betaFeedback

WorksMathematics

Abelian covers and the index of the centre in a group

Abstract

Let h(n)h(n) be the largest number of abelian subgroups needed to cover a group in which every set of more than nn elements contains two distinct elements that commute, and let ζ(n)ζ(n) be the largest index of the centre under the same constraint. Erdős asked for the growth of hh, Neumann proved that both quantities are finite, and Pyber bounded both ζζ and hh exponentially. We determine h(n)h(n) for every positive integer nn and prove that h(n)=1+ζ(n)h(n)=1+\sqrt{ζ(n)} whenever the constraint admits a nonabelian group. The covers are built directly from the commutator geometry of a scalar quotient, and the comparison with ζζ is made afterwards. Beyond finitely many small cases, both maxima are attained by the same groups, those isoclinic to an extraspecial 22-group of order 21+2⌊(n−1)/2⌋2^{1+2\lfloor(n-1)/2\rfloor}, confirming Pyber's prediction for the central index in this range.

Introduction

How many abelian subgroups are needed to cover a group in which every sufficiently large set contains two commuting elements? The aim is to turn this restriction on noncommutation into a sharp bound for the size of an abelian cover. Write ω(G)\omega(G) for the largest size of a set of pairwise noncommuting elements, called a clique, and a(G)a(G) for the least size of an abelian cover. A third invariant, the central index z(G)=[G:Z(G)]z(G)=[G:Z(G)], measures the size of the quotient on which commutation is determined. These satisfy

ω(G)≤a(G)≤z(G):\omega(G) \le a(G) \le z(G):

a clique meets each abelian subgroup in at most one point, and each central coset xZ(G)xZ(G) lies in ⟨x,Z(G)⟩\langle x,Z(G)\rangle.

Erdős asked for the optimal bound on a(G)a(G) under the constraint ω(G)≤n\omega(G) \le n [7 Problem 26]. Neumann [15] proved that this constraint bounds the central index. Thus both extremal functions

h(n)=max⁡ω(G)≤na(G),ζ(n)=max⁡ω(G)≤nz(G)h(n)=\max_{\omega(G)\le n}a(G),\qquad\zeta(n)=\max_{\omega(G)\le n}z(G)

are finite. Pyber [16] proved an exponential upper bound for ζ(n)\zeta(n), hence also for h(n)h(n). Extraspecial 2-groups give the exponential lower bound attributed to Isaacs [3 p. 40].

The covering question also has a graph formulation. In the noncommuting graph ΓG\Gamma_G, the vertices are the noncentral elements and edges join noncommuting pairs. Then ω(G)=ω(ΓG)\omega(G)=\omega(\Gamma_G) for nonabelian GG, and a(G)=max⁡{1,χ(ΓG)}a(G)=\max\{1,\chi(\Gamma_G)\}: a colour class generates an abelian subgroup, and a cover gives a colouring. Faber, Laver and McKenzie [8] developed this connection, while Mason [14] and Bertram [3] studied quantitative relations between covers and noncommuting sets. See also Abdollahi, Akbari and Maimani [1] for the graph-theoretic perspective. Thus the covering problem asks for the optimal bound on the chromatic number of noncommuting graphs in terms of their clique number.

Here we determine h(n)h(n) for every nn, classify its maximizers for n≥9n \ge9, and identify its exact relation to ζ(n)\zeta(n). The covering profile and its equality case are proved directly. The comparison with ζ\zeta uses only the central-index theorem [13 Theorems 1.1 and 1.2(v)]. For n≥9n \ge9, the maximizers all have binary symplectic commutation geometry, in the dimension specified below.

Theorem 1.1 (The covering profile). For every n≥3n \ge3.

h(n)=1+ζ(n).h(n) = 1 + \sqrt{\zeta(n)}.

Explicitly, h(n)=T(n)h(n) = T(n) for all n≥1n \ge1, where

n1,23≤n≤67,8n≥9T(n)1n101+2⌊(n−1)/2⌋\begin{array}{c|ccccc} n & 1,2 & 3 \le n \le6 & 7,8 & n \ge9 \\ \hline T(n) & 1 & n & 10 & 1 + 2^{\lfloor(n-1)/2 \rfloor} \end{array}

For n≥9n \ge9, the groups attaining h(n)h(n) under ω(G)≤n\omega(G) \le n are exactly those isoclinic to an extraspecial 22-group of order 21+2⌊(n−1)/2⌋2^{1+2\lfloor(n-1)/2 \rfloor}. These are also the groups attaining ζ(n)\zeta(n).

Why a square root appears. The identity in Theorem 1.1 reflects the geometry of the extremizers. Let K=FprK = \mathbb{F}_p^r. We call a group a finite-field determinant model if its central quotient is the additive group of K2K^2 and its commutator map is

β((x1,x2),(y1,y2))=ρ(x1y2−x2y1),\beta((x_1,x_2),(y_1,y_2)) = \rho(x_1y_2 - x_2y_1),

where ρ\rho is a nonzero Fp\mathbb{F}_p-linear map onto the derived subgroup, written additively. The determinant itself and its scalar trace versions are both included. The KK-lines give a spread, and every such group satisfies

z(G)=∣K∣2,a(G)=∣K∣+1=1+z(G)z(G) = |K|^2,\qquad a(G) = |K| + 1 = 1 + \sqrt{z(G)}

(Corollary 3.3). The theorem shows that, for each n≥3n \ge3, a group of this type simultaneously maximizes the covering number and the central index under ω(G)≤n\omega(G) \le n. In the small cases the relevant fields have orders 2,3,4,5,92,3,4,5,9; for n≥9n \ge9, the models are the binary scalar trace models. Thus a common geometric construction explains the relation between the two extremal functions.

For N≥9N \ge9, put B(N)=2⌊(N−1)/2⌋B(N) = 2^{\lfloor(N-1)/2 \rfloor}, so T(N)=1+B(N)T(N) = 1 + B(N). The function BB is nondecreasing, with B(N)≥2N/2−1B(N) \ge2^{N/2-1} and B(N+2)=2B(N)B(N + 2) = 2B(N). In particular T(N)>NT(N) > N: this holds at N=9,10N = 9,10 and persists in steps of two.

