susal,cnetd n obprie and nonbipartite, and connected, unstable, is it if X V ).) ¯ DOR )in 1) ( ssi obe to said is 1. BX n . aoia iatt obecover double bipartite canonical X fAut If ¯ fadol feither if only and if EMITROVI DE Introduction . = ) BX × If p S X hti ogun o3mdl 4. modulo 3 to congruent is that V 2 BX 1 ( sasbru fAut of subgroup a is scnetd obprie twin-free, nonbipartite, connected, is unstable X ⇐⇒ ) Aut 6= X { × ,ADDV IT MORRIS WITTE DAVE AND C, ´ . sdsonce,o sbipartite, is or disconnected, is 0 ftedrc product direct the if , X v p 1 X n where , } a automorphisms has ) sajcn to adjacent is sod or odd, is where , × S X 2 X then , p and susal (unless unstable is tbecircu- stable sayprime any is { < n . K reacirculant a orce 0 BX , 2 X 1 tis It . stable h “Cayley the of } the f ,or 8, iinrules nition w X is ti clear is it , eare We . × X is in unstable X. sthe is ple, . 2 ADEMIRHUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

Theorem 1.4 (Wilson [20, Appendix A.1] (and [16, p. 156])). Let X = Cay(Zn,S) be a circulant graph, such that n is even. Let Se = S ∩ 2Zn and So = S \ Se. If any of the following conditions is true, then X is unstable. (C.1) There is a nonzero element h of 2Zn, such that h + Se = Se. ′ (C.2 ) n is divisible by 4, and there exists h ∈ 1 + 2Zn, such that (a) 2h + So = So, and (b) for each s ∈ S, such that s ≡ 0 or −h (mod 4), we have s + h ∈ S. ′ (C.3 ) There is a subgroup H of Zn, such that the set R = { s ∈ S | s + H 6⊆ S }, is nonempty and has the property that if we let d = gcd R ∪ {n} , then n/d is even, r/d is odd for every r ∈ R, and either H * dZ n or H ⊆ 2dZn. × (C.4) There exists m ∈ Zn , such that (n/2) + mS = S. Remark 1.5. As will be explained in Remark 3.14, the statements (C.2′) and (C.3′) are slightly corrected versions of the original statements of The- orems C.2 and C.3 that appear in [20]. The correction (C.2′) is due to Qin-Xia-Zhou [16, p. 156]. Definition 1.6. We say that X has Wilson type (C.1), (C.2′), (C.3′), or (C.4), respectively, if it satisfies the corresponding condition of Theo- rem 1.4. In this terminology (modulo the corrections mentioned in Remark 1.5), Wilson [20, p. 377] conjectured that every nontrivially unstable circulant graph has a Wilson type. Unfortunately, this is not true, because counterexamples of order 24 were found by Qin-Xia-Zhou [16, p. 156] (cf. Obser- vation 6.1). Three of our results provide new conditions that force a circulant graph to be unstable. (These conditions provide infinitely many new counterexamples.) It seems likely that other conditions (and additional counterexamples) remain to be discovered. Our first condition includes (C.1) as the special case where K = Zn, ′ ′ includes (C.2 ) as the special case where K = 2Zn, and includes (C.3 ) as another special case (see Proposition 3.4).

Theorem 3.2 (cf. [20, Thms. 1, C.1, and C.3]). Let X = Cay(Zn,S) be a circulant graph. Choose nontrivial subgroups H and K of Zn, such that |K| is even, and let Ko = K \ 2K. If either (1) S + H ⊆ S ∪ (Ko + H) and H ∩ Ko = ∅, or (2) (S \ Ko)+ H ⊆ S ∪ Ko and either |H|= 6 2 or |K| is divisible by 4, then X is not stable. Our second condition is the following generalization of (C.4). AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 3

Proposition 3.7. Assume X = Cay(Zn,S) is a circulant graph of even order. If X =∼ Cay Zn, (n/2) + S , then X is unstable. Letting Se be the set of even elements of S, and So be the set of odd elements, as in Theorem 1.4, our third condition states that if Cay(Zn,Se) is unstable, and So is invariant under suﬃciently many translations, then X is unstable.

Proposition 3.12. Assume X = Cay(Zn,S) is a circulant graph of even order. If there exist permutations α and β of 2Zn, and a subgroup H of 2Zn, such that: (1) α 6= β,

(2) if the vertices u, v ∈ 2Zn are adjacent, then the vertices α(u) and β(v) are also adjacent, (3) w + H ⊆ S, for all odd w ∈ S, and

(4) α(v) − v ∈ H and β(v) − v ∈ H, for all v ∈ 2Zn, then X is unstable. Wilson’s conjecture is vacuously true for graphs of odd order: Theorem 1.10 (Fernandez-Hujdurovi´c[3] (or [12])). There are no nontrivially unstable circulant graphs of odd order. Therefore, only graphs of even order are of interest. Another of our main results shows that the conjecture is true in the easiest of these interesting cases: Theorem 5.1. If p is a prime number, then every nontrivially unstable circulant graph of order 2p has Wilson type (C.4). This allows us to provide an explicit description of the nontrivially unstable circulant graphs of such orders:

Corollary 5.7. Let X = Cay(Z2p,S) be a connected, twin-free, nonbipartite, circulant graph of order 2p, where p is prime, and let Se = S ∩ 2Z2p. × The graph X is nontrivially unstable if and only if there exists m ∈ Z2p, 2 such that m Se = Se, mSe 6= Se, and S = Se ∪ (n/2) + mSe . It also makes it possible to strengthen Theorem 1.10 to a determination of all the possible orders of nontrivially unstable circulant graphs: Corollary 5.8. For n ∈ Z+, there does not exist a nontrivially unstable circulant graph of order n if and only if either n is odd, or n< 8, or n = 2p, for some prime number p ≡ 3 (mod 4). Here is an outline of the paper. After this introduction comes a short section of preliminaries. Our main results, which were described above, are proved in Section 3 (Theorem 3.2 and Propositions 3.7 and 3.12) and Section 5 (Theorem 5.1 and its corollaries). In between, Section 4 discusses a lemma from [12]. Finally, Section 6 brieﬂy presents a few minor results that were obtained by computer exploration. 4 ADEMIRHUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

Acknowledgments. The work of Ademir Hujdurovi´cis supported in part by the Slovenian Research Agency (research program P1-0404 and research projects N1-0102, N1-0140, N1-0159, N1-0208, J1-1691, J1-1694, J1-1695, J1-2451 and J1-9110).

2. Preliminaries For emphasis, and ease of reference, we repeat a basic assumption from the first paragraph of the introduction (and add an exception): Assumption 2.1. All graphs in this paper are finite, undirected, and simple (no loops or multiple edges), except that loops will be allowed in Section 4 (see Assumption 4.1). Notation 2.2. For convenience, proofs will sometimes use the following abbreviation: n = n/2. Definition 2.3. Let S be a subset of an abelian group G, such that −s ∈ S for all s ∈ S, and 0 ∈/ S. (1) The Cayley graph Cay(G, S) is the graph whose vertices are the elements of G, and with an edge from v to w if and only if w = v + s for some s ∈ S (cf. [10, §1]). (2) For s ∈ S, we lets ˜ = (s, 1). (3) Note that if X = Cay(G, S), and we let S = { s˜ | s ∈ S }, then

BX = Cay G × Z2, S . e For s ∈ S, we say that an edge uv of BX is an s-edge if v = u ± s˜. e Remark 2.4. The reason that the set S in Cay(G, S) is not allowed to contain 0 is that Assumption 2.1 forbids the graphs in this paper from having loops. Deﬁnition 2.5 (Kotlov-Lov´asz [7]). A graph X is twin-free if there do not exist two distinct vertices that have exactly the same neighbors.

see Remark 2.6. It is easy to see (and well known) that Cay(Zn,S) is twin-free note if and only if there does not exist a nonzero h ∈ Zn, such that h + S = S. A.1 ! The notion of “block” is a fundamental concept in the theory of permutation groups, but we need only the following special case:

Deﬁnition 2.7 (cf. [5, pp. 12–13]). Let X = Cay(Zn,S) be a circulant graph. A nonempty subset B of V (BX) is a block for the action of Aut BX if, for every α ∈ Aut BX, we have either α(B)= B or α(B) ∩B = ∅.

see It is easy to see that this implies B is a coset of some subgroup H of Zn ×Z2, note and that every coset of H is a block. Indeed, the action of Aut BX permutes A.2 ! these cosets, so there is a natural action of Aut BX on the set of cosets. AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 5

Deﬁnition 2.8 (cf. [10, Defn. 3.1]). To say that a circulant graph X = ′ Cay(Zn,S) has the Cayley Isomorphism Property means that if S is a subset ∼ ′ × of Zn, such that X = Cay(Zn,S ), then there exists m ∈ Zn , such that S′ = mS. Theorem 2.9 (Muzychuk [13]). If n is either square-free, or twice a square- free number, then every Cayley graph on Zn has the Cayley Isomorphism Property.

3. Some conditions that imply instability The following known result is elementary, but, for the reader’s convenience, we briefly recall the proof. Lemma 3.1 ([9, Thm. 3.2], [11, Prop. 4.2]). Let X be a graph. If there exist permutations α and β of V (X), such that α 6= β and, for every edge uv of X, the vertex α(u) is adjacent to β(v), then X is unstable. The converse holds if BX is connected. Sketch of proof. (⇒) Define ϕ ∈ Aut(BX) by α(v), i if i = 0, ϕ(v, i)= (β(v), i if i = 1. Since α 6= β, we have ϕ∈ / Aut X × S2. (⇐) Let ϕ ∈ Aut BX, such that ϕ∈ / Aut X × S2. Since BX is connected, we know that its bipartition is unique, so the bipartition sets are blocks for the action of Aut BX. Therefore, after composing by a translation, we may assume ϕ Zn × {i} = Zn × {i} for i = 0, 1. So we may define permutations α and β of V (X) by ϕ(v, 0) = α(v), 0 and ϕ(v, 1) = β(v), 1 .

