On the Occurrence of the Complete Graph K
5in the Hasse Graph of a Finite Group.
ROLANDSCHMIDT(*)
Dedicated to Professor Guido Zappa on the occasion of his 90th birthday
1. Introduction.
In [7]we determined all finite groups with planar subgroup lattices and also those with planar Hasse graphs. Here a finite latticeLis called planar (see [4]) if it is possible to draw its Hasse diagram in the plane in such a way that no two line segments intersect; the Hasse graphLofLis its Hasse diagram considered as an undirected graph in the usual way (see §2) and for a finite groupG, the Hasse graphL(G)of its subgroup latticeL(G) is also called the Hasse graph ofG.
By Kuratowski's theorem, a finite graph is planar if and only if it does not contain a subdivision of the complete graph K5 or the complete bipartite graphK3;3as a subgraph (see [2, p. 24]). Figure 1 shows the graphsK5and K3;3; a subdivision is obtained from a graph by subdividing some of the edges,
(*) Indirizzo dell'A.: Mathematisches Seminar, Christian-Albrechts-UniversitaÈt zu Kiel, Ludewig-Meyn-Str. 4, 24098 Kiel Germany; e-mail: [email protected]
Figure 1
that is, by replacing the edges by paths having at most their endvertices in common. We were able to show in [7]that the Hasse graph of a finite group is planar if and only if it contains no subdivision ofK3;3as a subgraph.
The aim of the present paper is to determine all finite groups whose Hasse graph contains no subdivision ofK5as a subgraph; we shall call such a graphK5-free, for short. Forp-groups, we obtain the following results.
THEOREMA. LetjGj pn where2<p2Pand5n2N. Then L(G) is K5-free if and only if G is metacyclic of exponent at least pn 2.
THEOREMB. Let jGj 2n where 6n2N. Then L(G) is K5-free if and only if G is lattice-isomorphic to an abelian group of type (2n), (2n 1;2), or(2n 2;4), or G is one of the followingtwo groups
Gha;bja4b2n21;aba 1i;
Gha;bja81;b2n 3a4;aba 1i:
The assumptionn5 is needed in Theorem A since for every odd primep, there are two groups of orderp4withK5-free Hasse graph which are not metacyclic. Theorem B holds forn5 if one includes in the list the qua- ternion groupQ32of order 25and all metacyclic groups of order at most 24 (see Remarks 3.8 and 3.11).
It might be interesting to note that only rather few of theseK5-freep- groups have planar (orK3;3-free) Hasse graphs. In fact, we showed in [7]
that forp>2, a group of orderpnhas this property if and only if it is either cyclic or lattice-isomorphic to an abelian group of type (pn 1;p). Forp2, the same groups and, in addition, only the dihedral group of order 8, the quaternion groupsQ8andQ16, and the semidihedral group of order 16 have planar Hasse graphs. This latter group is the uniquep-group with planar Hasse graph and non-planar subgroup lattice.
In [7, Lemma 3.2]we showed that the Hasse graph of a cyclic group Ck of order kis K5-free if and only if it is planar, and that Ck has this property if and only ifkis of the formpn;pnqm, orpnqrwheren;m2N and p;q;rare primes; here the groups Cpn and Cpnqm even have planar subgroup lattices whereasL(Cpqr) is non-planar. Therefore the following two theorems complete the determination of all finite groups with K5- free Hasse graph.
THEOREMC. Let G be a finite nilpotent group which is not cyclic and not of prime power order. Then L(G)is K5-free if and only if G'CpQ8
where2<p2Por G'CpCqCqwith different primes p and q.
THEOREMD. Let G be a finite group which is not nilpotent. Then L(G) is K5-free if and only if GPK where P/G and K is cyclic, and one of the followingholds(where p;q;r are pairwise different primes and n2N).
(a) jPj p;jKj qnandjK:CK(P)j q3. (b) jPj p andjKj qr.
(c) jPj p2, P is cyclic,jKj qnandjK:CK(P)j q.
(d) jPj p2, P is elementary abelian, jKj qn; jK:CK(P)j q3, jCK(P)j q, andV(K=CK(P))operates irreducibly on P.
(e) jPj p2, P is elementary abelian,jKj qr and both minimal subgroups of K operate irreducibly on P.
(f) jPj p3, P is nonabelian of exponent p,jKj q and K operates irreducibly on P=F(P).
(g) P'Q8; jKj 3or9, G=CK(P)'SL(2;3).
(h) jPj p4; P is abelian of type (p2;p2); jKj q and K operates irreducibly on P=F(P).
Again, by [7], only few of these groups have planar Hasse graphs, namely those in (a) of Theorem D for whichjK:CK(P)j qand those in (d) for which jKj q. These groups, in fact, also have planar subgroup lattices whereas for all other groups in Theorems C and D, the Hasse graph is non-planar.
2. Notation and Preliminary Remarks.
In the whole paper,Gis a finite group andp;qare primes. We denote by L(G) the subgroup lattice ofG; soL(G) fXjXGgwith the set-theo- retical inclusion as relation and we write X\Y for the intersection and X[Yfor the join of the subgroupsX;YofG. Further we writeX<YifX is a maximal subgroup of Y.
Forn2Nand ap-groupG, we let VnV (G) hx2Gjxpn1i
n(G) hxpnjx2Gi and we denote by
Cnthe cyclic group of ordern
D2nthe dihedral group of order 2n (n2) Q2n the quaternion group of order 2n (n3) S2n the semidihedral group of order 2n(n4).
Further notation is standard (see [3]and [5]).
All graphs considered are undirected graphs. In particular, we define theHasse graph Lof a finite latticeLto be the (undirected) graph with vertex setLin which the unordered pairfx;ygof elements ofLis an edge if and only if one ofx;yis a lower neighbour of the other. The degree of a vertex is the number of edges in which it is contained. For a groupG, the Hasse graph L(G) of its subgroup lattice L(G) is also called the Hasse graph ofG.
A path g (of length n) from u to v in L is a sequence of elements x0;x1;. . .;xn of Lsuch thatx0u; xnv and for all ithere is an edge betweenxiandxi1inL, that is,xiis a lower or upper neighbour ofxi1. We usually writeg(x0;. . .;xn) for such a path and call the vertices dif- ferent from x0 and xn the internal vertices of g. Finally, a collection fg1;. . .;gmgof paths is calledinternally disjointif each internal vertex of gi(i1;. . .;m) lies on nogj (j6i).