The idea of the proof. The proof compares cliques and covers in the fibres of a quotient map. Call the largest clique size in a fibre its width. If the kernel is abelian, a fibre can be covered by exactly that many abelian subgroups. More generally, an abelian subgroup of index II in the kernel gives a cover of size at most II times the width. A clique in the quotient allows us to combine cliques in its fibres, so the sum of their widths is at most ω(G)\omega(G). Averaging over quotient cliques then bounds the total number of subgroups in the cover (§ 2).

For a pp-group, a scalar commutator form provides a symplectic quotient. In rank two, an abelian radical already gives a(P)=ω(P)a(P) = \omega(P) by Lemma 4.2. In the remaining cases, the symplectic group allows us to average fibre cliques (§ 4). When covering the fibres separately gives too many subgroups, we instead use subspaces isotropic for all commutator forms (§ 5). The model is the symplectic spread, which also constructs the sharp examples in § 3. For nonnilpotent groups, § 6 uses affine geometry in a minimal nonnilpotent quotient for the solvable case. A centralizer descent applies to every nonnilpotent group when N≥45N \ge45; nonsolvable groups with N≤44N \le44 are treated by the small-clique classification. For N≥9N \ge9, each construction gives a strict inequality except in the binary symplectic model with one commutator form. Tracking where equality is possible therefore identifies the maximizers.

The main distinction is between a scalar commutator form, which certifies noncommutation, and the full commutator map, which determines whether a subgroup is abelian. The former provides large cliques; the latter provides common isotropic subspaces and hence covers. Keeping both roles explicit allows the same quotient estimates to handle groups of different nilpotency classes. The case maps in § 4–§ 6 indicate where each estimate is used.

Classification input. The nonsolvable small-clique case uses the structural classification in [2 Theorem 1.1]; we do not claim a proof independent of the classification of finite simple groups (CFSG). The centralizer descent in § 6.1 uses only the classical Baer–Suzuki theorem.

Passing to finite groups. Two groups are isoclinic if isomorphisms of their central quotients and derived subgroups preserve the commutators of lifts of central cosets. A finite pp-group is extraspecial when its centre and derived subgroup coincide and have order pp. For an extraspecial 2-group the commutator is a nondegenerate alternating form with values in its derived subgroup of order two. Conversely, if G′G' has order two and G/Z(G)G/Z(G) is elementary abelian of order 22r2^{2r}, its commutator defines such a form. An isometry to the form of an extraspecial group of order 21+2r2^{1+2r} gives an isoclinism. We call this the binary symplectic model of dimension 2r2r.

By Zarrin’s theorem [22 Theorem 1.1(2)], a group of finite clique number is isoclinic to a finite group. Isoclinism preserves zz and ω\omega, since the central quotient and the commutation of its lifts determine them. To see that it preserves aa, enlarge each member of an abelian cover by the centre. The resulting subgroups are exactly the lifts of subgroups of the central quotient on which the commutator is trivial. We may therefore work with finite groups throughout.

Write N=ω(G)N=\omega(G). A noncommuting pair x,yx,y extends to a clique x,y,xyx,y,xy. Thus clique number two never occurs, and clique number one characterizes abelian groups. We use [x,y]=x−1y−1xy[x,y]=x^{-1}y^{-1}xy and write CmC_m for a cyclic group of order mm. The derived subgroup is G′=[G,G]G'=[G,G]; PP is generated by the pp-th powers. Centralizers of subsets are taken pointwise. Vector-space dimensions are over the indicated field; a line and a plane have vector dimensions one and two. Projective dimensions are specified explicitly. Orthogonality refers to the chosen scalar form unless a subscript specifies another form. Logarithms without a subscript are natural.

Cliques and covers in cosets

For a subset XX of a group, write ω(X)\omega(X) for the largest size of a clique contained in XX. We call ω(gU)\omega(gU) the width of the coset gUgU.

Lemma 2.1 (The abelian coset formula). If UU is abelian, then

ω(gU)=[U:CU(g)].(1)\omega(gU)=[U:C_U(g)]. \tag*{(1)}

Moreover, gUgU can be covered by exactly that many abelian subgroups.

Proof. For a,b∈Ua,b\in U, the elements ga,gbga,gb commute if and only if g(ab−1)=(ab−1)gg(ab^{-1})=(ab^{-1})g, by cancelling the two products and using commutativity in UU. Thus they commute exactly when a,ba,b lie in the same coset of CU(g)C_U(g) in UU. A clique takes at most one element from each resulting class, and any such choice is a clique. Each class consists of commuting elements and generates an abelian subgroup. Conversely, the clique gives a lower bound on the number of abelian subgroups in a cover.

If HH has an abelian subgroup of index II, partition gHgH into II cosets of that subgroup. Each has width at most ω(gH)\omega(gH). By the lemma, each can be covered by at most that many abelian subgroups. Thus gHgH costs at most Iω(gH)I\omega(gH). Throughout, the cost of a cover means its number of abelian subgroups.

Lifting a subgroup cover

For nonabelian GG, fix an abelian cover of a subgroup U≤GU \le G. For each coset gU≠UgU \ne U, its translates gAgA, with AA a member of this cover, cover gUgU. Each translate costs at most N−1N-1 subgroups. If gg fails to centralize AA, an element of AA extends a maximum clique in gAgA. If gg centralizes AA, the translate costs one. Thus

a(G)≤a(U)[1+([G:U]−1)(N−1)].(2)a(G) \le a(U)\left[1 + ([G:U]-1)(N-1)\right]. \tag*{(2)}

Adding widths across fibres

One fibre per nontrivial cyclic subgroup of the quotient suffices for the cover. To control their total width, we average over families whose cliques can be joined.

Lemma 2.2 (Clique and cover bounds in a quotient). Let π:G→Q\pi:G \to Q be a surjection with kernel HH. Write w(q)=ω(π−1(q))w(q)=\omega(\pi^{-1}(q)), and let II be the index of an abelian subgroup of HH.

(i) Two generators of the same cyclic subgroup of QQ have fibres of the same width, covered by the same families of abelian subgroups.

(ii) If every point of Xi⊆QX_i \subseteq Q fails to commute with every point of Xj⊆QX_j \subseteq Q for i≠ji \ne j, then

∑iω(π−1(Xi))≤ω(G).\sum_i \omega(\pi^{-1}(X_i)) \le\omega(G).

(iii) If D⊆QD \subseteq Q contains a generator of every nontrivial cyclic subgroup of QQ, then

a(G)≤a(H)+I∑q∈Dw(q).a(G) \le a(H) + I\sum_{q\in D} w(q).