Theorem 3.2 (cf. [20, Thms. 1, C.1, and C.3]). Let X = Cay(Zn,S) be a circulant graph. Choose nontrivial subgroups H and K of Zn, such that |K| is even, and let Ko = K \ 2K. If either (1) S + H ⊆ S ∪ (Ko + H) and H ∩ Ko = ∅, or (2) (S \ Ko)+ H ⊆ S ∪ Ko and either |H|= 6 2 or |K| is divisible by 4, then X is not stable. Proof (cf. proof of [20, Thm. 1]). Let h be a generator of H. We will deﬁne permutations α and β of Zn, such that Lemma 3.1 applies. (1) Deﬁne x + h if x ∈ 2K + H; x + h if x ∈ K + H; α(x)= β(x)= o (x otherwise; (x otherwise.

Note that 0 ∈/ Ko + H (because H ∩ Ko = ∅), so β(0) = 0. Since α(0) = h, this implies α 6= β. 6 ADEMIRHUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

Given an edge uv of X, we wish to show that α(u) is adjacent to β(v). We may assume that either u is moved by α or v is moved by β. In fact, we may assume that exactly one of the vertices is moved, for otherwise, β(v) − α(u) = (v + h) − (u + h)= v − u ∈ S.

This means we may assume that either u ∈ 2K + H or v ∈ Ko + H, but not both. Letting s = v − u ∈ S, this implies s∈ / Ko + H. Also, we have β(v) − α(u) ∈ (v + H) − (u + H) = (v − u)+ H = s + H, so we may write β(v) − α(u) = s + h′, for some h′ ∈ H. By the ﬁrst ′ assumption of (1), we know s + h ∈ S ∪ (Ko + H). Since s∈ / Ko + H, this implies s + h′ ∈ S, so α(u) is adjacent to β(v).

(2) If h ∈ 2K, then Ko + H = Ko, so

(S ∩ Ko)+ H ⊆ Ko + H = Ko.

Since, by the ﬁrst assumption of (2), we also have (S \ Ko)+ H ⊆ S ∪ Ko, this implies that (1) applies. Therefore, we may assume h∈ / 2K. Deﬁne

x + h if x ∈ 2K; x + h if x ∈ Ko; α(x)= x − h if x ∈ 2K + h; β(x)= x − h if x ∈ Ko + h; x otherwise x otherwise.

We claim that α 6= β. Note that α(0) = h. Therefore, if α = β, then we must have β(0) = h. Since 0 ∈/ Ko, this implies that 0 ∈ Ko+h (which means see h ∈ Ko) and −h = h (which means |h| = 2). Since |h| = 2 and h ∈ Ko, note we see that |H| = 2 and that |K| is not divisible by 4. This contradicts the A.3 ! second half of assumption (2), so the proof of the claim is complete. Given an edge uv of X, we wish to show that α(u) is adjacent to β(v). That is, we wish to show β(v) − α(u) ∈ S. We have v = u + s for some s ∈ S. We may assume β(v) − α(u) 6= v − u. In particular, we cannot have both α(u)= u and β(v)= v. Therefore,

either u ∈ 2K ∪ (2K + h) or v ∈ Ko ∪ (Ko + h).

Case 1. Assume u ∈ 2K ∪ (2K + h) and v ∈ Ko ∪ (Ko + h). We consider two different possibilities, but both of the arguments are very similar. Subcase 1.1. Assume u ∈ 2K. Then α(u) = u + h. Since β(v) − α(u) 6= v − u, this implies β(v) 6= v + h, so v∈ / Ko. By the assumption of Case 1, this implies v ∈ Ko + h, so β(v)= v − h. Hence, β(v) − α(u)= s − 2h. We have u ∈ 2K and v ∈ Ko + h, so s = v − u ∈ Ko + h, which means s − h ∈ Ko. Since h∈ / 2K, this implies s∈ / Ko and s − 2h∈ / Ko. Since AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 7 s∈ / Ko, the first assumption of (2) tells us s+H ⊆ S∪Ko. Since s−2h∈ / Ko, this implies s − 2h ∈ S. So α(u) is adjacent to β(v). Subcase 1.2. Assume u ∈ 2K + h. We have α(u) = u − h. Since β(v) − α(u) 6= v −u, this implies β(v) 6= v −h, so v∈ / Ko +h. By the assumption of Case 1, this implies v ∈ Ko, so β(v)= v + h. Hence, β(v) − α(u)= s + 2h. We have u ∈ 2K + h and v ∈ Ko, so s = v − u ∈ Ko − h, which means s + h ∈ Ko. Since h∈ / 2K, this implies s∈ / Ko and s + 2h∈ / Ko. Since s∈ / Ko, the first assumption of (2) tells us s+H ⊆ S∪Ko. Since s+2h∈ / Ko, this implies s + 2h ∈ S. So α(u) is adjacent to β(v). Case 2. Assume Case 1 does not apply. As in Case 1, we consider two different possibilities, but both of the arguments are very similar. Subcase 2.1. Assume u ∈ 2K ∪ (2K + h). Choose δ ∈ {0, 1}, such that u ∈ 2K + δh. We have α(u) = u + ǫh, where ǫ = 1 − 2δ, and we also have v∈ / Ko ∪ (Ko + h), since Case 1 does not apply, so β(v) = v. Since u ∈ 2K + δh, but u + s = v∈ / Ko + δh, we have s∈ / Ko. So the first assumption of (2) tells us s + H ⊆ S ∪ Ko, so s − ǫh ∈ S ∪ Ko. Since β(v) − α(u)= s − ǫh, then we may assume s − ǫh ∈ Ko (otherwise, α(u) is adjacent to β(v), as desired), so s ∈ Ko + ǫh. Then

v = u + s ∈ (2K + δh) + (Ko + ǫh)= Ko + (δ + ǫ)h = Ko + (1 − δ)h.

Since 1 − δ ∈ {0, 1}, but v∈ / Ko ∪ (Ko + h), this is a contradiction.

Subcase 2.2. Assume v ∈ Ko ∪ (Ko + h). Choose δ ∈ {0, 1}, such that v ∈ Ko + δh. We have β(v) = v + ǫh, where ǫ = 1 − 2δ, and we also have u∈ / 2K ∪ (2K + h), since Case 1 does not apply, so α(u) = u. Since v ∈ Ko + δh, but v − s = u∈ / 2K + δh, we have s∈ / Ko. So the ﬁrst assumption of (2) tells us s + H ⊆ S ∪ Ko, so s + ǫh ∈ S ∪ Ko. Since β(v) − α(u)= s + ǫh, then we may assume s + ǫh ∈ Ko (otherwise, α(u) is adjacent to β(v), as desired), so s ∈ Ko − ǫh. Then

u = v − s ∈ (Ko + δh) − (Ko − ǫh) = 2K + (δ + ǫ)h = 2K + (1 − δ)h. Since 1 − δ ∈ {0, 1}, but u∈ / 2K ∪ (2K + h), this is a contradiction. Remark 3.3. In the notation of Theorem 3.2, it is clear that

S ∩ (Ko + H) + H ⊆ (Ko + H)+ H = Ko + H. Therefore, the ﬁrst condition of 3.2(1) can be restated as:

S \ (Ko + H) + H ⊆ S. Wilson types (C.1), (C.2 ′), and (C.3′) are special cases of Theorem 3.2(2): ′ ′ Proposition 3.4. If Cay(Zn,S) has Wilson type (C.1), (C.2 ), or (C.3 ), then there are nontrivial subgroups H and K of Zn that satisfy the conditions given in part (2) of Theorem 3.2 (and |K| is even). 8 ADEMIRHUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

Proof. (C.1) Let K = Zn and H = hhi. Then

(S \ Ko)+ H = Se + hhi = Se ⊆ S, see so the ﬁrst condition of 3.2(2) is satisﬁed. Also, since h ∈ 2Zn = 2K, it note must be true that either |H|=2 6 or |K| is divisible by 4. ′ A.4 ! (C.2 ) Let H = hhi and K = 2Zn. (Note that |H| > 2, since h is odd and n is divisible by 4.) Then

Ko = { x ∈ Zn | x ≡ 2 (mod 4) }.

We will show that (S \ Ko)+ H ⊆ S ∪ Ko. We may assume h ≡ 1 (mod 4), by applying the graph automorphism x 7→ −x if necessary. Fix some s ∈ S \ Ko. Suppose, first, that s 6≡ 0 (mod 4) (and recall that s∈ / Ko, so s 6≡ ′ 2 (mod 4)), so s is odd. This means s ∈ So, so we see from C.2 (a) that s + 2kh ∈ S for all k ∈ Z. If s + 2kh ≡ 3 (mod 4), then s + (2k + 1)h ∈ S ′ (by C.2 (b)). If s + 2kh ≡ 1 (mod 4), then s + (2k + 1)h ∈ Ko. Thus, we have s + H ⊆ S ∪ Ko. Now, suppose s ≡ 0 (mod 4). Then s + h ∈ S (by C.2′(b)). Now, since s + h 6≡ 0 (mod 4), the previous case tells us that s + h + H ⊆ S ∪ Ko. Since h + H = H, this means s + H ⊆ S ∪ Ko. (C.3′) Let K = hRi = hdi. Since n/d is even, we know that K has even order. Then, since r/d is odd for every r ∈ R, we see that R ⊆ Ko. By the definition of R, this means (S \ Ko)+ H ⊆ S, so the first condition of 3.2(2) is satisfied. Also, since either H * dZn or H ⊆ 2dZn, we know that either H * K or H ⊆ 2K. If H * K and |H| = 2, then it is clear that H ∩ K = {0} ⊆ 2K. see Thus, in both cases, we have H ∩ K ⊆ 2K, which easily implies that either note |H|=2 6 or |K| is divisible by 4. A.5 ! There is a strong converse to Proposition 3.4 when n is not divisible by 4:

Proposition 3.5. If X = Cay(Zn,S), H, and K satisfy the conditions of Theorem 3.2(2), and n is not divisible by 4, then X has Wilson type (C.1). Proof. Since n is not divisible by 4, it is not possible for |K| to be divisible by 4, so the second half of Theorem 3.2(2) tells us that |H| > 2. This implies that He := H ∩2Zn is nontrivial. Also, since n is not divisible by 4, we know that Ko ∩ 2Zn = ∅, so

Se + He = (S ∩ 2Zn)+ He ⊆ (S \ Ko)+ H ⊆ S ∪ Ko ⊆ S ∪ (Zn \ 2Zn).