If V is a subgraph of L which is a subdivision of K5, then there exist 5 points x1;. . .;x52V and an internally disjoint set of 10 paths gij fromxitoxj (1i<j5) in V, and hence inL. Conversely, every subset G fx1;. . .;x5g of L with jGj 5 together with an internally disjoint set of paths gij from xi to xj (1i<j5) in L yields a subgraphV ofL which is a subdivision ofK5. We shall call such a set G a K5-set in L and the members of G are called theK5-points ofV.
They are uniquely determined by V since all the other vertices of V have degree 2 in V.
If we want to show, for a certain groupG, thatL(G)has such a sub- graph V, we usually only give aK5-setGand describe the nontrivial paths gijbetween the members ofG; the trivial ones being edges or, for example, one of thep1 possible paths of length 2 between the bottom and the top of an elementary abelian section of orderp2inG.
Finally, as mentioned in the introduction, we call a graphK5-freeif it contains no subdivision ofK5as a subgraph.
3. p-Groups and Subdivisions ofK5.
In this section we shall determine all finite p-groups whose Hasse graphs areK5-free. We start with the following trivial, but fundamental, result.
LEMMA3.1. If G is elementary abelian of order p3, then L(G)contains a subdivision of K5.
PROOF. Let G hai hbi hci where o(a)o(b)o(c)p and let G : f1;hbi;ha;bi;hb;ci;Gg. There are five edges between these points in L(G) and the other five paths needed are (1;habi;ha;bi), (1;hbci;hb;ci), (1;habci;hab;ci;G), (hbi;hb;aci;G), and (ha;bi;hai;ha;ci;hci;hb;ci). So G together with these paths yields a subdivision ofK5inL(G). p REMARK 3.2. By [7, Theorem A], all other groups of order p3 have planar subgroup lattices except the nonabelian group H of exponent p (p>2). Since all subgroups of orderpdifferent fromZ(H) have degree 2 in the graph L(H), it is easy to see that L(H) isK5-free. So the ele- mentary abelian group is the unique group of orderp3whose Hasse graph
contains a subdivision ofK5. p
Lemma 3.1 shows that ifGis ap-group withK5-free Hasse graph, then every subgroup of G is generated by two elements. Groups with this property were studied by Blackburn [1]; most of them are metacyclic.
LEMMA 3.3. Let jGj pn; n5. If L(G) is K5-free, then G is meta- cyclic and has exponent at least pn 2.
PROOF. We show first thatGis metacyclic. By Lemma 3.1 and [1, The- orems 4.1 and 5.1], this is true except possibly whenGis a 3-group of maximal class. (Blackburn has a weaker assumption in his Theorem 4.1, so he gets a further possibility forp>2, namely a groupGof orderp5in whichV(G) is nonabelian of orderp3and exponentpandG=V(G) is cyclic. However, if we takex2Gsuch thatG hxiV(G) and a maximal subgroupSofGcontain- ing x, then F(S) is a cyclic normal subgroup of order p2 in G and so F(S)V(G)=F(S)\V(G) is elementary abelian of orderp3.) So suppose, for a contradiction,thatGisa3-groupofmaximalclass.ThenGhasafactorgroupof order35which is also of maximal class.Weshow thatthe Hasse graph of sucha group contains a subdivision ofK5; this will be the desired contradiction.
So letG0be of order 35and class 4. Then it is well-known [3, pp. 370-371]
thatG0 has normal subgroupsGiwithG0>G1>G2>G3>G4>1 such thatG1is metacyclic,G3V(G1)F(G1) andG2(G0)0is abelian. LetM be a maximal subgroup ofG0such thatM6G1. ThenMhas maximal class [3, p. 374]and so F(M)G3 since G0 has only one normal subgroup of order 32. Similarly, F(G2)G4Z(G0). IfM=G4 would contain a cyclic subgroup H=G4 of order 32, then H would be abelian and hence H\G2Z(M), a contradiction sinceMhas maximal class. ThusM=G4has exponent 3. Now we claim that G : fM;G1;G2;G3;G4g is a K5-set in
L(G0). Here we use 3 nontrivial paths, namely (M;G0;G1);(M;M1;M2;G4) with maximal subgroupsM16G2 of M andM26G3 of M1 (which exist sinceM=G4 has exponent 3), and (G1;SF(G1);S;V(S);1;G4) whereSis a cyclic subgroup of order 32ofG1not contained inG2and not containingG4 (which exists sinceV(G1)F(G1)). In addition, 4 paths are edges and in the elementary abelian factors M=G3;G1=G3, andG2=G4 there are inter- mediate subgroups not yet used. So we get a subdivision ofK5 inL(G0) and this completes the proof thatGis metacyclic.
It remains to be shown that expGpn 2. So suppose, for a contra- diction, that expGpn 3. SinceGis metacyclic,GNXwhereN/Gand N and X are cyclic. Then jN:N\Xj jG:Xj p3 and jX:N\Xj
jG:Nj p3. So G=N\X contains a subgroup isomorphic to a semi- direct product of a cyclic groupN1of orderp3by a cyclic groupX1of order p3. Forp>2, such a group has modular subgroup lattice and is therefore lattice-isomorphic to an abelian group of type (p3;p3) [5, 2.3.1 and 2.5.9].
Forp2; AutC8is elementary abelian of order 4 and henceX1induces an automorphism of order at most 2 inN1. So to get a contradiction, we show that the Hasse graph of a semidirect product H haihbi, wherehai/H, o(a)o(b)p3 andbp2CH(a), contains a subdivision ofK5. For this let G: fH1;. . .;H5g where H1 hap2i, H2 hap2;bp2i, H3 hap;bp2i, H4 hap;bpiandH5 ha;bpi. Clearly,H5is abelian of type (p3;p2) and we have 4 edges fHi;Hi1g, 2 nondiagonal paths (H1;hapi;hai;ha;bp2i;H5) and (H2;hap2;bpi;H4), 3 diagonal paths (H1;hapbp2i;H3); (H3;habpi;H5) and (H1;1;hap2bp2i;hapbpi;hap2;apbpi;H4), and finally the path (H2;hbp2i;hbpi;hbi;hap2;bi;hap;bi;H;H5). SoL(H)contains a subdivision of K5. But L(G) has a subgraph isomorphic to such anL(H), a contra-
diction. Thus expGpn 2. p
Forp>2, the converse of Lemma 3.3 also holds. To prove this, we need some preliminaries which will also be used in the casep2.