Suppose a random subset of DD contains each q∈Dq \in D with probability at least δ>0\delta> 0. If lifts of its distinct points are always noncommuting, then ∑q∈Dw(q)≤ω(G)/δ\sum_{q\in D} w(q) \le\omega(G)/\delta.

Proof. For (i), choose an exponent inducing the required power on the cyclic subgroup and coprime to ∣G∣|G|, using the Chinese remainder theorem. The power map is a bijection between the fibres and preserves the cyclic subgroup generated by each element. This preserves both commutation and membership in abelian subgroups. For (ii), take a maximum clique in each inverse image. Their union is a clique in GG. For (iii), cover HH and each selected fibre, then use (i). For every realization of the random subset, the cliques in its fibres can again be joined. Their total size is at most ω(G)\omega(G), and its expectation is at least δ∑q∈Dw(q)\delta\sum_{q\in D}w(q).

Lifting the cyclic subgroups of G/Z(G)G/Z(G) gives a(G)≤max⁡(1,z(G)−1)a(G) \le\max(1,z(G)-1). For a pp-group of central index pk>1p^k>1 this cost is at most

Cp(k):=pk−1p−1,C_p(k):=\frac{p^k-1}{p-1},

since each nontrivial cyclic subgroup has at least p−1p-1 generators.

Centralizers in a narrow coset

The factor II makes the quotient estimate useful only if the kernel has a large abelian subgroup. To find one, we look inside a narrow coset. The centralizers of a maximum clique cover that coset, so one of them meets a large part of the kernel. The same cover also controls the intersection of the centralizers.

Lemma 2.3 (Centralizer partitions). Let DD be a maximum clique in a finite subset XX of GG. The sets X∩CG(x)X \cap C_G(x), x∈Dx \in D, cover XX. If the sum of their sizes is at most ∣X∣|X|, they partition XX into commuting sets.

Proof. An uncovered point would extend DD. If the sizes sum to at most ∣X∣|X|, the cover is a partition. A noncommuting pair in the part at xx would replace xx and enlarge DD, since neither point commutes with any other vertex.

We call a cover irredundant if deleting any member leaves an uncovered point.

Lemma 2.4 (Centralizers in a coset). Let U≤GU \le G, g∈Gg \in G, and let XX be a maximum clique in gUgU. Then

U∩CG(X)=Z(U)∩CU(g).U \cap C_G(X) = Z(U) \cap C_U(g).

The sets CG(x)∩gUC_G(x) \cap gU, x∈Xx \in X, form an irredundant coset cover. Each x∈gUx \in gU satisfies ω(CU(x))≤ω(gU)\omega(C_U(x)) \le\omega(gU), and some x∈Xx \in X satisfies [U:CU(x)]≤∣X∣[U : C_U(x)] \le|X|. If UU is a nonabelian pp-group, then ω(gU)≥p+1\omega(gU) \ge p + 1.

Proof. Suppose a∈Ua \in U centralizes XX but fails to commute with y∈gUy \in gU. Replace each vertex xx that commutes with yy by xaxa. The replacements remain in the coset and preserve mutual commutators because aa centralizes XX. Now no vertex commutes with yy, so we can adjoin yy, a contradiction. Thus aa centralizes all of gUgU, which is equivalent to belonging to Z(U)∩CU(g)Z(U) \cap C_U(g).

The parts of the cover in Lemma 2.3 are xCU(x)xC_U(x), and each vertex belongs only to its own part. At least one part has size ∣U∣/∣X∣|U|/|X| or more, giving the index bound. Multiplication by xx carries a clique in CU(x)C_U(x) into gUgU with its commutation relations unchanged. Translating the cover by g−1g^{-1} on the left and then inverting gives an irredundant cover of UU by right cosets of the CU(x)C_U(x).

If UU is a nonabelian pp-group, a one-part cover would contradict the intersection formula. Each subgroup in the cover is therefore proper. At most pp parts could cover only by partitioning UU into pp cosets of index pp. Two distinct index-pp subgroups are normal and multiply to UU, so any two of their cosets intersect. Hence all the subgroups in the partition would coincide. Their intersection lies in Z(U)Z(U), which would make U/Z(U)U/Z(U) cyclic.

Tomkinson’s bound for irredundant coset covers [20] now gives

[U:Z(U)∩CU(g)]≤ω(gU)!.(3)[U : Z(U) \cap C_U(g)] \le\omega(gU)! . \tag*{(3)}

In particular a coset of a nonabelian subgroup has width at least three.

Suppose x∈gUx \in gU centralizes an abelian subgroup AA of index II in UU. We can use the fibre width to cover UU itself. For each u∈Uu \in U, the elements of AA commuting with uu are exactly those commuting with xuxu. The abelian coset formula then gives

ω(uA)=[A:CA(u)]=[A:CA(xu)]=ω(xuA)≤ω(gU).\omega(uA) = [A : C_A(u)] = [A : C_A(xu)] = \omega(xuA) \le\omega(gU).

Covering the II cosets of AA in UU therefore costs at most Iω(gU)I\omega(gU).

Symplectic geometry and the sharp examples

When an alternating form determines commutation in a central quotient, cliques bound the admissible profile and isotropic subspaces lift to abelian subgroups. We construct both.

An alternating form bb is symplectic if the only vector pairing to zero with every vector is zero. A clique for bb is a set whose distinct vectors pair nontrivially. In dimension two, the basic example is the determinant pairing. Two nonzero vectors pair to zero exactly when they lie on the same line. Thus a clique is obtained by choosing one nonzero vector on each line. The following lemma extends this count to larger dimensions. For odd primes, the lower bound pr+1pr+1 is due to Chin [5], in the equivalent formulation for extraspecial pp-groups. The binary maximum is due to Isaacs [3 p. 40]. We include proofs for completeness.

Lemma 3.1 (Symplectic cliques). Let (V,b)(V,b) be symplectic of dimension 2r>02r > 0 over Fp\mathbb{F}_p. There is a clique of size pr+1pr+1. For p=2p=2, the maximum is 2r+12r+1, and every maximum clique spans VV and has sum zero.

Proof. To combine cliques X,YX,Y in orthogonal summands, choose y0∈Yy_0 \in Y and use (X+y0)∪(Y∖{y0})(X+y_0)\cup(Y\setminus\{y_0\}). The translate retains its original pairings, and its pairings with the remaining vertices of YY are those of y0y_0. The new clique has size ∣X∣+∣Y∣−1|X|+|Y|-1. Starting with p+1p+1 vertices per symplectic plane gives pr+1pr+1 vertices.