Since He ⊆ 2Zn, we also know that Se + He ⊆ 2Zn. Therefore, we conclude that Se + He ⊆ Se, so X has Wilson type (C.1). Remark 3.6. By combining Propositions 3.4 and 3.5, we see that if X has a Wilson type, and n is not divisible by 4, then X must have Wilson type (C.1) or (C.4).

Proposition 3.7. Assume X = Cay(Zn,S) is a circulant graph of even order. If X =∼ Cay Zn,S + (n/2) , then X is unstable. AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 9

Proof. If α is an isomorphism from Cay(Zn,S) to Cay(Zn,S + n), then see Lemma 3.1 applies with β(x)= α(x)+ n. note A.6 ! Remarks 3.8. (1) Proposition 3.7 is a generalization of Wilson type (C.4). To see this, note that if n + mS = S, then mS = S + n. Also, it is well known × ∼ that if m ∈ Zn , then Cay(Zn,S) = Cay(Zn,mS) (cf. [6, Lem. 3.7.3, p. 48]). Therefore, we conclude that Cay(Zn,S) =∼ Cay(Zn,S + n), so Proposition 3.7 applies. (2) M. Muzychuk [14] found an eﬃcient method to check whether two circulant graphs (such as Cay(Zn,S) and Cay(Zn,S + n)) are isomorphic. Here is a family of examples that illustrate Proposition 3.7: Example 3.9. Let n = 2p2, where p is prime and p ≡ 1 (mod 4), and 2 choose c ∈ Z, such that c ≡−1 (mod p). Fix some a ∈ Zn of order p, and let S = Se ∪ So, where

Se = ±2+ hai ∪ {±a}⊆ 2Zn, ′ So = ±2+ hai ∪ {±ca}⊆ 2Zn, ′ So = n + So ⊆ 1 + 2Zn. Then

(1) Cay(Zn,S) =∼ Cay Zn,S + (n/2) , so Proposition 3.7 implies that Cay(Z ,S) is (nontrivially) unstable, but n (2) Cay(Zn,S) does not have a Wilson type.

Proof. (1) Choose a set R of coset representatives for hai in Zn, such that R + n = R, and deﬁne α: Zn → Zn by α(r + x)= r + cx for r ∈ R and x ∈ hai.

It suffices to show that if v, w ∈ Zn, such that v −w ∈ S, then α(v)−α(w) ∈ S + n. First, consider two vertices v = r + x and w = r + y that are in the same coset of hai. Then the definition of S implies that v − w = ±a, so y = x ± a, so ′ α(v) − α(w)= r + cx − r + c(x ± a) = ±ca ∈ So = So + n. Next, suppose v ∈ w + n + hai. Assume, without loss of generality, that v ∈ 1+2Zn and w ∈ 2Zn. Write v = r +n+x and w = r +y with r ∈ R and x,y ∈ hai. The definition of S implies that v − w = n ± ca, so y = x ± ca, so (using the fact that c2 ≡−1 (mod p)) we have 2 α(v) − α(w)= r + n + cx − r + c(x ± ca) = n ± c a = n ∓ a ∈ n + Se. We may now assume that vand w are in two different cosets of ha, ni. Then, from the definition of S, we see that every vertex in v +hai is adjacent 10 ADEMIR HUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

to every vertex in w + hai. Since α(v) ∈ v + hai and α(w) ∈ w + hai, it is therefore obvious that α(v) is adjacent to α(w). (2) The proof is by contradiction. ′ ′ Suppose, first, that Cay(Zn,S) has Wilson type (C.1), (C.2 ), or (C.3 ). Then Remark 3.6 tells us that the graph actually has Wilson type (C.1), so ∼ h + Se = Se for some non-zero h ∈ 2Zn. Since 2Zn = Zp2 , we know that |h| is divisible by p. Since Se is a union of cosets of hhi, this implies that |Se| is divisible by p, which contradicts the fact that |Se| = 2|a| +2=2p + 2. We may now assume that the graph is of Wilson type (C.4). Then we × can find m ∈ Z2p2 , such that mS + n = S. Since n is odd, this implies ′ see mSe + n = So, so mSe = So. By passing to the quotient group 2Zn/hai, we ′ note conclude that m ≡±1 (mod p). So ma = a∈ / So. This contradicts the fact A.7 ′ ! that mSe = So. It is shown in [4] that every nontrivially unstable circulant graph of valency ≤ 7 has a Wilson type, so the following examples have minimal valency among those that do not have a Wilson type: Example 3.10. Let n := 3 · 2ℓ, where ℓ ≥ 4 is even, and let n n S := ±3, ±6, ± , ± 3 . 12 2 n o Then the circulant graph X := Cay(Zn,S) has valency 8 and is nontrivially unstable, but does not have a Wilson type. see note Proof. It is easy to see that X is connected, nonbipartite, and twin-free. For A.8 ! convenience, let a = n/12. (unstable) Let ρ0 : 2Zn → Z2 be the homomorphism with kernel 4Zn. Then define ρ: Zn → Z2 by

ρ0(v) if v ∈ 2Z ; ρ(v)= n (ρ0(v + 1) otherwise. Finally, we let m := (n/6) − 1, and deﬁne α(v)= mv + ρ(v)n. ∼ We will show that α is an isomorphism from X to X = Cay(Zn,S + n), so Proposition 3.7 shows that X is unstable. Thus, given v ∈ Zn and s ∈ S, we wish to show that (3.11) α(v + s) − α(v) ∈ S + n.

see From the choice of m, we have m · 3 = n − 3, and a straightforward note calculation shows that m(n +3) = −3 (in Zn), so A.9 ! m{±3, n ± 3} + hni = {±3, n ± 3}⊆ S. Since it is clear from the deﬁnition of α that α(v + s) − α(v) ∈ ms + hni, this implies that (3.11) holds for s ∈ {±3, n ± 3}. AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 11

To deal with the remaining elements ±6 and ±a of S, ﬁrst note that ℓ−2 6 ∈/ 4Zn and a = 2 ∈ 4Zn (since ℓ ≥ 4), so ρ0(6) = 1 and ρ0(a) = 0. Hence, for all v ∈ Zn, we have ρ(v ± 6) = ρ(v)+1 and ρ(v ± a)= ρ(v). Therefore, we have α(v ± 6) − α(v)= m(v ± 6) + ρ(v ± 6)n − mv + ρ(v)n = ± 6m + ρ(v ± 6) − ρ(v) n = ±(n − 6) + n ≡±6+ n (mod n) ∈ S + n. Also note that 2ℓ−1 ≡ 8 (mod 12) (because ℓ − 1 ≥ 3 is odd), so we have m = 2ℓ−1 − 1 = 6k + 1, where k is odd. Therefore α(v ± a) − α(v)= m(v ± a)+ ρ(v ± a)n − mv + ρ(v)n = ±ma + ρ(v ± a) − ρ(v) n = ±(6k + 1)a + 0n = ±(kn + a) = ±a + n ∈ S + n. (no Wilson type) First, suppose that the graph has Wilson type (C.1), (C.2′), or (C.3′). By Proposition 3.4, this implies there are subgroups H and K of Zn that satisfy the conditions of Theorem 3.2(2). In particular, at least one coset of H is completely contained in S. We claim that H = hni. Let h = n/|H|, so h is a divisor of n, and H = hhi. Since S contains a coset of H, we know that at least one of any h consecutive elements x + 1,x + 2,...,x + h of Zn is an element of S. Since see (n − 3) − (n/12) > n/3, this easily implies h = n, which completes the proof note of the claim. A.10! Since 6 + n∈ / S and (n/12) + n∈ / S, we then see from the ﬁrst half of 3.2(2) that 6 ∈ Ko and n/12 ∈ Ko. This is impossible, because n/12 is divisible by 4, but 6 is not. Now suppose that the graph has Wilson type (C.4). This means there is × some m ∈ Zn , such that mS = S + n. Since n is even, this implies mSo = So + n, so (perhaps after replacing m with −m, we have m · 3 ∈ {3, n + 3}. But then m · 6+ n = 2(m · 3) + n ∈ 2{3, n + 3} + n = {6+ n}. Since 6 + n∈ / S, this is a contradiction.

Proposition 3.12. Assume X = Cay(Zn,S) is a circulant graph of even order. If there exist permutations α and β of 2Zn, and a subgroup H of 2Zn, such that: 12 ADEMIR HUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

(1) α 6= β,

(2) if the vertices u, v ∈ 2Zn are adjacent, then the vertices α(u) and β(v) are also adjacent, (3) s + H ⊆ S, for all odd s ∈ S, and

(4) α(v) − v ∈ H and β(v) − v ∈ H, for all v ∈ 2Zn, then X is unstable. ′ ′ Proof. Deﬁne permutations α and β of Zn by stipulating that α′(v)= β′(v)= v if v is odd, whereas α′(v)= α(v) and β′(v)= β(v) if v is even. ′ see It is straightforward to verify that if uv is an edge of X, then α (u) is adjacent note to β′(v), so Lemma 3.1 tells us that X is unstable. A.11! The following result is the special case where Cay(2Zn,S∩2Zn) has Wilson type (C.4).