DEFINITION3.4. LetGbe a group.
(a) We denote byL5(G) the set of all subgroupsHofGfor which there exists aK5-setG inL(G) such thatH2G.
(b) For every subsetLofL(G) containingL5(G) we define the graph L^in the following way: the set of vertices ofL^isLand a two element subset fH;KgofLis an edge if and only if there exists a path (X0;. . .;Xr) inL(G) such thatX0H,XrK andXi2L(G)n Lfor alli1;. . .;r 1.
Then we have the following trivial result.
LEMMA3.5. Let L5(G) L L(G). IfGis a K5-set in L(G), then it is also a K5-set inL.^
PROOF. Let G fH1;. . .;H5g. SinceL5(G) L, clearly G L. Fur- thermore there exist internally disjoint paths gij from Hi to Hj in L(G) (1i<j5). If we remove in these paths all the elements X2L(G)n L, we get internally disjoint paths^gijfromHitoHjin the graph L. Thus^ Gis aK5-set inL.^ p LEMMA3.6. Let p2P; 2m2Nand G hai hbiwhere o(a)pm and o(b)p2. Then L(G)is K5-free.
PROOF. Let A: hai;B: hbi and L: fAiBjj0im;0j2g whereAi: hapm iiis the subgroup of orderpiofAandBj: hbp2jiis the subgroup of orderpjofB. We want to apply Lemma 3.5 with this latticeLand therefore have to show thatL5(G) L.
For this letH2L5(G) and suppose first thatG=His cyclic. Since every element ofL5(G) clearly has degree at least 4 in the graphL(G), it follows thatHis not cyclic. Suppose, for a contradiction, thatHhas type (pk;p) for some k1. Ifk1, then HV(G) andG=Hwould not be cyclic. Thus k2 and Hhas exactly one noncyclic andp cyclic maximal subgroups.
There are 4 internally disjoint paths fromHto the otherK5-points of aK5- set inL(G). SinceG=His cyclic, at least two of these paths have to use cyclic maximal subgroups ofH, at most one of these can go on toF(H). So at least one path has to stop at a maximal cyclic subgroupXofHor to go on to a subgroupY 6HcoveringX. In the first case,X2L5(G) and again there exists Y 6H covering X. Clearly, Y is cyclic; otherwise V(G) Y\HX, a contradiction. Hence XF(Y)F(G) (see [3, p. 269]) and soHXV(G)F(G); but this contradicts the fact thatG=His cyclic.
ThusHcannot have type (pk;p) and sinceGis abelian of type (pm;p2), the type ofHmust be (pk;p2) for somek2. ThereforeHcontains the unique subgroupV2(G) of type (p2;p2) ofG. We have shown:
(1) IfH2L5(G) andG=His cyclic, thenV2(G)H.
It follows thatHAiBfor somei2 and soH2 L. SinceL(G) is self-dual [5, p. 454], (1) also implies that a cyclic subgroupK inL5(G) is contained in V
2(G) and hence KAi2 L (where im 2). Finally, if neither H nor G=H is cyclic, then V(G)HF(G) and hence HAiB12 L(where 1im 1). So we have shown that
(2) L5(G) L.
We now determineL. By Definition 3.4, every edge in^ L(G)between two elements of L is also an edge inL. Furthermore, every elementary^ abelian sectionAi1Bj1=AiBj (0i<m; 0j<2) yields an edge betweenAiBjandAi1Bj1. Finally, there is an edge betweenAiand Ai2B(0im 2) since there is a diagonal path in the factor group Ai2B=Ai'Cp2Cp2 given by
hapm ii; hapm i1bpi; hapm i 2bi; hapm i2b;bpi; hapm i 2i hbi : We claim that there are no more edges inL:^
(3) The edges fAiBj;Ai1Bjg(i<m), fAiBj;AiBj1g (j<2); fAiBj;Ai1Bj1g (i<m; j<2), and fAi;Ai2Bg (im 2) are all the edges of the graphL.^
For this, letH;Kbe inLwithH6Kand let (H;X1;. . .;Xr;K) be a path inL(G)such thatXi2L(G)n Lfor alli(so thatfH;Kgis an edge inL).^ Clearly, we may assume thatr1 and we suppose first thatHAk for some k. Then km 1; and if km 1, it follows that r1 and KAmB1, so that fH;Kg is one of the edges in (3). Therefore let km 2. Then we prove by induction onithat fori1;. . .;r,
(a) H<Xi<Ak2B:L,
(b) A\XiHHB\Xi ifXiis cyclic, and (g) Xi<L ifXiis not cyclic.
This is clear fori1. So suppose it holds for somei<rand consider first the case thatXiis cyclic. IfXi1<Xi, then by (a),HXi1and since Xi162 L, we have thatH<Xi1. SinceL(Xi) is a chain, (b) holds forXi1. IfXi<Xi1andXi1 is cyclic, then again Xi1 satisfies (b) sinceXidoes;
furthermore, Xi1Ak2B since o(b)p2. Finally, if Xi<Xi1 and Xi1 is not cyclic, then Xi1XiV(G)L. Furthermore, since Xi16Ak1B1, we have jXi:Hj p2 and so Xi1<L. Now suppose that Xi is not cyclic. Then V(G)Xi and since Xi62 L, we have V2(G)6XiandXi6F(G); in particular,G=Xiis cyclic. So ifXi<Xi1, it follows that Xi1XiV2(G)L. By (g), Xi<L and so Xi1L, a con- tradiction. Thus Xi1<Xi<L. Since Ak2B16Xi6Ak1B, every cyclic maximal subgroup ofXisatisfies (b) and the only noncyclic maximal subgroup ofXiisF(L)Ak1B12 L. So we have shown that (a) - (g) hold. IfXr is cyclic and jXr:Hj p2, then (b) shows that neither F(Xr) norXrV(G) belong toL. ThusjXr:Hj pandKXrV(G)Ak1B1. IfXr is not cyclic, then by (g),Xr is one of thep 1 maximal subgroups
different from Ak2B1 and Ak1B of L. Then G=Xr is cyclic and it follows thatK L orKF(L)Ak1B1. In all cases, fH;Kg is one of the edges in (3).