Over F2\mathbb{F}_2, a clique of size kk has a Gram matrix with diagonal zero and all other entries one. Its kernel consists of the vectors whose entries all equal their sum. The rank is therefore kk for even kk and k−1k-1 for odd kk. It cannot exceed 2r2r, so k≤2r+1k\leq2r+1. At equality the rank is 2r2r, so the clique spans VV. Its sum is orthogonal to the clique and is therefore zero.

We call a maximum clique for a binary symplectic form a symplectic simplex. A subspace on which an alternating form vanishes on every pair is called totally isotropic. When commutation in a central quotient is determined by such a form, the abelian subgroups containing the centre are exactly the lifts of these subspaces.

Lemma 3.2 (A cover from a symplectic spread). Suppose the central quotient of GG is a vector space of dimension 2r>02r>0 over Fp\mathbb{F}_p, and commutation of lifts is determined by a symplectic form on that space. Then a(G)=pr+1a(G)=p^r+1.

Proof. A totally isotropic subspace UU satisfies U≤U⊥U\leq U^\perp, so dim⁡U≤r\dim U\leq r. It contains at most pr−1p^r-1 nonzero vectors. Dividing the total p2r−1p^{2r}-1 by pr−1p^r-1 gives the lower bound pr+1p^r+1.

For the cover, write K=FprK=\mathbb{F}_{p^r} and view K2K^2 as a space over Fp\mathbb{F}_p. Use the form

((x1,x2),(y1,y2))⟼Tr⁡K/Fp(x1y2−x2y1).((x_1,x_2),(y_1,y_2))\longmapsto\operatorname{Tr}_{K/\mathbb{F}_p}(x_1y_2-x_2y_1).

The trace pairing is nondegenerate, so this form is symplectic. Every KK-line is totally isotropic because the determinant vanishes on it. The pr+1p^r+1 lines partition the nonzero vectors of K2K^2. Such a partition into totally isotropic subspaces of dimension rr is called a symplectic spread. All nondegenerate alternating forms over Fp\mathbb{F}_p of dimension 2r2r have symplectic bases, so the partition gives a spread in the original quotient. Lifting its members gives the required cover.

Corollary 3.3 (Finite-field determinant models). Every finite-field determinant model with central quotient K2K^2 satisfies a(G)=∣K∣+1=1+z(G)a(G)=|K|+1=1+\sqrt{z(G)}.

Proof. Write ∣K∣=pr|K|=p^r. The ∣K∣+1|K|+1 lines over KK are isotropic for the determinant and hence for its image under ρ\rho. Their preimages form an abelian cover. For the converse, choose a linear functional λ\lambda on the derived subgroup such that λρ≠0\lambda\rho\neq0. The form λρ∘det⁡\lambda\rho\circ\det is nondegenerate over Fp\mathbb{F}_p: for v≠0v\neq0, the map w↦det⁡(v,w)w\mapsto\det(v,w) is onto KK. Every abelian subgroup, enlarged by the centre, therefore has isotropic image of dimension at most rr. Counting nonzero vectors as in Lemma 3.2 requires at least (p2r−1)/(pr−1)=pr+1(p^{2r}-1)/(p^r-1)=p^r+1 members.

Groups attaining the bounds

The constructions above give all lower bounds in Theorem 1.1. Let kk be the allowed clique size.

For k=1,2k=1,2 the abelian model attains the bound. For 3≤k≤63\leq k\leq6, take UT⁡3(Fq)\operatorname{UT}_3(\mathbb{F}_q), the group of upper unitriangular matrices, with q=k−1∈{2,3,4,5}q=k-1\in\{2,3,4,5\}. Its central quotient is Fq2\mathbb{F}_q^2, and the commutator is the determinant xy′−x′yxy'-x'y. The q+1q+1 lines lift to an abelian cover; one nonzero point on each line gives a clique of size q+1=kq+1=k.

For k=7,8k=7,8, use an extraspecial 3-group with a four-dimensional symplectic central quotient. By the duality between the symplectic generalized quadrangle W(3,q)W(3,q) and the parabolic quadric Q(4,q)Q(4,q), a set of pairwise nonorthogonal projective points corresponds to a partial spread of the quadric. Tallini’s bound [19]2, together with the nonexistence of ovoids in W(3,q)W(3,q) for odd qq, bounds its size by q2−q+1q^2-q+1; see [12], Theorem 3.3 and § 6. At q=3q=3 this is seven, attained by combining two plane cliques. The clique number is therefore seven, and the spread in Lemma 3.2 gives an optimal cover of size ten.

For k≥9k \ge9, take an extraspecial 2-group of order 21+2r2^{1+2r} with r=⌊(k−1)/2⌋r = \lfloor(k-1)/2 \rfloor. Its central index is 4r4^r, its clique number is 2r+12r+1, and its covering number is 2r+12^r+1, by Lemmas 3.1 and 3.2.

The scalar quotient of a pp-group

A finite nilpotent group is a product of its Sylow subgroups. We first cover each prime-power factor using a scalar commutator form. We will combine their covers at the end of § 5.

Case map. For a scalar radical RR, write d=dim⁡(P/R)=2jd = \dim(P/R) = 2j. The binary cases are treated in § 5; the odd-prime cases below are organized as follows. In the last row, [R:Z(R)]=pm[R : Z(R)] = p^m.

Radical and primeRange or constraintCovering argument
Any radical, p=2p = 2All NNBinary case map in § 5.
Abelian noncentral, pp oddj=1j = 1; or N≥p2j+1N \ge p^2j + 1Lemma 4.2; otherwise the fibre average in § 4.2.
Central, one component, pp oddj=1j = 1; or p≥5p \ge5, j≥2j \ge2The spread of Lemma 3.2, with N≥pj+1N \ge pj + 1.
Central, one component, p=3p = 3j=2j = 2; or j≥3j \ge3The ten-member cover; otherwise thirteen-point cliques and the spread.
Central, several components, pp oddN≥p2(j−1)+1N \ge p^2(j - 1) + 1Lemma 4.6; the common centralizer planes in dimension four.
Nonabelian, pp oddN≥(pj+1)max⁡{p+1,m+1}N \ge(p^j + 1)\max\{p+1,m+1\}Lemma 4.5 and the fibre average.

Table 2.