Corollary 3.13. Assume X = Cay(Zn,S) is a circulant graph with n ≡ ∗ 0 (mod 4). If there exists m ∈ Zn, such that mSe + n = Se, and So + 2(m − 1)Zn = So + n = So, then X is unstable.

Proof. Deﬁne permutations α and β of 2Zn by α(x)= mx and β(x)= mx+ n. (The assumption that n ≡ 0 (mod 4) implies that n is even, so the image of β is in 2Zn.) Let H = h2(m − 1)Zn, ni, so So + H = So, by assumption. Then Proposition 3.12 applies, since α(v) − v = (m − 1)v ∈ 2(m − 1)Zn ⊆ H and β(v) − v = (m − 1)v + n ⊆ H. Remark 3.14. The statement of Theorem 1.4 includes corrected versions (C.2′) and (C.3′) of the original statements of Theorems C.2 and C.3 that appear in [20]. (C.2′) Y.-L. Qin, B. Xia, and S. Zhou [16] corrected the original hypothesis (C.2) by adding condition (b). (Note that complete graphs with more than two vertices are stable [16, Eg. 2.2], but satisfy condition (a) with h = 1, so this condition alone does not imply instability, even when n is divisible by 4.) (C.3′) The original statement of hypothesis (C.3) in [20, p. 376] includes the extraneous hypothesis that d> 1, and neglects to state the requirements that n/d is even and that either H * dZn or H ⊆ 2dZn. (Explanation: [20, Thm. 1] does not require the Cayley graph generated by the red edges to be disconnected, so there is no need to assume d > 1. The proof of [20, Thm. C.3] uses the assumption that r/d is odd to conclude that each component of Cay(Zn,R) is bipartite; this requires each r ∈ R to be an element of even order in Zn, which means that n/gcd(r, n) is even. Conditions (1) and (2) near the bottom of [20, p. 360] translate to the requirement that either H * dZn or H ⊆ 2dZn.) AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 13

It was mentioned above that if n > 2, then the complete graph Kn is stable [16, Eg. 2.2]. However, if n is even, then Kn satisﬁes the conditions in the original statement of (C.3) (with H = hn/2i, R = {n/2}, d = n/2, and r/d = 1 for the unique element r of R). This shows that the additional conditions in (C.3′) cannot be deleted.

4. The main lemma of [12] Assumption 4.1. For the proof of Corollary 4.7, it will be helpful to tem- porarily relax our standing assumption that all graphs are simple (see As- sumption 2.1). Namely graphs are allowed to have loops (but not multiple edges) in this section. The following elementary observation is stated only for automorphisms in [12], but the same proof applies to isomorphisms. + Lemma 4.2 (cf. [12, Lem. 2.2]). Let m ∈ Z , and let X1 = Cay(G1,S1) and X2 = Cay(G2,S2) be Cayley graphs, such that (1) G1 and G2 are abelian, and

(2) for j = 1, 2, we have ms 6= mt for all s,t ∈ Sj, such that s 6= t. If ϕ is any isomorphism from X1 to X2, then ϕ is also an isomorphism from Cay(G1,mS1) to Cay(G2,mS2), where mSj = { ms | s ∈ Sj }.

Proof (cf. proof of [12, Lem. 2.2]). Write m = p1p2 · · · pr, where each pi is prime, and let mi = p1p2 · · · pi for 0 ≤ i ≤ r. We will prove by induction on i that ϕ is an isomorphism from Cay(G1,miS1) to Cay(G2,miS2) The base case is true by assumption, since m0Sj = 1Sj = Sj. For v, w ∈ Gj, let #(v, w) be the number of walks of length pi from v to w in Cay(Gj,mi−1Sj). These walks are in one-to-one correspondence with the pi-tuples (s1,s2,...,spi ) of elements of mi−1Sj, such that

s1 + s2 + · · · + spi = w − v.

Since Gj is abelian, any cyclic rotation of (s1,s2,...,spi ) also corresponds to a walk from v to w. Therefore, the set of these walks can be partitioned into sets of cardinality pi, unless w = pis + v, for some s ∈ mi−1Sj, in which case there is a walk of the form v,s + v, 2s + v,...,pis + v = w. (Also note that s is unique, if it exists, by assumption (2).) Hence, we see that

#(v, w) 6≡ 0 (mod pi) ⇐⇒ v is adjacent to w in Cay(Gj,pimi−1Sj).

Since pimi−1 = mi, the desired conclusion that ϕ is an isomorphism from Cay(G1,miS1) to Cay(G2,miS2) now follows from the induction hypothesis that ϕ is an isomorphism from Cay(G1,mi−1S1) to Cay(G2,mi−1S2) (and the observation that isomorphisms preserve the value of the function #).

Corollary 4.3. Let X = Cay(Zn,S) be a circulant graph of even order, let ϕ be an automorphism of BX, and let S′ = { s′ ∈ S | s′ + (n/2) ∈/ S }. 14 ADEMIR HUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

′ Then ϕ is an automorphism of Cay(Zn × Z2, 2S × {0}). Proof. For every s ∈ S, we have { (t, 1) ∈ G × {1} | 2(t, 1) = 2(s, 1) } = {(s, 1), (s + n, 1)},

since n is the unique element of order 2 in Zn. Hence, the set { (t, 1) ∈ S × {1} | 2(t, 1) = 2(s, 1) } has cardinality 1 (which is odd) if s + n∈ / S, and has cardinality 2 (which is see even) otherwise. Therefore, the desired conclusion is obtained by applying note the proof of Lemma 4.2 with G1 = G2 = Zn × Z2 and m = 2, since A.12! 2(S′ × {1}) = 2S′ × {0}. Corollary 4.4 ([12, Rem. 3.1], [19, Thm. 23.9(a), p. 58]). Let ϕ be an automorphism of a Cayley graph Cay(G, S), and let m ∈ Z+. If G is abelian and gcd m, |G| = 1, then ϕ is an automorphism of Cay(G,mS).

Proof. Apply Lemma 4.2 with G1 = G2 = G and S1 = S2 = S. To verify hypothesis (2) of the lemma, note that the map x 7→ mx is a bijection on G, since gcd m, |G| = 1.

Corollary 4.5. Let α be an automorphism of BX, where X = Cay(Zn,S) is a circulant graph, let s,t ∈ S, and let k,ℓ ∈ Z+. Assume (1) α maps some s-edge to a t-edge, (2) ks ∈ S, (3) k ≡ ℓ (mod gcd |s|, |t| ), and (4) gcd k, |s| = gcd ℓ, |t| = 1. Then ℓt ∈ S. Proof. By assumption (3), there exists m ∈ Z+, such that m ≡ k (mod s) and m ≡ ℓ (mod |t|). Then assumption (4) implies that m is relatively prime to |s| and |t|, so, by Dirichlet’s Theorem on primes in arithmetic progres- sions, we may choose a prime p> 2n, such that pm ≡ 1 (mod lcm |s|, |t| ). This implies pks = s and pℓt = t. Since p is relatively prime to |Zn × Z2|, we know from Corollary 4.4 that every automorphism of BX is an automorphism of Cay Zn×Z2,p S×{1} . Since (s, 1) = p(ks, 1) ∈ p S × {1} , and α maps some s-edge to a t-edge, this implies that t ∈ pS. Since t = pℓt, and multiplication by p is a bijection see note on Zn × Z2, this implies ℓt ∈ S. A.13! Here are two interesting special cases:

Corollary 4.6. Let α be an automorphism of BX, where X = Cay(Zn,S) is a circulant graph, and let s,t ∈ S. If α maps some s-edge to a t-edge, and either gcd(|s|, |t| = 1, or S contains every element that generates hsi (e.g., if |s| ∈ {1, 2, 3, 4, 6}), then S contains every element that generates hti. AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 15

Proof. Let ℓt be a generator of hti, so gcd ℓ, |t| = 1. It suffices to find k ∈ Z+, such that ks ∈ S, k ≡ ℓ (mod gcd |s|, |t| ), and gcd k, |s| = 1, for then Corollary 4.5 tells us thatℓt ∈ S. If gcd(|s|, |t| = 1, we may let k = 1. Since gcd ℓ, |t| = 1, we know that ℓ is relatively prime to gcd |s|, |t| , so there is some k ∈ Z+, such that k ≡ ℓ (mod gcd |s|, |t| ) and gcd k, |s| = 1. (For example, we could take k to be a large prime.) If S contains every element that generates hsi, then ks ∈ S. Recall that the following corollary’s assumption that X is loopless will automatically be satisfied in all of the following sections of the paper (see Assumption 2.1).

Corollary 4.7. Let α be an automorphism of BX, where X = Cay(Zn,S) is a connected, nonbipartite, loopless, circulant graph, and let t ∈ Zn. If α(0, 1) = (t, 1), then S does not contain any generator of the subgroup hti. c c c Proof. Let S = G \ S be the complement of S, and let X = Cay(Zn,S ). Since BX is connected, it is easy to see that every automorphism of BX is see also an automorphism of BXc. note Note that 0 ∈ Sc, since X is assumed to be loopless. Therefore, we may A.14! let s = 0 (so |s| = 1) to conclude from Corollary 4.6 that Sc contains every generator of hti. The following result is stated only for automorphisms in [12], but essen- tially the same proof applies to isomorphisms.