Since L(G) is self-dual, the above argument also covers the case that HAkBfor somek. It remains to consider the case thatHAkB1for somek. Here, ifk0 orkm, thenr1 andKA1BorK Am 1, respectively. So suppose that 1km 1. Again using the duality of L(G), we may assume thatX1is a cyclic maximal subgroup ofAkB1. Then F(X1)Ak 1and so eitherKAk 1 or (Ak 1;X1;. . .;Xr;K) is one of the paths considered above. In the latter case, K is one of the groups Ak, Ak 1B1,Ak1B, as we have shown; sofH;Kgis one of the edges in (3).
We finally show that the graph (4) L^ is planar.
For this we just map the elementAiBj ofLto the point (j;i) inR2. Then we connect pairs of points of the form f(r;s);(r;s1)g, f(r;s);(r1;s)gandf(r;s);(r1;s1)gin the obvious way by straight line segments. Finally we have to connect the points (0;s) with (2;s2) and we can do this using the straight line segment from (0;s) to ( 2s 2; s 2) together with the lower half of the circle with center ( s;0) and radius (s2)
p2
. Since these circles do not intersect, this is a planar representation of the graphL; a similar one is shown in Figure 2 for^ m4.
Figure 2
It is obvious that Lemma 3.6 follows from (4) and Lemma 3.5. For, if L(G) would contain a subdivision of K5, then so would L; but a planar^
graph contains no subdivision ofK5. p
Now we prove Theorem A; we also determine the occurring groups.
THEOREM3.7. LetjGj pnwhere2<p2Pand5n2N. Then the followingproperties are equivalent.
(a) L(G)is K5-free.
(b) G is metacyclic andexpGpn 2.
(c) G is either abelian of type(pn);(pn 1;p), or(pn 2;p2), or G is one of the followingnonabelian groups:
(c4) G ha;bjapn1 bp1; aba1pn2i, (c5) G ha;bjapn2 bp2 1; aba1pn 3i, (c6) G ha;bjapn2 bp2 1; aba1pn 4i, (c7) G ha;bjap2bpn2 1; aba1pi, (c8) G ha;bjap31; bpn3ap2; aba1pi.
PROOF. By Lemma 3.3, (a) implies (b), and that (c) implies (b) is ob- vious. So suppose that (b) holds. Then again by Iwasawa's and Baer's the- orems [5, 2.3.1 and 2.5.9], since p>2; L(G) is modular and therefore is isomorphic to the subgroup lattice of an abelianp-group which clearly also has exponent at leastpn 2. Thus (a) follows from Lemma 3.6 and [7, The- orem A]. Furthermore, Gis one of the groups in (c) if expGpn 1. If expGpn 2andGhas a cyclic normal subgrouphaiof orderpn 2, then by [5, 2.3.11],haihas a complementhbiinGandGis abelian or the group in (c5) or (c6). Finally, suppose that expGpn 2 and G has no cyclic normal subgroup of orderpn 2. By [5, 2.3.18]there are an abelian normal subgroup A, an element b of order pn 2 in G and s2N such that GAhbi and aba1ps for all a2A. Then jA:A\ hbij jG:hbij p2 and since G06 hbi, it follows thats1 andA=A\ hbiis cyclic. Soxxbx1pfor hxi A\ hbiand hencejA\ hbij p. It follows easily thatGis the group in
(c7) or (c8). p
Forp>2, Theorem 3.7 gives the groups of orderpnwithK5-free Hasse graphs ifn5. Of course, it is not difficult to determine also the groups of orderp4with this property. We give the result without proof.
REMARK3.8. LetjGj p4; p>2. ThenL(G)isK5-free if and only if Gis metacyclic or one of the following groups:
(a) G hx;y;zjxpypzp2 1[y;z];yxyzp;zxzyi, (b) G hx;y;zjxpypzp2 1[y;z];yxyzsp;zxzyi wherep5 andsis a quadratic nonresidue modulop,
(c) G hx;y;zjy3z91[y;z];x3z3;yxyz 3;zxzyi.
We turn to the casejGj 2n. Here, for simplicity, we shall prove our main result only for n6. But for this, we first have to study certain groups of order 25.
LEMMA3.9. If G is a dihedral or semidihedral group of order25or if G ha;bja8b41;aba 1i, then L(G)contains a subdivision of K5. PROOF. Suppose first thatG'D32orG'S32. Then in both cases,G has exactly three maximal subgroupsA;D;M such thatA haiis cyclic, D'D16 and M is quaternion or dihedral; furthermore Z(G) ha8iand G=Z(G)'D16. LetG: fZ(G);ha4i;U;V;DgwhereUandV are the two dihedral subgroups of order 8 ofD. SinceU\V ha4i, we see that there are 5 edges inL(G) between members ofG. Further obvious paths are (ha4i;ha2i;D) and (D;G;M;M1;M2;Z(G)) with noncyclic M1<M and ha4i 6M2<M1. SinceU=Z(G) andV=Z(G) are elementary abelian of or- der 4, there are trivial paths (Z(G);U1;U) and (Z(G);V1;V) with U16 ha4i 6V1and there are further noncyclic maximal subgroupsU2ofU and V2 of V. So we finally get a path (U;U2;X;1;Y;V2;V) with jXj jYj 2 andX6Z(G)6Y. All these paths are internally disjoint and soGis aK5-set inL(G).
Now let G ha;bja8b41;aba 1i. Then Z(G) ha4;b2i and G=hb2i 'D16. FurthermoreGhas three maximal subgroupsM hai hb2i andM1;M2 for whichMi=hb2i 'D8 (i1;2). ClearlyF(G)M1\M2
ha2;b2iand we letG: fM1;M2;F(G);Z(G);hb2ig. This time there are 4 edges between members ofG, two trivial paths fromMitoZ(G) and a third noncyclic maximal subgroupHiinMiwhich yields a path fromMitohb2ifor i1;2. The final two paths are (M1;G;M2) and (hb2i;1;ha4i;ha2i;F(G)). So
Gis aK5-set inL(G). p
LEMMA3.10. Let6n2Nand suppose that G is a nonabelian group of order2nsuch that L(G)is K5-free. Then G is one of the followinggroups.