Proposition 4.1 (Nilpotent groups). If GG is nilpotent and N=ω(G)N = \omega(G), then a(G)≤T(N)a(G) \le T(N). For N≥9N \ge9, equality requires the binary symplectic model of dimension N−1N - 1.

The quotient given by one commutator form

Let PP be a nonabelian pp-group and set N=ω(P)N = \omega(P). A single scalar commutator value will suffice to certify noncommutation of lifts.

The commutators are central modulo γ3(P)=[P′,P]\gamma_3(P) = [P',P]. The abelian pp-group P′/γ3(P)P'/\gamma_3(P) is nontrivial: otherwise the lower central series would be constant from P′P', contrary to nilpotence. Choose a nonzero homomorphism ξ:P′/γ3(P)→(Fp,+)\xi: P'/\gamma_3(P) \to(\mathbb{F}_p,+), and also write ξ\xi for its composite with the quotient map on P′P'.

Applying ξ\xi to commutators gives the pairing

bξ(x,y)=ξ([x,y]).b_{\xi}(x,y) = \xi([x,y]).

Centrality of the commutators in P/γ3(P)P/\gamma_3(P) makes the pairing additive in each argument. Its radical is the subgroup

R=Rξ={x∈P:bξ(x,y)=0 for all y∈P}.R = R_{\xi} = \{x \in P : b_{\xi}(x,y) = 0 \text{ for all } y \in P\}.

It is normal and contains P′PpZ(P)P'P^pZ(P). Hence V=P/RV = P/R is a vector space over Fp\mathbb{F}_p, and the induced form is symplectic by the definition of RR. We call bξb_{\xi} a scalar commutator form and RR its scalar radical. A nonzero pairing in VV certifies noncommutation of every pair of lifts, so a scalar clique lets us add its fibre widths.

Scalar radicals need not be central, even in class two: RξR_{\xi} tests only ξ([x,y])=0\xi([x,y]) = 0, whereas Z(P)Z(P) requires [x,y]=1[x,y] = 1 for every yy. For example, in D8×D8D_8 \times D_8, projecting the commutator onto the first component gives the nonabelian radical Z(D8)×D8Z(D_8) \times D_8; here D8D_8 is the dihedral group of order eight.

Lemma 4.2 (The terminal rank-two case). If a scalar commutator form of PP has rank two and an abelian radical RR, then a(P)=ω(P)a(P)=\omega(P).

Proof. For x∉Rx\notin R, the perpendicular of xRxR in the symplectic plane P/RP/R is its own line. Thus CP(x)=⟨x,CR(x)⟩C_P(x)=\langle x,C_R(x)\rangle is abelian. Two commuting elements outside RR have the same centralizer: each centralizer is abelian and hence contained in the other.

Choose one representative outside RR for each distinct centralizer. They form a clique, and their centralizers cover P∖RP\setminus R. If they cover RR as well, the clique and cover have the same size. Otherwise an uncovered point of RR extends the clique, and adding RR extends the cover. Hence a(P)=ω(P)a(P)=\omega(P).

Adding the widths of the fibres

Write d=2jd=2j for the dimension of VV. Select one nonzero vector on each line LL of VV. Its fibre has width wLw_L, and a cover of that fibre serves the whole line by Lemma 2.2(i). To bound the total width, take the lines of a symplectic clique of size pj+1p^j+1 and move them by a uniformly chosen symplectic transformation. The selected vectors on the resulting lines still form a clique. Transitivity on lines gives inclusion probability (pj+1)/Cp(2j)(p^j+1)/C_p(2j). Lemma 2.2(iii) now gives

∑LwL≤Cp(2j)Npj+1.\sum_L w_L \le\frac{C_p(2j)N}{p^j+1}.

If RR has an abelian subgroup of index II, the same lemma gives the common bound

a(P)≤a(R)+ICp(2j)Npj+1.(4)a(P)\le a(R)+I\frac{C_p(2j)N}{p^j+1}. \tag*{(4)}

Lemma 4.3 (A clique meeting a hyperplane once). In a symplectic space (V,b)(V,b) of dimension 2r>02r>0 over Fp\mathbb{F}_p, every hyperplane meets some clique of size pr+1p^r+1 in exactly one point.

Proof. For a prescribed hyperplane e⊥e^\perp, choose ff with b(e,f)=1b(e,f)=1, and let V0V_0 be the orthogonal complement of ⟨e,f⟩\langle e,f\rangle. Take a clique D0D_0 of size pr−1+1p^{r-1}+1 in V0V_0, with D0={0}D_0=\{0\} if r=1r=1. Then

(D0+f)∪{e+tf:t∈Fp}(D_0+f)\cup\{e+tf:t\in\mathbb{F}_p\}

is a clique. The first part inherits the pairings of D0D_0. In the second part they are differences of distinct coefficients of ff, and the cross pairings are all b(f,e)b(f,e). Only ee lies in e⊥e^\perp.

Lemma 4.4 (Coset covers of pp-groups). If UU is a finite pp-group and the right cosets HixiH_ix_i, 1≤i≤t1\le i\le t, form an irredundant cover of UU, then [U:⋂iHi]≤pt−1[U:\bigcap_i H_i]\le p^{t-1}.

Proof. We argue by induction on tt. At each step, we restrict the cover to a coset disjoint from the members already selected. The restricted cover uses fewer members, so induction controls the index of the next intersection. The case t=1t=1 is immediate.

For t>1t>1, all HiH_i are proper. Choose a maximal subgroup MM containing H1H_1. It is normal of index pp. Right-translate the cover so that H1x1H_1x_1 misses MM, and choose an irredundant subcover of MM, with index set I1I_1. Put B1=⋂i∈I1HiB_1=\bigcap_{i\in I_1}H_i. Since 1∉I11\notin I_1, induction gives [M:M∩B1]≤p∣I1∣−1[M:M\cap B_1]\le p^{|I_1|-1}.

Write D=⋂iHiD=\bigcap_iH_i. If the intersection BjB_j of the subgroups selected so far is not yet DD, the selected cosets do not cover UU, by irredundancy. Choose a point yy that they miss. They miss the whole coset BjyB_jy, because BjB_j lies in every selected subgroup. Restrict the remaining cover to BjyB_jy and translate by y−1y^{-1}. An irredundant subcover, indexed by Ij+1I_{j+1}, consists of cosets of Bj∩HiB_j\cap H_i in BjB_j. Its indices are all new, so induction bounds the next intersection by

[Bj:Bj+1]≤p∣Ij+1∣−1,Bj+1=Bj∩⋂i∈Ij+1Hi.[B_j:B_{j+1}]\le p^{|I_{j+1}|-1},\qquad B_{j+1}=B_j\cap\bigcap_{i\in I_{j+1}}H_i.