Corollary 4.8 (cf. [12, Thm. 1.1]). Assume X1 = Cay(Zn,S1) and X2 = Cay(Zn,S2) are twin-free, connected, circulant graphs of odd order. If ϕ is any isomorphism from BX1 to BX2, such that

ϕ(Zn × {0})= Zn × {0}, then there is an isomorphism α: X1 → X2, such that ϕ(x, i)= α(x), i for all (x, i) ∈ BX1. Proof. This follows quite easily from Lemma 4.2 (with m = n + 1), by the see argument in the proof of [12, Thm. 1.1]. note A.15! 5. Unstable circulant graphs of order 2p Theorem 5.1. If p is a prime number, then every nontrivially unstable circulant graph of order 2p has Wilson type (C.4). The proof of this theorem will use several lemmas that are stated in greater generality than is needed here, because they may be useful for un- derstanding the unstable circulant graphs of other square-free orders. 16 ADEMIR HUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

Lemma 5.2. Let X = Cay(Zn,S) be a connected, nonbipartite, circulant graph, such that n ≡ 2 (mod 4). Then X has Wilson type (C.1) if and only if BX has an automorphism α, such that

(1) α∈ / Aut X × S2, and

(2) α ﬁxes 2Zn × Z2 (setwise). Proof. (⇒) This is the easy half of the proof (and does not require the see assumption that n ≡ 2 (mod 4)). If h + Se = Se, then we may deﬁne note α ∈ Aut(BX) by A.16! (x + h, i) if x ≡ i (mod 2), α(x, i)= ((x, i) if x 6≡ i (mod 2).

(⇐) Since α∈ / Aut X × S2, there is some v ∈ Zn, such that α(v, 1) 6= α(v, 0) + (0, 1). After conjugating α by a translation that moves (v, 0) to (0, 0), we may assume v = 0. This means that α(0, 0) = (0, 0) and α(0, 1) 6= (0, 1). Let Se = S ∩ 2Zn, and let Xe = Cay(2Zn,Se), so the subgraph of X induced by 2Zn × Z2 is BXe. Then assumption (2) implies that α restricts to an automorphism of BXe. ′ Now, let Xe be the connected component of Xe that contains 0. Since n ≡ 2 (mod 4), we know that n/2 is odd, so |hSei| is odd. This implies that ′ BXe is connected, and is therefore a connected component of BXe. Since ′ ′ BXe contains the ﬁxed point (0, 0), we conclude that BXe is α-invariant. ′ Therefore, α restricts to an automorphism of BXe. Since α ﬁxes (0, 0), then it follows from Theorem 1.10 that α(0, 1) = (0, 1). This is a contradiction.

Lemma 5.3 (Klin-Muzychuk, 1995, personal communication). Let

• X = Cay(Zn,S) be a circulant graph of order n, • G be an abelian group of order n, • e ∈ Z, such that eg = 0 for all g ∈ G, and • m ∈ Z, such that m ≡ 1 (mod e) and gcd(m,n) = 1. If X is isomorphic to some Cayley graph Cay(G, T ) on G, then S = mS. Proof (Klin-Muzychuk). Let ζ be a primitive nth root of unity, and, for is 1 ≤ i ≤ n, let λi = s∈S ζ , so λ1, λ2, . . . , λn are the eigenvalues of X [2, Prop. 3.5, p. 16]. Since gcd(m,n) = 1, there is a Galois automorphism α of the field extensionPQ[ζ]/Q, such that α(ζ) = ζm [8, Thm. 6.3.1, p. 278]. Since eg = 0 for all g ∈ G (and G is abelian), we know that each eigenvalue of Cay(G, T )isasumof eth roots of unity. (The range of any homomorphism χ: G → C× consists of eth roots of unity, so this is a consequence of the well-known formula in [1, Cor. 3.2] that generalizes the above formula for the eigenvalues of a circulant graph.) Since m ≡ 1 (mod e), this implies that every eigenvalue of Cay(G, T ) is fixed by α. So λ1, λ2, . . . , λn are fixed AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 17 by α. For 1 ≤ i ≤ n, this means

is is mis is ζ = λi = α(λi)= α ζ = ζ = ζ . ∈ ∈ ! ∈ ∈ Xs S Xs S Xs S sXmS ij Since the Vandermonde matrix [ζ ]1≤i,j≤n is invertible [18] (and therefore see has linearly independent rows), this implies that S = mS. note A.17! Lemma 5.4. Assume X = Cay(Zn,S) is a nontrivially unstable, circulant graph of even order, and there exists α ∈ Aut BX, such that α ﬁxes the two cosets of 2Zn × {0} that are in Zn × {0}, but interchanges the two cosets that are in Zn × {1}. If X does not have Wilson type (C.1), then: (1) n ≡ 2 (mod 4),

(2) hS ∩ 2Zni = 2Zn, and

(3) X =∼ Cay(Zn,S + (n/2)) (so Proposition 3.7 applies). Proof. Let

• Se = S ∩ 2Zn, • So = S \ Se,

• Ge = 2Zn × Z2, • Go = h2Zn ×{0}, (1, 1)i = h(1, 1)i = 2Zn ×{0} ∪ (2Zn +1)×{1} , • Xo = Cay(2Zn,So + n), • Be = BXe = Cay Ge,Se × {1} be the subgraph of BX induced by G , e • Bo = Cay Go,So × {1} be the subgraph of BX induced by Go, and • n = n/2 (see Notation 2.2). By assumption,

the restriction of α to Ge is an isomorphism from Be to Bo. (1) [The proof of this part of the lemma does not require the assumption that X does not have Wilson type (C.1).] Suppose n 6≡ 2 (mod 4). Then gcd(1 + n,n) = 1. Therefore, since Go is cyclic, and ng = 0 for all g ∈ Ge, we see from Lemma 5.3 that

So = (1+ n)So = So + n.

This means that Bo has twins. More precisely, since |n| = 2, each equiva- lence class of vertices with the same neighbors has even cardinality. So the same must be true in the isomorphic graph Be, which means there exists an element (h1, h2) of order 2 in Ge, such that

Se × {1} + (h1, h2)= Se × {1}.

Since the element of order 2 in the cyclic group 2Zn is unique (and the equation 1 + h2 = 1 implies that h2 = 0), we conclude that Se is invariant under translation by n. We already know that the same is true for So, so 18 ADEMIR HUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

we conclude that S + n = S. This contradicts the assumption that X is nontrivially unstable (and therefore twin-free). (2) Since n ≡ 2 (mod 4), we know that n is odd. So Zn = 2Zn ∪(2Zn +n). For convenience, label the 4 cosets of 2Zn × {0} by i i Ce = 2Zn × {i} and Co = (2Zn + n) × {i} for i ∈ Z2. 0 1 0 1 0 0 Then Ge = Ce ∪ Ce and Go = Ce ∪ Co . By assumption, α ﬁxes Ce and Co , 1 1 see but interchanges Ce and Co . note The function ϕ(x, i) = (x + in, i) is an isomorphism A.18! from Bo = Cay Go,So × {1} to BXo = Cay 2Zn × Z2, (So + n) × {1} . By composing ϕ with the restriction of α to Ge (that is, by restricting the map ϕ α(x, i) to Ge), we obtain an isomorphism from BXe to BXo. Since X does not have Wilson type (C.1), we know that Xe is twin-free (see Remark 2.6), so we see from Theorem 1.10 that each connected component of Xe is isomorphic to a connected component of Xo. Hence, these two connected components must have the same order, which means that the two subgroups hSei and hSo + ni of Zn have the same order, and are therefore equal. So hSe, ni ⊇ hSe ∪ Soi = hSi = Zn. Hence, hSei is a subgroup of index ≤ 2 in Zn, and must therefore be all of 2Zn. This establishes (2). (3) Let us begin by making a part of the above proof of (2) more concrete. It was established there that composing ϕ with the restriction of α to Ge is see an isomorphism from BXe to BXo. Since Xe is twin-free, we therefore see note from Corollary 4.8 that there is an isomorphism α0 : Xe → Xo, such that A.19! α(x, 1) = α0(x)+ n, 1 for x ∈ 2Zn. ′ see Similarly, if we let ϕ (x, i) = x + (1 − i)n, i , then restricting the map note α ϕ′(x, i) −(n, 0) to G yields an isomorphism from BX to BX . So there e o e A.20! is an isomorphism α1 : X → X , such that o e α(x, 1) = α1(x)+ n, 1 for x ∈ 2Zn. By comparing the two formulas for α(x, 1), we see that α0 = α1. ′ Now, deﬁne α : Zn → Zn by

α0(x) if x ∈ 2Z , α′(x)= n (α0(x + n)+ n if x∈ / 2Zn.

Since S = Se ∪ So (and therefore S + n = (So + n) ∪ (Se + n)), it will suffice to show that ′ α is an isomorphism from Cay(Zn,Se) to Cay(Zn,So + n) and ′ α is an isomorphism from Cay(Zn,So) to Cay(Zn,Se + n). see note The first is easy to see from the definition of α′, because A.21! AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 19

α0 is an isomorphism from Xe = Cay(2Zn,Se) to Xo = Cay(2Zn,So + n).

To establish the second, let s ∈ So. If x ∈ 2Zn, then ′ ′ α (x + s)= α0 (x + s)+ n + n (deﬁnition of α , since x + s∈ / 2Zn) α0 = α1 is an isomorphism ∈ α0(x)+ Se + n from Xo to Xe ′ ′ = α (x) + (Se + n). (deﬁnition of α , since x ∈ 2Zn)

Similarly, if x∈ / 2Zn, then deﬁnition of α′, α′(x + s)= α (x + s) 0 since x + s ∈ 2Z n = α0 (x + n) + (s + n) (n + n = 0) α = α is an isomorphism ∈ α (x + n)+ S 0 1 0 e from X to X o e = α0(x + n)+ n + Se + n (n + n = 0) deﬁnition of α′, = α ′(x)+ S + n . e since x∈ / 2Z n ′ ′ ′ Since α (x+s) ∈ α (x)+(Se +n) in both cases, we conclude that α is an isomorphism from Cay(Zn,So) to Cay(Zn,Se + n), as desired. This completes the proof of (3). The following result gathers the most important conclusions of Lem- mas 5.2 and 5.4.