(a) G ha;bja2n1b21;aba12n2i (b) G ha;bja2n2b41;aba12n3i (c) G ha;bja2n2b41;aba12n4i (d) G ha;bja4 b2n21;aba 1i (e) G ha;bja8 1;b2n 3a4;aba 1i
PROOF. By Lemma 3.3,Gis metacyclic and has exponent at least 2n 2. If expG2n 1, then eitherGis the group with modular subgroup lattice in (a) orGis dihedral, quaternion, or semidihedral [3, p. 90]. In the latter three cases,G=Z(G)'D2n1which is impossible by Lemma 3.9 sincen6. Thus Gsatisfies (a).
So let expG2n 2. SinceGis metacyclic, there existsN/Gsuch that N haiis cyclic andGNhxifor somex2G. Consider first the case that no such N has order 2n 2 and let HNhx2i. Since ax2ar with r1(mod 4);L(H) is modular [5, 2.3.4]and since His generated by two elements of order at most 2n 3, it follows from [5, 2.3.5]that expH2n 3. Hence there existsb2GnHsuch thato(b)2n 2. Then GNhbi and jN:N\ hbij jG:hbij 4. Clearly, abat for some t2N and if t1(mod 4), then [a;b]2 hbiandhbi/G which would contradict our as- sumption in this case. So t61(mod 4) and it follows that jN\ hbij 2.
Thus either jNj 4 andGis the group in (d) orjNj 8; b2n3 a4 and aba 1 oraba3. In the latter case, (ab2n4)ba3b2n 4(ab2n4) 1 and we may replaceabyab2n4 to obtain the group in (e).
It is left to consider the case that GNhxi where N hai/G and o(a)2n 2. Againaxat for somet2Nand ift1(mod 4), then by [5, 2.3.4], L(G) is modular. Since expG24,G is not hamiltonian; then [5, 2.3.11]implies that N has a complement hbiin G. Thus Gis one of the groups in (b) or (c).
So lett61(mod 4). Then againjN\ hxij 2 and we show that this case cannot occur. Ifxinduces an automorphism of order 2 inN, thenaxa 1 oraxa 12n3and soG=hx2iis a dihedral or semidihedral group of order 2n 1 or 2n 2. This contradicts Lemma 3.9 except when n6 and N\ hxi 61. But in this case, G=ha8i is the third group of order 25 in Lemma 3.9, again a contradiction. Thus x induces an automorphism of order 4 in N. Since AutN has only two cyclic subgroups of order 4 [3, p. 84]andt61(mod 4), it follows that aba 12n 4 forbxorbx 1. Then (ab)4a2n3b4; so ifN\ hbi 61, we have (ab)41 and can replaceb byab. Thus we may assume thatN\ hbi 1. ThenG=ha2n 3;b2iis semi- dihedral of order 2n 2 and Lemma 3.9 implies that n6. So G ha;bja16b41;aba3i and G has the following maximal sub- groups: H: ha;b2i has modular subgroup lattice, S: ha2;bi and T: ha2;abi. Put W:V2(H) ha4;b2i and Z:Z(G) ha8i and let
G: fZ;W;F(G);S;Tg. There are 3 edges (connectingF(G) toW;S;T)
and 4 obvious paths (S;G;T);(Z;Zhb2i;W);(Z;ha4i;ha2i;F(G)) and (Z;1;hb2i;hbi;ha8;bi;ha4;bi;S). Since G=W 'D8, there are trivial paths (S;S1;W) and (T;T1;W) with T16F(G)6S16 ha4;bi. Finally, (ab)2
a12b2and so ifX h(ab)2i, thenT=X'D8andZ<X. Thus there exists a noncyclic subgroup T2=X of order 4 with T2 6T1 and a path (T;T2;T3;X;Z) so that all the paths considered are internally disjoint. So L(G)contains a subdivision ofK5, a contradiction. p PROOF OFTHEOREM B. Let jGj 2n where 6n2N. Again by [5, 2,3.1 and 2.5.9], the groups in (a) - (c) of Lemma 3.10 have modular sub- group lattice and are therefore lattice-isomorphic to the abelian groups mentioned in Theorem B. So ifL(G)isK5-free, then by Lemmas 3.3 and 3.10,Gis one of the groups given in this theorem. Conversely, ifGis lat- tice-isomorphic to an abelian group of type (2n);(2n 1;2), or (2n 2;4), then by Lemma 3.6 and [7, Theorem A],GisK5-free.
So, finally, let Gbe one of the other two groups given in Theorem B.
Then in both cases, b22Z(G); let hb2i :Z and H: ha;b2i. Then His abelian of exponent 2n 3and hence there exists a complementhcitoZinH;
soHZ hciwitho(c)4.
Clearly,G=Z'D8. So everyx2GnHsatisfiesx22Zandhaihxi G since H is the unique maximal subgroup of G containing hai. Then H hai(H\ hxi) haihx2iand sincex22Z, it follows thathx2i Z. Since every subgroup ofGnot contained inHcontains an elementx2GnH, it follows that L(G) consists of the subgroups ofH, four cyclic subgroups X1;. . .;X4 containing Zas a maximal subgroup, two maximal subgroups M1andM2different fromH, andG.
Now it is easy to see that L5(G)L(H). For, these cyclic sub- groups Xi have degree 2 inL(G), G has degree 3, and both Mi have degree 4; however, also Mi cannot be a member of a K5-set since two of the four paths from Mi to the other members would have to start with a cyclic maximal subgroup of Mi and then use F(Mi), which is impossible.
Now suppose, for a contradiction, thatG is aK5-set inL(G). Then by Lemma 3.5, it is also a K5-set inL^ whereL L(H). Clearly,L^ contains L(H) and a further edge can only connect two of the three groups Z;Z hc2i;Hsince no other subgroup of His covered by an element of L(G)nL(H). Since there are already edges fromZ hc2itoZand toHin L(H), the only additional edge isfZ;Hg. LetZbe a cyclic 2-group con- tainingZas a maximal subgroup and considerHZ hcias a subgroup of index 2 in G:Z hci. In the subdivision ofK5 given byG in L, we^ replace the edgefZ;Hg- if it appears in one of the paths - by the path (Z;Z;Z hc2i;G;H). Then we get a subdivision of K5 in L(G); but this
contradicts Lemma 3.6. p
As in the casep>2 it is not difficult to determine the 2-groups of order less than 26 withK5-free Hasse graph. Again we give the result without proof.