Each step adds a nonempty set of previously unused indices, so the process terminates at Bs=DB_s=D. Since D≤MD\le M, multiplying the index bounds gives

[U:D]≤p[M:M∩B1][B1:D]≤p1+∑j=1s(∣Ij∣−1).[U:D]\le p[M:M\cap B_1][B_1:D]\le p^{1+\sum_{j=1}^{s}(|I_j|-1)}.

If the index 11 remains unused, then ∑j∣Ij∣≤t−1\sum_j|I_j|\le t-1 and s≥1s\ge1, so the exponent is at most t−1t-1. If it is used, then s≥2s\ge2, since 1∉I11\notin I_1, and ∑j∣Ij∣≤t\sum_j|I_j|\le t gives the same bound.

Lemma 4.5 (Fibres over a noncentral radical). Let RR be the radical of a scalar commutator form of rank 2j2j on PP, and put σ=pj+1\sigma=p^j+1.

  1. If RR is abelian and noncentral, then ω(P)≥pj+1\omega(P)\ge p^j+1.

  1. If RR is nonabelian and [R:Z(R)]=pm[R:Z(R)]=p^m, then the width of each fibre gRgR is at least p+1p+1 and at least m+1+log⁡p[Z(R):CZ(R)(g)]m+1+\log_p[Z(R):C_{Z(R)}(g)]. Consequently ω(P)≥σmax⁡{p+1,m+1}\omega(P)\ge\sigma\max\{p+1,m+1\}, and ω(P)≥σ(m+1)+1\omega(P)\ge\sigma(m+1)+1 if Z(R)≠Z(P)Z(R)\ne Z(P).

Proof. For (1), choose a∈R∖Z(P)a\in R\setminus Z(P) and a hyperplane of P/RP/R containing CP(a)/RC_P(a)/R. By Lemma 4.3, some symplectic clique of size σ\sigma meets the hyperplane once. Each of its pjp^j other vertices has a fibre of width at least pp by (1), since every lift moves aa. The remaining fibre contributes at least one.

For (2), let ww be the width of gRgR. Lemma 2.4 gives w≥p+1w\ge p+1 and an irredundant cover of RR by ww cosets. Their subgroups intersect in Z(R)∩CR(g)Z(R)\cap C_R(g), of index at most pw−1p^{w-1} by Lemma 4.4. This index is also pm[Z(R):CZ(R)(g)]p^m[Z(R):C_{Z(R)}(g)], giving the second width bound. Summing the widths over a clique of size σ\sigma gives ω(P)≥σmax⁡{p+1,m+1}\omega(P)\ge\sigma\max\{p+1,m+1\}. If Z(R)≠Z(P)Z(R)\ne Z(P), choose gg moving a point of Z(R)Z(R). Then g∉Rg\notin R and gRgR has width at least m+2m+2. Some symplectic clique of size σ\sigma contains gRgR, by transitivity on nonzero vectors, giving the last bound.

Odd primes

Choose a scalar form of least nonzero rank. The fibre average will suffice except in the central ternary cases, where we use larger cliques or a common isotropic spread.

Abelian noncentral radical. Lemma 4.2 gives a(P)=Na(P)=N when j=1j=1. For j≥2j\ge2, Lemma 4.5 gives N≥pj+1N\ge p^j+1. To compare the fibre-cover bound with the profile, note that N2−N/2N2^{-N/2} decreases for N≥3N\ge3, since each step multiplies it by (1+1/N)/2<1(1+1/N)/\sqrt{2}<1. Using B(N)≥2N/2−1B(N)\ge2^{N/2-1}, we may compare at N=pj+1N=p^j+1. There (pj+1)/(pj+1)<p(p^j+1)/(pj+1)<p, so

a(P)−1B(N)<2pp−1(p22p2/2)j<1.\frac{a(P)-1}{B(N)}<\frac{\sqrt{2}p}{p-1}\left(\frac{p^2}{2p^{2/2}}\right)^j<1.

The factor in parentheses is at most 1/21/2. For p=3p=3 this follows from 9<27/29<2^{7/2}. For p≥5p\ge5 use p2<2p≤2pp/2−1p^2<2p\le2p^{p/2-1}. The first inequality holds at 55 and persists in steps of two, since p+2<2pp+2<2p.

Central radical and one commutator component. If R=Z(P)R=Z(P), then PP has class two and P/Z(P)P/Z(P) is elementary abelian. Its commutator subgroup is elementary abelian as well, because [x,y]p=[xp,y]=1[x,y]^p=[x^p,y]=1. Write β\beta for the alternating commutator map on P/Z(P)P/Z(P) with values in P′P'. Composing β\beta with a linear functional on P′P' gives a scalarization of β\beta. When P′P' has dimension one, a symplectic spread gives exactly pj+1p^j+1 subgroups. In a scalar plane the spread costs exactly N=p+1N=p+1. For j≥2j\ge2 and p≥5p\ge5, we compare the spread size with the profile at the scalar clique bound. If jj is even, p2<2jp^2<2^j gives pj<B(pj+1)p^j<B(pj+1). If jj is odd, j≥3j\ge3 and

pjB(pj+1)=2(p2−p/2)j≤2(5 2−5/2)3=125128.\frac{p^j}{B(pj+1)}=\sqrt{2}\left(p2^{-p/2}\right)^j\le\sqrt{2}\left(5\,2^{-5/2}\right)^3=\frac{125}{128}.

Here p2−p/2p2^{-p/2} decreases as pp increases in steps of two.