Corollary 5.5. Let X = Cay(Zn,S) be a nontrivially unstable, circulant graph, such that n ≡ 2 (mod 4), and such that 2Zn × {0} is a block for the action of Aut BX. Then either X has Wilson type (C.1), or X is isomorphic to Cay Zn,S + (n/2) (so Proposition 3.7 applies). Proof. Since X is unstable, there exists α ∈ Aut BX, such that α∈ / Aut X × S2. By composing with a translation, we may assume that α fixes (0, 0). Since 2Zn × {0} is a block for the action of Aut BX, this implies that α fixes 2Zn × {0}. It must also fix the bipartition set Zn × {0}, so it fixes the difference Zn × {0} \ 2Zn × {0} = (2Zn + 1) × {0}. And then α either fixes the two remaining cosets, or interchanges them. In the first case, Lemma 5.2 tells us that X has Wilson type (C.1). In the second case, Lemma 5.4(3) provides the desired conclusion. The following result presents some useful special cases. Recall that the “Cayley Isomorphism Property” was defined in Definition 2.8. Corollary 5.6. Assume X is as in Corollary 5.5. Also assume that either (1) X has the Cayley Isomorphism Property, or

(2) Xe has the Cayley Isomorphism Property, or 20 ADEMIR HUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

(3) n is square-free, or

(4) the valency of Xe is ≤ 5. Then X has Wilson type (C.1) or (C.4). Proof. Assume X does not have Wilson type (C.1). Then Corollary 5.5 tells us that X =∼ Cay(Zn,S + n). (1) If X has the Cayley Isomorphism Property, this implies there is some × m ∈ Zn , such that S + n = mS, so X has Wilson type (C.4). (2) Assume that Xe has the Cayley Isomorphism Property. Let α0 and α1 be the isomorphisms in the proof of Lemma 5.4(3). Since Xe has the Cayley × Isomorphism Property, we have α0(x) = m ϕ(x), for some m ∈ Zn , and some ϕ ∈ Aut Xe. Then

mSe = m ϕ(Se) − ϕ(0) (ϕ ∈ Aut Xe) = α0(Se) − α0(0) (α0(x)= m ϕ(x)) =∼ = So + n (α0 : Xe → Xo). × Now, since ϕ ∈ Aut Xe = Aut Cay(2Zn,Se) and m ∈ Zn , Corollary 4.4 implies

ϕ ∈ Aut Cay(2Zn,mSe) = Aut Cay(2Zn,So + n) = Aut Xo. Therefore

m(So + n)= m ϕ(So + n) − ϕ(0) (ϕ ∈ Aut Xo) = α0(So + n) − α0(0) (α0(x)= m ϕ(x)) α = α is an isomorphism = S , 0 1 e from X to X o e so mSo = Se + n. Therefore

mS = m(Se ∪ So)= mSe ∪ mSo = (So + n) ∪ (Se + n)= S + n. So X has Wilson type (C.4). (3) The order of X is square-free, so Theorem 2.9 tells us that X has the Cayley Isomorphism Property. Therefore (1) applies. (4) It is known [10, §7.2] that every connected, circulant graph of valency ≤ 5 has the Cayley Isomorphism Property. (A proof for valency 4 can also be found in [15, Thm. 5.4].) Therefore (2) applies.

Proof of Theorem 5.1. Let X = Cay(Z2p,S) be a nontrivially unstable, circulant graph of order n = 2p. It is easy to see, by inspection, that there are no nontrivially unstable, circulant graphs of order 4, so p is odd. Therefore n = 2p is square-free. Let S′ = S \ (S + p). Since X is twin-free, we know that S +p 6= S (see Remark 2.6), which means that 2S′ is nonempty. ′ Case 1. Assume 2S 6= {0}. Since 2Z2p has order p, which is prime, AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 21

′ every nonzero element is a generator. So 2S generates 2Z2p. We also know from Corollary 4.3 that every automorphism of BX is an automorphism of ′ Cay Z2p × Z2, 2S × {0} . By combining these two facts, we conclude that 2Z 2p × {0} is a block for the action of Aut BX. Therefore, Corollary 5.6(3) applies, so X has Wilson type (C.1) or (C.4). However, 2Z2p =∼ Zp has no nontrivial, proper subgroups, so it is obvious that X does not have Wilson type (C.1). Therefore, it must have Wilson type (C.4), which is exactly what we needed to prove. Case 2. Assume 2S′ = {0}. This means that S′ = {p}, so p ∈ S. Since X is unstable, we may let α be an automorphism of BX, such that α(0, 1) = (t, 1) with t 6= 0. For all x ∈ Zn, we see from Corollary 4.7 that if |x| = |t|, then x∈ / S. Since p ∈ S, this tell us that |t|= 6 2. So |t| is either p or 2p. Therefore, either S does not contain any element of order p, or S does not contain any element of order 2p. However, since 2S′ = {0}, we also know that s + p ∈ S for all s ∈ S \ {p}. Also note that |s| = p ⇐⇒ |s + p| = 2p. Putting this together, we conclude that S = {p}. This contradicts the fact that the nontrivially unstable graph X must be connected.

Corollary 5.7. Let X = Cay(Z2p,S) be a circulant graph of order 2p, where p is an odd prime, and let Se = S ∩ 2Z2p. The graph X is unstable if and × only if either it is trivially unstable, or there exists m ∈ Z2p, such that 2 m Se = Se, mSe 6= Se, and S = Se ∪ (n/2) + mSe . Proof. (⇐) If X is not trivially unstable, the conditions imply that X has Wilson type (C.4). (⇒) Assume X is nontrivially unstable. We conclude from Theorem 5.1 that X has Wilson type (C.4), so there is some m ∈ Z2p, such that S = mS + p. Since p is odd, this implies that So = mSe + p (where So = S \ Se) and 2 Se = mSo + p = m(mSe + p)= m Se. If mSe = Se, then So = Se + p, so S = S + p, which contradicts the fact that nontrivially unstable graphs are twin-free. Corollary 5.8. For n ∈ Z+, there does not exist a nontrivially unstable circulant graph of order n if and only if either n is odd, or n< 8, or n = 2p, for some prime number p ≡ 3 (mod 4). ∼ Proof. (⇒) If n is not prime, then 2Zn = Zn/2 has a nontrivial, proper subgroup A. Choose some b ∈ 2Zn \ A, and let S = {±1}∪ (±b + A), so Se := S ∩ 2Zn = ±b + A. Then X = Cay(Zn,S) has Wilson type (C.1), so it is unstable. If n is prime, and n 6≡ 3 (mod 4) (and n ≥ 8), then n ≡ 1 (mod 4), × 2 so there exists m ∈ Zn , such that m = −1. Let S = {±1, n ± m}, so 22 ADEMIR HUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

Se := S ∩ 2Zn = {±m + n} and S = mS + n. Then X = Cay(Zn,S) has Wilson type (C.4), so it is unstable. In either case, X is also connected (because 1 ∈ S) and nonbipartite (because Se 6= ∅). Hence, if X is not nontrivially unstable, then it must not be twin-free, so there is a nonzero h ∈ Zn, such that h+S = S. Note that in both cases, So = {±1} and since n> 4 and h is nonzero, it cannot happen that {±1} + h = {±1}. It follows that So + h = Se and Se + h = So (and h is odd). Since So + 2h = So (and S0 = {±1}), we must have 2h = 0, which means h = n, so Se = So + n = {n ± 1}. If n is not prime, then, since {n ± 1} = Se = ±b + A, we must have h2i ⊆ A. Since n > 4, this implies |b + A| = |A| ≥ n/2 > 2, which is a contradiction. If n is prime, we must have m = ±1 (since Se = {n±m}, which contradicts the fact that m2 = −1. (⇐) We prove the contrapositive: supposing there does exist a nontrivially unstable circulant graph of order n, we will show that n is odd, that n ≥ 8, and that n/2 is not a prime number that is congruent to 3 (mod 4). The fact that n is odd is immediate from Theorem 1.10. Also, it is easy to see, by inspection, that there are no nontrivially unstable circulant graphs of order 2 or 4; so n ≥ 6. Now suppose X = Cay(Z2p,S) is a nontrivially unstable circulant graph of order 2p, where p is prime, and p ≡ 3 (mod 4). (This includes the case where n = 6.) We will show that this leads to a contradiction. By Theorem 5.1, we know that X has Wilson type (C.4), so there is some × m ∈ Z2p, such that S = mS +n. Write m = mom2, where mo has odd order × (as an element of the group Z2p), and the order of m2 is a power of 2. Since × Z2p is cyclic of order p − 1 ≡ 2 (mod 4), there are no elements of order 4 in × Z2p, so m2 ∈ {±1}. Since S = −S, this implies S = m2S, so we conclude that S = moS + n. After repeatedly multiplying both sides of this equation k by mo, we see that S = mo S + n for any odd number k, including k = |mo|. Hence, we have S = S +n. This contradicts the fact that X is twin-free.

6. Computational results S. Wilson [20, p. 377] mentioned: “There are 3274 circulant graphs which are non-trivially unstable and have no more than 38 vertices.” However, we performed computations that produce a diﬀerent number: there seem to be 3576 such graphs (up to isomorphism). The interested reader can reproduce our results in only a few minutes by running the Sagemath1 code or Magma code or Maple code that is available in the ancillary ﬁles directory of this paper on the arxiv. Analysis of these graphs establishes:

1Sagemath code can be run online at https://cocalc.com/ AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 23

Observation 6.1. Every nontrivially unstable circulant graph of order less than 40 has a Wilson type, except the following six graphs (up to isomorphism), all of order 24: (1) Cay(Z24, {±2, ±3, ±8, ±9, ±10}), (2) Cay(Z24, {±2, ±3, ±8, ±9, ±10, 12}), (3) Cay(Z24, {±1, ±2, ±5, ±7, ±8, ±10, ±11}), (4) Cay(Z24, {±1, ±2, ±5, ±7, ±8, ±10, ±11, 12}), (5) Cay(Z24, {±1, ±2, ±3, ±5, ±7, ±8, ±9, ±10, ±11}), (6) Cay(Z24, {±1, ±2, ±3, ±5, ±7, ±8, ±9, ±10, ±11, 12}). The first of these graphs appears explicitly in [16, p. 156], and it seems that some (or all) of the others were also known to the authors of that paper. Remark 6.2. Several additional days of computer time extended the ex- haustive calculations to order 50. (However, these calculations are certainly not definitive, because they were performed only once, so their correctness has not been verified.) Lists of the additional 67725 nontrivially unstable graphs have been deposited in the ancillary files directory of this paper on the arxiv. We note that the instability of every example that was found is explained by the instability conditions in Section 3. On the other hand, the calculations uncovered 316 additional examples of nontrivially unstable circulant graphs with no Wilson type: 52 of order 40, 262 of order 48, and 2 of order 50.