REMARK3.11. LetjGj 2n;n5. ThenL(G)isK5-free if and only if one of the following holds:
(a) n4 andGis metacyclic,
(b) Gis one of the groups in Theorem B forn5, (c) G'Q32.
4. Proof of Theorem C.
It remains to determine the nonprimary finite groups with K5-free Hasse graphs. For this, we first look at direct products of lattices.
LEMMA4.1. Let LL1L2with finite lattices Li. Then Lcontains a subdivision of K5if one of the followingholds.
(a) L1 contains L(C2C2)as a sublattice andjL2j 3.
(b) L1 contains L(C4C2)as a sublattice andjL2j 2.
PROOF. (a) We may assume that L1L(H) f1;M1;M2;M3;Hg wherejHj 4 andjMij 2 fori1;2;3 and thatL2is a chainA<B<C of length 2. Then
G: fHA;1B;M1B;M2B;HBg
is aK5-set inLsince there are 5 trivial paths between members ofGin the interval [HB=1B], furthermore 4 obvious paths fromHAtoKB viaKA, and finally the path (M1B;M1C;HC;M2C;M2B).
(b) Here we may assume that L1L(H) where H hai hbi with o(a)4,o(b)2 and thatL2 f0;Igis a chain of length 1. Then
G : f10;ha2i 0;ha2;bi 0;ha2i I;ha2;bi Ig
is aK5-set inL since there are 5 edges between members ofG and we have the further paths (10;ha2bi0;ha2;bi0);(10;1I;ha2iI), (10;hbi0;hbiI, ha2;biI), (ha2i0;habi0;habiI;HI;ha2;biI),
and (ha2;bi0;H0,hai0;haiI;ha2iI). p
There is a similar result for the Hasse graph of a semidirect product of two groups. This graph contains a subdivision ofK5 if a certain smaller
configuration occurs in the Hasse graph of the complement. We first give this configuration a name.
DEFINITION4.2. A (3,6)-gonin a graph is a set of 3 distinct pointsR;S;T together with an internally disjoint set of 6 paths between them, one be- tweenRandS, two betweenSandT, and three betweenRandT.
Since at most one path between two points can be an edge, there are intermediate points in one of the paths betweenSandTand in two of the paths betweenRandT(see Figure 3).
LEMMA 4.3. Let GNK where 16N/G and N\K1. If L(K) contains a (3,6)-gon, then L(G)contains a subdivision of K5.
PROOF. Let R;S;T be the points of a (3,6)-gon in L(K), let e;d1;d2;r1;r2;r3be the appropriate paths fromRtoS,StoT, andRtoT with interior points U;V;W of d2;r2;r3, respectively. For every path g(X0;. . .;Xr), we letg 1:(Xr;. . .;X0) be the path with the same end- points in opposite direction. The isomorphism theorem for groups implies that if all theXiare subgroups ofK, then^g:(NX0;. . .;NXr) is a path in L(G)and that the setf^e;^d1;^d2;^r1;^r2;^r3gis internally disjoint. Finally, for every subgroup X of K, we let g(X)(X0;X1;. . .;Xr) be a fixed path in L(G) from X to NX such that Xi<Xi1 for all i. By Dedekind's law, Xi(N\Xi)X and hence Xi\KX for all i. This shows that for X6YK, the pathsg(X) andg(Y) are disjoint.
Figure 3
We now let G: fR;S;T;NR;NTg. There are obvious paths e;d1;r1 betweenR;S;T,g(R) betweenRandNR,g(T) betweenTandNT, and^r1 between NRand NT. Furthermore,g(S) and^e 1 yield a path from Sto NR(viaNS), and we get a path fromStoNTusingd2fromStoU, then g(U), and finally^d2 from NU to NT. Similarly we get paths from R to NTviar2,g(V);^r2and fromTtoNRviar31;g(W);^r31. ThusGis aK5-set
in L(G). p
We shall need the following result on groups without (3,6)-gons.
PROPOSITION4.4. LetKbe a group such thatL(K)contains no (3,6)- gon. ThenKhas one of the followingproperties (wherepandqare primes and2n2N).
(a) K is cyclic.
(b) K is abelian of type(2n;2)or lattice-isomorphic to this group.
(c) K is elementary abelian of order p2or nonabelian of order pq.
(d) K is isomorphic to Q8; C4C4, or CpC2C2 where p>2.
PROOF. LetKbe a minimal counterexample and suppose first thatKis a 2-group. The trivial subgroupTand the two noncyclic subgroupsR;Sof order 4 clearly yield a (3,6)-gon inL(D8). SoD8cannot be involved inKand [5, 2.3.3]
shows thatL(K) is modular. SinceL(D8)is a subgraph ofL(C2C2C2), K is generated by two elements and so, again by [5, 2.3.1 and 2.5.9], either K'Q8orL(K)'L(H) whereHis abelian of type (2n) or (2n;2m). However, ifC8C4'H0H, thenRV2(H0);TV(H0);S V
2(H0) would yield a (3,6)-gon inL(H), a contradiction. It follows thatK 'H'C4C4or that m1. In every case,Kis one of the groups in (a) - (d), a contradiction. SoKis not a 2-group.
Now suppose, for a contradiction, that there are subgroupsA/PK such that P=Ais elementary abelian of orderp2;p>2, or nonabelian of orderpq;p>q;pandqprimes. Suppose first thatPK. Then sinceKis a counterexample, A61. LetA<S<K. IfSwould have a maximal sub- group B6A, thenRK,TA andSwould yield a (3,6)-gon inL(K) with 5 paths inside [K=A]and the sixth connectingStoTviaBandT\B.
This contradiction shows that all subgroups strictly betweenAandK are cyclic of prime power order; sinceA61, it follows thatKis ap-group all of whose maximal subgroups are cyclic. Hence K'Q8 (see [3, p. 311]), a contradiction sincep>2.