At p=3p=3 and j=2j=2, the spread costs ten and the clique number is seven by § 3.1. For j≥3j \ge3, we need larger cliques than the plane count provides. Ceria, De Beule, Pavese and Smaldore [4 Theorem 3.7] give thirteen projective points forming a clique in dimension six over F3\mathbb{F}_3: more generally, for c∈Fq×c \in\mathbb{F}_q^{\times}, their q2+q+1q^2+q+1 points

Px=(x,xq,xq2,cx−q2,cx−q,cx−1),x∈Fq3×,Nm⁡Fq3/Fq(x)=1,P_x=(x,x^q,x^{q^2},cx^{-q^2},cx^{-q},cx^{-1}),\qquad x\in\mathbb{F}_{q^3}^{\times},\qquad\operatorname{Nm}_{\mathbb{F}_{q^3}/\mathbb{F}_q}(x)=1,

lie in the Fq\mathbb{F}_q-subgeometry {[a:aq:aq2:bq2:bq:b]:(a,b)≠(0,0)}\{[a:a^q:a^{q^2}:b^{q^2}:b^q:b]:(a,b)\ne(0,0)\}, where a,b∈Fq3a,b\in\mathbb{F}_{q^3}, of PG⁡(5,q3)\operatorname{PG}(5,q^3) and are pairwise nonorthogonal for the alternating form b(v,w)=∑i=13(viw7−i−v7−iwi)b(v,w)=\sum_{i=1}^{3}(v_iw_{7-i}-v_{7-i}w_i), since

b(Px,Py)=cTr⁡Fq3/Fq(x/y−y/x)=cNm⁡Fq3/Fq(1−y/x)≠0(x≠y).b(P_x,P_y)=c\operatorname{Tr}_{\mathbb{F}_{q^3}/\mathbb{F}_q}(x/y-y/x)=c\operatorname{Nm}_{\mathbb{F}_{q^3}/\mathbb{F}_q}(1-y/x)\ne0\qquad(x\ne y).

Taking q=3q=3 gives the required thirteen points. Write j=3u+vj=3u+v, with 0≤v≤20\le v\le2, and combine uu such cliques with vv plane cliques by Lemma 3.1. This gives N≥12u+3v+1≥4j−1N\ge12u+3v+1\ge4j-1. Thus B(N)≥4j/2>3jB(N)\ge4^{j/2}>3^j, since 43/2>334^{3/2}>3^3 and the bound grows faster than the spread cost.

Central radical and several commutator components. A subspace isotropic for one form may contain points with noncommuting lifts. The cover must therefore use subspaces isotropic for every commutator form. To compensate for any increase in their number, we use a second scalarization to construct a larger clique. The first form separates the fibres, and the second lets us take a whole line of points in most of them.

Lemma 4.6 (Cliques from two commutator forms). Let β:V×V→W\beta:V\times V\to W be alternating bilinear over Fp\mathbb{F}_p, with values spanning WW, dim⁡V=2r\dim V=2r and dim⁡W≥2\dim W\ge2. If some scalarization b=λ∘βb=\lambda\circ\beta is nondegenerate, then r≥2r\ge2 and there are p2(r−1)+1p^{2(r-1)}+1 vectors whose β\beta-values on distinct pairs are all nonzero.

Proof. Take an independent functional μ\mu on WW and put c=μ∘βc=\mu\circ\beta. Write c(v,x)=b(Sv,x)c(v,x)=b(Sv,x). The operator SS is nonscalar because the values of β\beta span WW. Thus some v,Svv,Sv are independent, since an operator preserving every line is scalar.

On the symplectic quotient v⊥b/⟨v⟩v^{\perp_b}/\langle v\rangle, the functional c(v,⋅)c(v,\mathord\cdot) is well defined and nonzero, since Sv∉⟨v⟩Sv\notin\langle v\rangle. This also shows that the quotient has positive dimension, so r≥2r\ge2. Choose a clique of size pr−1+1p^{r-1}+1 meeting its kernel once, by Lemma 4.3. Lift its vertices to v⊥bv^{\perp_b} and replace every lift xx outside the kernel by the line x+Fpvx+\mathbb{F}_pv. The bb-pairings between different fibres are still nonzero. Within a line the cc-pairing is a nonzero multiple of c(v,x)c(v,x), so all its points can be used. This gives 1+p2(r−1)1+p^2(r-1) vertices.

Applying the lemma to β\beta gives N≥p2(j−1)+1N\ge p^2(j-1)+1. Lifting the lines gives a cover of size Cp(2j)C_p(2j). For p≥5p\ge5 and j≥3j\ge3, the inequality p2<2pp^2<2^p gives p2j<2pjp^{2j}<2^{pj}. Since j≥2j\ge2 and p≥5p\ge5, p2(j−1)≥2pjp^2(j-1)\ge2pj, so the line count is below B(N)B(N). For p=3,j≥3p=3,j\ge3, the bound N≥9(j−1)+1N\ge9(j-1)+1 gives B(N)≥2916j−3B(N)\ge2^9 16^{j-3}, while the line count is less than (93/22)9j−3(9^{3/2}2)9^{j-3}.

Let j=2j=2. The least-rank choice implies that every nonzero scalarization of β\beta on the four-space is nondegenerate. The Pfaffian bound [10], Lemma 3 gives dim⁡P′≤2\dim P'\le2, and hence dim⁡P′=2\dim P'=2. For x≠0x\ne0, the map β(x,⋅):V→P′\beta(x,\mathord\cdot):V\to P' is onto: a functional annihilating its image would give a nonzero scalarization with xx in its radical. Thus Cβ(x)=ker⁡β(x,⋅)C_\beta(x)=\ker\beta(x,\mathord\cdot) is a plane. It is isotropic for β\beta, since it is spanned by xx and by a vector commuting with xx. For each nonzero y∈Cβ(x)y\in C_\beta(x) we then have Cβ(x)≤Cβ(y)C_\beta(x)\le C_\beta(y), with equality by dimension. The planes partition the nonzero vectors of VV; their p2+1p^2+1 lifts give an abelian cover. Since N≥p2+1≥10N\ge p^2+1\ge10 by Lemma 4.6, this cover is strictly below the profile.

Nonabelian radical. Write [R:Z(R)]=pm[R : Z(R)] = p^{m} with m≥2m \ge2, and put σ=pj+1\sigma= pj + 1 and M=max⁡{p+1,m+1}M = \max\{p + 1, m + 1\}. Lemma 4.5 gives N≥σMN \ge\sigma M. At most Cp(m)C_p(m) cyclic lifts cover RR, and one such lift has index at most pm−1p^{m-1} in RR. Since Cp(m)<23pm−1≤pm−1N/σC_p(m) < \frac{2}{3}p^{m-1} \le p^{m-1}N/\sigma, (4) gives

a(P)≤Cp(m)+pm−1Cp(2j)Nσ≤pm−1(1+Cp(2j))Nσ.a(P) \le C_p(m) + p^{m-1}\frac{C_p(2j)N}{\sigma} \le p^{m-1}(1+C_p(2j))\frac{N}{\sigma}.