References [1] L. Babai, Spectra of Cayley graphs. J. Combin. The- ory Ser. B 27 (1979), no. 2, 180–189. MR 0546860, doi:10.1016/0095-8956(79)90079-0 [2] N. Biggs, Algebraic graph theory. Cambridge University Press, London, 1974. MR 0347649, doi:10.1017/CBO9780511608704 [3] B. Fernandez and A. Hujdurović, Canonical double covers of circulants (preprint, 2020). https://arxiv.org/abs/2006.12826 [4] A. Hujdurović, ¯D. Mitrović, and D. W. Morris, Automorphisms of the double cover of a circulant graph of valency at most 7 (preprint, 2021). https://arxiv.org/abs/2108.05164 [5] J. D. Dixon and B. Mortimer, Permutation Groups. Springer, New York, 1996. MR 1409812, doi:10.1007/978-1-4612-0731-3 [6] C. Godsil and G. Royle, Algebraic Graph Theory. Springer, New York, 2001. MR 1829620, doi:10.1007/978-1-4613-0163-9 [7] A. Kotlov and L. Lovász, The rank and size of graphs. J. Graph Theory 23 (1996), no. 2, 185–189. MR 1408346, doi:10.1002/(SICI)1097-0118(199610)23:2<185::AID-JGT9>3.0.CO;2-P 24 ADEMIR HUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

[8] S. Lang, Algebra. Springer-Verlag, New York, 2002. MR 1878556, doi:10.1007/978-1-4613-0041-0

[9] J. Lauri, R. Mizzi, R. Scapellato, Unstable graphs: a fresh outlook via TF-automorphisms. Ars Math. Contemp. 8 (2015), no. 1, 115–131. MR 3281124, doi:10.26493/1855-3974.534.934

[10] C. H. Li, On isomorphisms of ﬁnite Cayley graphs—a survey. Discrete Math. 256 (2002), no. 1-2, 301–334. MR 1927074, doi:10.1016/S0012-365X(01)00438-1

[11] D. Maruˇsiˇc, R. Scapellato, and N. Zagaglia Salvi, A characterization of particular symmetric (0,1) matrices. Linear Algebra Appl. 119 (1989), 153–162. MR 1005241, doi:10.1016/0024-3795(89)90075-X

[12] D. W. Morris, On automorphisms of direct products of Cayley graphs on abelian groups. Electronic J. Combin. 28(3) (2021), #P3.5. doi:10.37236/9940

[13] M. Muzychuk, On Ad´am’s´ conjecture for circulant graphs (corrigen- dum). Discrete Math. 176 (1997), no. 1-3, 285–298. MR 1477298, doi:10.1016/S0012-365X(97)81804-3

[14] M. Muzychuk, A solution of the isomorphism problem for circulant graphs. Proc. London Math. Soc. (3) 88 (2004), no. 1, 1–41. MR 2018956, doi:10.1112/S0024611503014412

[15] M. Muzychuk, M. Klin, and R. P¨oschel, The isomorphism problem for circulant graphs via Schur ring theory, in: A. Barg and S. Litsyn, eds., Codes and association schemes (Piscataway, NJ, 1999), pp. 241– 264. Amer. Math. Soc., Providence, RI, 2001. MR 1816402, doi:10.1090/dimacs/056

[16] Y.-L. Qin, B. Xia, and S. Zhou, Stability of circulant graphs, J. Combin. Theory Ser. B 136 (2019) 154–169. MR 3926283, doi:10.1016/j.jctb.2018.10.004

[17] Wikipedia, Bipartite double cover. https://en.wikipedia.org/wiki/Bipartite_double_cover

[18] Wikipedia, Vandermonde matrix. https://en.wikipedia.org/wiki/Vandermonde_matrix

[19] H. Wielandt, Finite Permutation Groups. Academic Press, New York, 1964. MR 0183775, doi:10.1016/C2013-0-11702-3

[20] S. Wilson, Unexpected symmetries in unstable graphs. J. Com- bin. Theory Ser. B 98 (2008), no. 2, 359–383. MR 2389604, doi:10.1016/j.jctb.2007.08.001 AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 25

University of Primorska, UP IAM, Muzejski trg 2, 6000 Koper, Slovenia and University of Primorska, UP FAMNIT, Glagoljaˇska 8, 6000 Koper, Slovenia Email address: [email protected]

University of Primorska, UP FAMNIT, Glagoljaˇska 8, 6000 Koper, Slovenia Email address: [email protected]

Department of Mathematics and Computer Science, University of Leth- bridge, Lethbridge, Alberta, T1K 3M4, Canada Email address: [email protected] 26 ADEMIR HUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

Appendix A. Notes to aid the referee

A.1. We prove the contrapositive. (⇒) Suppose there exists h ∈ Zn, such that h + S = S. We claim that 0 and −h have exactly the same neighbors, so Cay(Zn,S) is not twin-free: v is a neighbor of 0 ⇐⇒ v ∈ S ⇐⇒ h + v ∈ S ⇐⇒ v − (−h) ∈ S ⇐⇒ v is adjacent to −h.

(⇐) Suppose v and w have exactly the same neighbors in Cay(Zn,S), and let h = v − w (so w = v − h). We claim that h + S = S: S = { x − w | x is a neighbor of w } = { x − w | x is a neighbor of v } = { x − (v − h) | x is a neighbor of v } = h + { x − v | x is a neighbor of v } = h + S.

A.2. Let Gˆ be the regular representation of Zn × Z2. Since Gˆ is contained in Aut BX, it is clear that B is also a block for the action of Gˆ. Hence, [5, Exer. 1.5.6, p. 13] tells us that B is an orbit of some subgroup Hˆ of Gˆ. If we let H be the corresponding subgroup of Zn × Z2, then the Hˆ orbit of an element b of Zn × Z2 is precisely b + H, so it is a coset of H. We see from [5, Exer. 1.5.3, p. 13] that every coset of H is a block. Furthermore, since these cosets form a partition of Zn × Z2, we see that they are the entire “system of blocks” containing B, so [5, Exer. 1.5.3, p. 12] tells us that Aut BX acts on this set of blocks.

A.3. Since H = hhi and |h| = 2, it is clear that |H| = 2. If |K| is divisible by 4, then (since K is a subgroup of Zn, and is therefore cyclic) we know that K has an element k of order 4. So |2k| = 2. Since the element of order 2 in Zn is unique, we conclude that h = 2k ∈ 2K. This contradicts the fact that h ∈ Ko.

A.4. Since h ∈ 2K, we may write h = 2k, for some k ∈ K. If |h| = 2, this implies |k| = 4, so |K| must be divisible by 4. AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 27

A.5. If |H| = 2, then H = hni. Also, since |K| is even, we know that n ∈ K. So n ∈ H ∩ K ⊆ 2K, which means there is some k ∈ K, such that 2k = n. Then |k| = 4, so |K| is divisible by 4.

A.6. Suppose uv is an edge of X. Since α is an isomorphism from Cay(Zn,S) to Cay(Zn,S+n), this implies that α(u) is adjacent to α(v) in Cay(Zn,S+n), which means α(u) − α(v) ∈ S + n. So

α(u) − β(v)= α(u) − α(v)+ n = α(u) − α(v) + n ∈ (S + n)+ n = S.

∼ ∼ A.7. Since 2Zn/hai = Zp2 /pZp2 = Zp, there is a homomorphism ϕ: 2Zn → ′ Zp, such that ϕ(2) = 1 and ker ϕ = hai. Then ϕ(Se)= {±1, 0} and ϕ(So)= {±1, 0}. Therefore, in Zp, we have ′ {±m, 0} = m · {±1, 0} = m · ϕ(Se)= ϕ(mSe)= ϕ(So)= {±1, 0}, so m ≡±1 (mod p).

A.8. Let a = 2ℓ−2 = n/12 ∈ S. • Since gcd(3, a) = 1, we know that X is connected. • Since 3 is odd and a is even (because ℓ> 2), we know that X is not bipartite. • Suppose X is not twin-free. Then Remark 2.6 provides some nonzero h ∈ Zn, such that h + S = S. Since |S| = 8, this implies that |h| is a divisor of 8, so |h| is even, which implies that n ∈ hhi. Therefore, we have n + S = S. However, the only even elements of S are ±6 and ±n/12, so it is clear that n + 6 ∈/ S (because ℓ > 3). This is a contradiction. 28 ADEMIR HUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

A.9. We have n n m(n +3) = − 1 + 3 6 2 = (2 ℓ−1 −1)(3 · 2ℓ−1 + 3) = 3(2ℓ−1 − 1)(2ℓ−1 + 1) = 3(22ℓ−2 − 1) = 3 · 22ℓ−2 − 3 = 2ℓ−2 n − 3 ≡−3 (mod n).

A.10. Since n ≥ 3 · 24 = 48 > 36, have n 6n − 36 − n 4n n (n − 3) − = > = . 12 12 12 3 If ℓ ≥ 5, then n/12 ≥ (3 · 25)/12 = 8 > 6, so S does not contain any elements between n/12 and n−3. Therefore, the preceding inequality implies that h>n/3. Since h is a divisor of n, we conclude that h = n. A bit of additional argument is required in the special case where ℓ = 4 (because n/12 = 4 < 6). However, in this case, we have (n − 3) − 6 = 21 − 6 = 15. This is not a divisor of 48 = n, so this diﬀerence is not equal to h. Therefore, 6 and n − 3 cannot be two consecutive elements of a coset of H, so we once again must have h ≥ (n − 3) − (n/12) > n/3.