ThusP<Kand the minimality ofKimplies thatA1 and thatPis a maximal subgroup of K. Let x2KnP . If PxP and x normalizes a
minimal subgroupMofP, thenR1,TPtogether withSP\ hxiif P\ hxi 61 andSMifP\ hxi 1 yield a (3,6)-gon inL(K)with 5 paths inside [P=1], the sixth path being (S;hxi;K;T) or (S;Shxi;K;T), respec- tively. This is a contradiction; so ifPxP, thenxoperates irreducibly on P. In particular,hxi \P1; nowRP; T1 and SK yield a (3,6)- gon, again a contradiction. It follows that Px6P. If P\Px61, again R1; TP and SP\Px would yield a (3,6)-gon with sixth path (S;Px;K;T). So, finally, P\Px1 for all x2KnP; but then again RP; T1 andSKyield a (3,6)-gon. This is the desired contradiction which shows that for p>2, neither CpCp nor a nonabelian group of orderpq(p>q2P) are involved inK.
It follows that Sylow p-subgroups are cyclic; moreover, Burnside's theorem [3, p. 419]implies that K has a normal p-complement. The in- tersection of all these normalp-complements (2<p2P) is a normal Sylow 2-subgroup HofK with cyclic factor group. The minimality ofK implies thatHis one of the groups in (a) - (d) andKHCwith cyclic groupCof odd order. IfCwould operate nontrivially onH, then it would also operate nontrivially on H=F(H) (see [3, p. 275]). So jH=F(H)j 4 and it would follow that the alternating groupA4would be involved inK. ButL(A4)has a (3,6)-gon given by RA4; T1 and the subgroupS of order 4. This contradiction shows thatKHC.
SinceKis not a 2-group,C61; sinceKis a counterexample,His not cyclic. So jH:F(H)j 4 and if F(H)61, then RH; TF(H) and STCwould yield a (3,6)-gon inL(K), a contradiction. ThusjHj 4.
Since K is a counterexample, C contains a subgroup D'CpCq or E'Cp2, p and q primes. However, in L(HD) we get a (3,6)-gon RH; SDandT1 with obvious paths. InL(HE) we may take RE; SV(E) andTHS; here, ifA;B;Care the three subgroups of order 2 of H, we have the paths (S;1;A;H;T);(S;AS;T) and (R;AR;HR;T);(R;BR;BS;T);(R;CR;CS;T) to get a
(3,6)-gon. This is a final contradiction. p
Since one of the three points of a (3,6)-gon has degree at least 5 in the graph, it is easy to see that the Hasse graphs of the groups in (b) - (d) of Proposition 4.4 contain no (3,6)-gon and that the cyclic groups with this property are those with planar subgroup lattice and those whose order is the product of 4 pairwise different primes. We shall not need this.
An immediate consequence of 4.1 - 4.4 is the following result on co- prime direct products which we shall use quite often. First of all, it com-
pletes the determination of nilpotent groups withK5-free Hasse graphs, that is, it implies Theorem C.
LEMMA4.5. Let GAB where A616B and(jAj;jBj)1. If L(G) is K5-free, then either G is cyclic or GPK wherejPj p and K is isomorphic to Q8or elementary abelian of order q2or nonabelian of order qr with pairwise different primes p;q;r.
PROOF. Suppose thatGis not cyclic. Then one of the two factorsAand B, sayB, is not cyclic. By [6, Theorem 1.1],L(B) containsL(C2C2) as a sublattice. By [5, 1.6.4],L(G)'L(A)L(B) and so Lemma 4.1, (a) shows that jL(A)j 2. ThusjAj p for some prime pand now Lemma 4.1, (b) yields thatL(B) does not containL(C4C2) as a sublattice. By Lemma 4.3, L(B) contains no (3,6)-gon and so Proposition 4.4 implies thatGis one of the groups in Lemma 4.5 orB'CqC2C2 withq626p. But in this case,G'(CpCq)(C2C2) which would contradict Lemma 4.1,(a). p PROOF OFTHEOREMC. IfL(G)isK5-free, then by Lemma 4.5,Gis one of the groups given in the theorem. Conversely, ifGPQwithjPj p is one of these groups, thenL(G) contains only four points of degree at least 4, namelyZ(Q);Q;PZ(Q);GifQ'Q8and 1;Q;P;GifQ'CqCq.
Hence there is noK5-set inL(G). p
5. Proof of Theorem D.
LEMMA5.1. All the groups in (a) - (h) of Theorem D have K5-free Hasse graphs.
PROOF. In most cases this is rather obvious; in some cases it even follows already from the fact that there do not exist 5 subgroups ofGwith degree at least 4 inL(G). This, for example, holds ifGsatisfies (a). For, thenCK(P) and the subgroups properly containingPCK(P) have degree p1 orp2, whereas all other subgroups have degree at most 3; since jK:CK(P)j q3, there are at most 4 subgroups of degree at least 4. And ifG satisfies (b), then Ghas precisely 4 subgroups of degree at least 4. If G satisfies (d), then in addition to CK(P) and the subgroups properly con- taining PCK(P), also 1;P, and PCK(P) have degree at least 4. But every path between one of the groups 1;P, orCK(P) and one of the groups containingPCK(P) has to useCK(P) orPCK(P). Since we would need
at least 4 paths of this type for aK5-set, it follows that L(G) isK5-free.
Similarly, ifGsatisfies (e), then the only subgroups of degree at least 4 are 1;P;PQ;PR;GwhereKQRwithjQj qandjRj r; however, there is no path fromPtoGwhich does not usePQ;PR, or 1. Finally, ifGsatisfies (f), thenq6 jp 1 since K is irreducible onP=F(P). Hence K centralizes F(P) and it follows that 1;F(P);P;Gare the only subgroups of degree at least 4 inL(G).
Now suppose thatGsatisfies (c). Then subgroups of orderqnhave degree 2 and subgroups of orderpqnofGhave degreep2 inL(G); however, all but 2 of the paths from such a subgroup to the other members of aK5-set would have to start with a cyclic subgroup of orderqnand then useCK(P), which is impossible. So if we let H:PCK(P), it follows that L5(G) L(H)[ fGg :L. SinceCK(P) hxifor every x2GnH, the edges of L^ are the edges of L(H) together with fH;CK(P)g;fH;Gg;fCK(P);Gg and fV(P)CK(P);Gg. Thus L^ is planar and by Lemma 3.5, L(G) is K5-free.
Now let G be the group in (g) with jKj 9. Then there are only 6 subgroups of degree at least 4, namely F(P);P;CK(P);F(P)CK(P), PCK(P), andG. SoL5(G) LwhereLis the set of these 6 subgroups. It is easy to see thatP;CK(P) andGhave degree 3 inL. By Lemma 3.5,^ L(G) isK5-free. The other group in (g) is isomorphic toG=CK(P).
Finally, suppose thatGsatisfies (h). Then every maximal subgroupM ofPhas degreep2; however, all but 2 of the paths fromMto a member of a K5-set would have to start with a cyclic maximal subgroup ofMand then use F(M). This is impossible. SinceK is also irreducible onF(P)
V(P), a dual argument shows that alsoF(M) cannot be a member of aK5- set. Furthermore,F(P)Khas degreep22, but again all but 2 paths would start with cyclic subgroups of orderqand contain 1, which is impossible.
This leaves only 1;F(P);P;Gas possible members ofL5(G). So there is no
K5-set inL(G). p
We now prove in a couple of steps that every nonnilpotent finite group withK5-free Hasse graph is one of the groups in (a) - (h) of Theorem D.
First we study groups with a normal Sylowp-subgroup. For this we need two simple results on small groups.
LEMMA 5.2. Let p;q2P such that p6q and suppose that GPQ where P is an elementary abelian normal subgroup of order p2,jQj q and [P;Q]61. If L(G)is K5-free, then Q operates irreducibly on P.
PROOF. If not, then by Maschke's theorem, PP1P2 where Pi/G(i1;2) and P1Q, say, is nonabelian of order pq. Then G: f1;P2;P;P1Q;Gg is aK5-set inL(G); here 7 paths are trivial in the sense of §2 and since p>q, there are further paths (P;P1;P1Q), (P2;P2Q;Q;P1Q) and (1;Qx;P2Qx;G) for somex2P. This contradicts our
assumption. p
LEMMA5.3. Let GPK where P/G; jPj p or p2, p does not divide jKjand CK(P)1. If L(G)is K5-free, thenjL(K)j 4.
PROOF. Suppose, for a contradiction, thatjL(K)j 5. IfPis cyclic, we may assume thatjPj p. ThenKis cyclic, sojKj>8 and hencejPj>10. If Pis not cyclic, then by Lemma 5.2, every nontrivial subgroup ofKoperates irreducibly onP. HencejKjis odd andjL(K)j 5 implies thatjPj>10. In both cases,NG(K)K; so there are at least 10 complements toPinGand any two of them intersect trivially (see [5, 4.1.1]).
By assumption, there are pairwise different subgroups Hi satisfying P<HiG(i1;. . .;4); let H51. We choose 10 pairwise different complements Kij (1i<j5) to P in G and paths dij in L(Kij) from Hi\Kij toHj\Kij via the join of the two groups. Then the set of paths (Hi;dij;Hj) for 1i<j4 and (Hi;di5) for 1i4 is internally dis- joint. SofH1;. . .;H5gis aK5-set inL(G), a contradiction. p LEMMA5.4. Let GPK where P is a normal p-subgroup, p a prime, K a p0-group and [P;K]61. If L(G) is K5-free, then there are primes q;r such that either
(a) K is a cyclic q-group andjK:CK(P)j q3, or (b) K is cyclic of order qr.
PROOF. LetGbe a minimal counterexample. Since then alsoG=F(P) is a counterexample [3, p. 275],F(P)1 and Lemma 3.1 implies that
(1) jPj porPis elementary abelian of orderp2.
Let K0:CK(P). By Lemma 5.3, jL(K=K0)j 4. Since the subgroup lattice of every noncyclic group contains L(C2C2) as a sublattice [6, Theorem 1.1], it follows thatK=K0is cyclic and that there are primesq;r such that
(2) K=K0is cyclic of orderqrorqkwhere 1k3.
Finally, by Lemma 4.3,
(3) L(K) contains no (3,6)-gon.
Now suppose first thatK=K0 is cyclic of orderqr; q6r. SinceGis a counterexample, K061. By Proposition 4.4, K is cyclic; otherwise K'CqC2C2 and PH with C2C2'HK would be a counter- example of order less thanjGj. LetQandRbe the subgroups betweenK0
andKsuch thatjQ:K0j qandjR:K0j r. The minimality ofGimplies thatQandRsatisfy (a) or (b) of the lemma; sinceK061, it follows thatjK0j is a prime andjK0j 6r, say. Lemma 4.3, applied withNK0, yields that L(PR) contains no (3,6)-gon; by Proposition 4.4, jPj p. We claim that G : f1;K0;PQ;PR;Gg is aK5-set in L(G). Here only 3 of the requir- ed paths are trivial, but, in addition, there are 3 obvious paths (1;P;PK0;PQ); (K0;Q;PQ) and (K0;R;PR). Since p5, there exist a;b;c;d2Psuch thatK;Ka;Kb;Kc;Kdare pairwise different. So if we let S be the subgroup of order r in K, we have 4 further paths (1;Sa;PS;PR); (1;Sb;K0Sb;Kb;G); (K0;Qc;Kc;G) and (PQ;PQ\Kd;Kd, PR\Kd;PR). Since any two different conjugates ofKintersect inK0, the set of these paths is internally disjoint. SoG is aK5-set, a contradiction.
It follows thatK=K0is cyclic of orderqkwhere 1k3. We claim that k1 and
(4) K is elementary abelian of order q2 or nonabelian of order qt withq<t2P.
For this, suppose first thatKis aq-group. Ifk2, then every maximal subgroup ofKwould operate nontrivially onPand hence would be cyclic by the minimality of G. So K would be isomorphic toCqCq or Q8 [3, p. 311], which would be impossible since k2. Thus k1 and hence F(K)K0. SoF(K)/Gand the minimality ofGimplies thatF(K)1. By Lemma 3.1, (4) holds. - Now suppose thatKis not aq-group. IfKST with S'C2C2 andjTj 2P, then the minimality ofGwould imply that tq; soSK0 andGPTS, contradicting Lemma 4.5. Thus Propo- sition 4.4 implies that (4) holds or thatKis cyclic. In the latter case, letQbe the Sylowq-subgroup ofK. Then Lemma 4.3, applied with aq-complement N in K0, yields that L(PQ) contains no (3,6)-gon; by Proposition 4.4, jPQj pq. It follows thatk1 andGPQK0. Now Lemma 4.5 implies that jK0jis a prime; but this is impossible since G is a counterexample.
Thus (4) holds in all cases.
LetQKsuch thatjQj qandQ6K0. By Lemma 4.3, applied with NK0; L(PQ) contains no (3,6)-gon and Proposition 4.4 implies