A.11. If u and v are odd, then α′(u)= u and β′(v)= v, so it is obvious that α′(u) is adjacent to β′(v). Also, if u and v are even, then it is immediate from (2) that α′(u) is adjacent to β′(v). So we may assume that s := u − v is odd. Then we see from (3) that s + H ⊆ S. From (4) (and the fact that α′(x) = β′(x) = x ∈ x + H when x∈ / 2Zn), we also know that ′ ′ α (x) ∈ x + H and β (x) ∈ x + H, for all x ∈ Zn. Therefore α′(u) − β′(v) ∈ (u + H) − (v + H) = (u − v)+ H = s + H ⊆ S, so α′(u) is adjacent to β′(v). AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 29

A.12. For v, w ∈ V (BX), let #(v, w) be the number of walks of length 2 from v to w in BX. These walks are in one-to-one correspondence with the elements of

Wv,w := { (s1,s2) ∈ S × S | s1 + s2 = w − v }, so b b #(v, w)= |Wv,w|.

Since Zn × Z2 is abelian, we have s2 + s1 = s1 + s2, so the map π : (s1,s2) 7→ (s2,s1) is a permutation of W. Let b b b b Fv,w := { (s,s) | s ∈ S, 2s = w − v } be the set of ﬁxed points of π in its action on Wv,w. Since the cardinality of every orbit of π is either 1 or 2, we see thatb

|Wv,w| ≡ |Fv,w| (mod 2).

We claim that |Fv,w| is either 0, 1, or 2. (So |Fv,w| is odd if and only if |Fv,w| = 1.) To see this, suppose (s,s) and (t,t) are two diﬀerent elements of Fv,w. This means that s 6= t and 2s = w − v = 2t, so 2s = 2t. Then t − s must be an element of order 2, so t = s + n (because n is the only element of order 2 in Zn). This completes theb proof of theb claim. Furthermore, the argument establishes that

′ |Fv,w| = 1 ⇐⇒ w − v ∈ 2S . So e ′ |Fv,w| is odd ⇐⇒ w − v ∈ 2S . Combining the above facts implies that e #(v, w) is odd ⇐⇒ w − v ∈ 2S′. Since every automorphism of BX must preserve the value of #(v, w), and e 2S′ = 2S′ × {0}, this implies that every automorphism of BX is an auto- ′ morphism of Cay Zn × Z2, 2S × {0} . e

A.13. Since t ∈ pS, we have t = ps, for some s ∈ S. However, we also have t = pℓt. So pℓt = ps. Since multiplication by p is a bijection on Zn × Z2, this implies ℓt = s ∈ S. 30 ADEMIR HUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

A.14. For all v, w ∈ Zn, we have: (v, 0) is adjacent to (w, 1) in BX X = Cay(G, S), and ⇐⇒ v − w ∈ S definition of BX) ⇐⇒ v − w∈ / Sc (definition of Xc) (v, 0) is not adjacent Xc = Cay(G, Sc), and ⇐⇒ . to (w, 1) in BXc definition of BXc) Therefore, if α is an automorphism of BX that fixes Zn × {0} and Zn × {1} (setwise), then α is also an automorphism of BXc. Now, let α be any automorphism of BX. Since the bipartition of BX is unique (because BX is connected), we know that α either fixes the two bipartition sets (setwise) or interchanges them. If it fixes the sets, then the preceding paragraph implies that α is an automorphism of BXc, as desired. Now, suppose α interchanges the two bipartition sets. Let α′ be the composition of α with the translation by (0, 1). Then α′ is an automorphism of BX that fixes the two bipartition sets, so it is an automorphism of BXc. Since α differs from α′ only by a translation (which is an automorphism of BXc), we conclude that α is an automorphism of BXc.

A.15. Since ϕ(Zn × {0})= Zn × {0}, there is a bijection α: Zn → Zn, such that ϕ(x, 0) = α(x), 0 for all x ∈ Zn. Let m = n +1. Then m is even (because n is odd), so mS1 = mS1 × {0}. Also, for all x ∈ Zn we have nx = 0, so mx = x; therefore mS1 = S1. Then, since m = n+1 is obviously relatively prime to n, Corollary 4.4e tells us that ϕ is an isomorphism from Cay(Zn ×Z2,S1 ×{0}) to Cay(Zn ×Z2,S2 ×{0}). By the definition of α, this implies that α is an isomorphism from Cay(Zn,S1) to Cay(Zn,S2). We already know (from the definition of α) that ϕ(x, 0) = α(x), 0 for all x ∈ Z . The fact that X2 is twin-free now implies that we also have n ϕ(x, 1) = α(x), 1 . To see this, let N1 be the set of neighbors of x in X1, so N2 := α(N1) is the set of neighbors of α(x) in X2 (since α is an isomorphism). Then the set of neighbors of (x, 1) is N1 × {0} (by the definition of BX1), so, since ϕ is an isomorphism, the set of neighbors of ϕ(x, 1) in BX2 is

ϕ N1 × {0} = α(N1) × {0} = N2 × {0}.

Thus, if we write ϕ(x, 1) = (y, 1), then N2 is the set of neighbors of y in X2. Since this is the same as the set of neighbors of α(x), and X2 is twin-free, we conclude that y = α(x). So ϕ(x, 1) = (y, 1) = α(x), 1 , as claimed. AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 31

A.16. First note that h must be even, since h+Se = Se (and Se is nonempty, since X is not bipartite). Therefore, α ﬁxes the two sets {x ≡ i (mod 2)} and {x 6≡ i (mod 2)}. Since α is obviously a bijection on each of these sets, we conclude that α is a bijection. Now, suppose (x, 0) is adjacent to (y, 1), and let s = y − x ∈ S. • If x is odd and y is even, then α ﬁxes (x, 0) and (y, 1), so it is obvious that α(x, 0) is adjacent to α(y, 1). • If x is even and y is odd, then α translates (x, 0) and (y, 1) by (h, 0). Since this translation is an automorphism of BX, it is again obvious that α(x, 0) is adjacent to α(y, 1).

• Finally, suppose x and y have the same parity, which means s ∈ Se. Then α ﬁxes one of the two vertices, and translates the other by (h, 0), so

α(y, 1) − α(x, 0) = (y, 1) − (x, 0) ± (h, 0) ∈ Se × {1} ± (h, 0) = Se ± h × {1}= Se × {1}⊆ S × {1}, so α(x, 0) is adjacent to α(y, 1).

ij A.17. Let A = [ζ ]0≤i,j≤n−1 be the Vandermonde matrix. Also let n−1 n v = [v0, v1, . . . , vn−1] = [vi]i=0 ∈ Z and n−1 n w = [w0, w1, . . . , wn−1] = [wi]i=0 ∈ Z , where 1 if i ∈ S, vi = (0 otherwise, and 1 if i ∈ mS, wi = (0 otherwise. Then n−1 n−1 vA = ζis ∈ Cn and wA = ζis ∈ Cn. " ∈ # " ∈ # Xs S i=0 sXmS i=0 is is Since s∈S ζ = s∈mS ζ for all i, we conclude that vA = wA. Since A is invertible, this implies v = w, which means S = mS. P P 32 ADEMIR HUJDUROVIC,´ ¯DOR¯DE MITROVIC,´ AND DAVE WITTE MORRIS

0 A.18. It is obvious from the deﬁnition that ϕ ﬁxes Ce . (Indeed, it acts as 1 1 the identity on this set.) Since n is odd, we also see that ϕ maps Co to Ce . 0 1 0 1 Hence, ϕ maps Go = Ce ∪ Co to Ce ∪ Ce = Zn × Z2 (and is a bijection). Now, suppose (x, 0) is adjacent to (y, 1) in Bo. This means that y−x ∈ So. Then

ϕ(y, 1) − ϕ(x, 0) = (y + n, 1) − (x, 0) = (y − x)+ n, 1 ∈ (So + n) × {1}, so ϕ(x, 0) is adjacent to ϕ(y, 1) in Cay 2Zn × Z2, (So + n) × {1} , as desired.

A.19. Corollary 4.8 provides an isomorphism α0 : Xe → Xo, such that

ϕ α(x, 1) = α0(x), 1 .

From the deﬁnition of ϕ, we conclude that α(x,1) = α0(x)+ n, 1 .

′ 0 0 1 1 A.20. Let β(x, i) = α ϕ (x, i) − (n, 0), so β(Ce ) = Ce and β(Ce ) = Ce , which means that β is a bijection on G . Note that e β(y, 1) − β(x, 0) = α ϕ′(y, 1) − α ϕ′(x, 0) = α(y, 1) − α(x + n, 0).

Suppose (x, 0) is adjacent to (y, 1) in BXo. Then y − x ∈ So + n, so y − (x + n) ∈ So ⊆ S. Since α ∈ Aut BX, then α(y, 1) − α(x + n, 0) ∈ S × {1}.

1 0 Furthermore, since (y, 1) ∈ Ce and (x + n, 0) ∈ Co , we know that α(y, 1) ∈ 1 0 Co and α(x + n, 0) ∈ Co , so β(y, 1) − β(x, 0) = α(y, 1) − α(x + n, 0) 1 0 ∈ S × {1} ∩ (Co − Co ) 1 = S × {1} ∩ Ce = S e × {1}.

So β is an isomorphism from BXo to BXe. AUTOMORPHISMS OF THE DOUBLE COVER OF A CIRCULANT 33

A.21. Suppose v − w ∈ Se. Then v and w have the same parity, so the calculations of α′(v) and α′(w) both use the same formula, so either ′ ′ α (v) − α (w)= α0(v) − α0(w) ∈ So + n or ′ ′ α (v) − α (w)= α0(v + n)+ n − α0(w + n)+ n = α 0(v + n) − α0(w + n) ∈ So + n

(because (v + n) − (w + n)= v − w ∈ Se).