The ratio of the last bound to 2N/2−12^{N/2-1} is largest at N=σMN = \sigma M, because N2−N/2N2^{-N/2} decreases. Since 1+Cp(2j)≤(p+2)p2j−21+C_p(2j) \le(p+2)p^{2j-2}, this gives

a(P)B(N)≤ρθj−1,ρ=2M(p+2)pm−12−(p+1)M/2,θ=p22−pM/2.\frac{a(P)}{B(N)} \le\rho\theta^{j-1}, \qquad\rho= 2M(p+2)p^{m-1}2^{-(p+1)M/2}, \qquad\theta= p^2 2^{-pM/2}.

Here θ≤9/64\theta\le9/64: M≥4M \ge4, and p2/4pp^2/4^p decreases for odd p≥3p \ge3. For p≥5p \ge5, use m≤M−1m \le M-1 and p≤582(p+1)/2p \le\frac{5}{8}2^{(p+1)/2}, which holds at 55 and persists in steps of two. Then

ρ≤2(p+2)p2M(5/8)M≤14256(5/8)6<1,\rho\le\frac{2(p+2)}{p^2}M(5/8)^M \le\frac{14}{25}6(5/8)^6 < 1,

since the two factors decrease for p≥5p \ge5 and M≥6M \ge6, respectively. If p=3,m=2p=3,m=2, we have ρ=120/28<1\rho=120/2^8<1. If p=3,m≥3p=3,m\ge3, then M=m+1M=m+1 and ρ=(10/9)M(3/4)M≤45/32\rho=(10/9)M(3/4)^M\le45/32, since M(3/4)MM(3/4)^M decreases for M≥4M\ge4. This suffices when j≥2j\ge2, as (45/32)(9/64)<1(45/32)(9/64)<1.

It remains to take p=3,j=1,m≥3p=3,j=1,m\ge3. Here σ=C3(2)=4\sigma=C_3(2)=4, and the first cost bound gives a(P)<3m−1(N+2)a(P)<3^{m-1}(N+2). If Z(R)≠Z(P)Z(R)\ne Z(P), Lemma 4.5 gives N≥4m+5N\ge4m+5. At N=4m+6N=4m+6, the inequality 3m−1(N+2)≤B(N)3^{m-1}(N+2)\le B(N) reduces to 3m−1(m+2)≤4m3^{m-1}(m+2)\le4^m. This holds at m=3m=3 and persists since 3(m+3)≤4(m+2)3(m+3)\le4(m+2). It also holds at N=4m+5N=4m+5, where BB has the same value, and persists in steps of two, since BB doubles. If Z(R)=Z(P)Z(R)=Z(P), then z(P)=3m+2z(P)=3^{m+2} and cyclic lifts give

a(P)≤C3(m+2)<3m+2/2≤22m+1≤B(N).a(P)\le C_3(m+2)<3^{m+2}/2\le2^{2m+1}\le B(N).

The middle inequality starts at m=3m=3 and persists as mm increases; the last uses N≥4m+4N\ge4m+4. Thus a(P)≤T(N)a(P)\le T(N) for odd pp. When N≥9N\ge9, the bounds above are all strict, including the costs NN and ten.

Binary groups and the equality model

Case map. Write ν=log⁡2z(P)\nu=\log_2 z(P). First use cyclic lifts when N≥2ν+1N\ge2\nu+1. In the remaining range N≤2νN\le2\nu, test the radicals in the order shown below. Here d=dim⁡(P/R)d=\dim(P/R).

Radical or commutatorRangeCovering argument
AnyN≥2ν+1N\ge2\nu+1Cyclic lifts; the small values N≤8N\le8 are included below.
Some nonabelian radicalN≤2νN\le2\nuThe centralizer-cover estimate in § 5.1.
All radicals abelian; rank two occursAll remaining NNLemma 4.2: a(P)=Na(P)=N.
Abelian noncentral radical; all ranks ≥4\ge4N≥2d+5N\ge2d+5The uniform fibre average.
SameN≤2d+4N\le2d+4Lemma 5.1; separate fibres, except for d=4d=4, N=9,10N=9,10, handled by Lemma 5.2.
All radicals central; one componentAll remaining NNLemma 3.2; the equality model.
All radicals central; several componentsν≥10\nu\ge10; or ν=4,6,8\nu=4,6,8The amplified clique forces the cyclic-lift range; otherwise Lemma 5.4.

Table 3.

For the first row, cyclic lifts give a(P)≤2ν−1<B(N)a(P) \le2^\nu- 1 < B(N) when N≥9N \ge9. If N≤8N \le8, then ν≤3\nu\le3: the cost is at most three for ν≤2\nu\le2, and seven for ν=3\nu= 3, which requires N≥7N \ge7.

In the remaining rows, a nonabelian radical raises every fibre width; an abelian noncentral radical raises widths in directions that act nontrivially. With central radicals, we use the commutator forms themselves.

A nonabelian radical

Suppose some scalar radical RR is nonabelian. Put A=Z(R)A = Z(R) and write [R:A]=2m[R : A] = 2^m and [A:Z(P)]=2s[A : Z(P)] = 2^s. Thus ν=d+m+s\nu= d + m + s, with d=dim⁡(P/R)≥2d = \dim(P/R) \ge2 and m≥2m \ge2.

The radical can be covered by 2m−12^m - 1 cyclic lifts, and one such lift gives an abelian subgroup of index at most 2m−12^{m-1} in RR. Thus the fibre bound gives

a(P)≤2m−1+2m−1(2d−1)Nd+1.a(P) \le2^m - 1 + 2^{m-1}\frac{(2^d - 1)N}{d + 1}.

By Lemma 4.5, the width of gRgR is at least m+1+log⁡2[A:CA(g)]m + 1 + \log_2[A : C_A(g)]. If A=Z(P)A = Z(P), the widths over a simplex give N≥(d+1)(m+1)≥2(d+m)+1=2ν+1N \ge(d + 1)(m + 1) \ge2(d + m) + 1 = 2\nu+ 1, since dm≥d+mdm \ge d + m. This contradicts N≤2νN \le2\nu.

Now A≠Z(P)A \ne Z(P). Conjugation on AA factors through P/RP/R. The simplex spans P/RP/R, so at least one of its lifts acts nontrivially on AA. Number the lifts g1,…,gd+1g_1,\ldots,g_{d+1} so that gd+1g_{d+1} is such a lift. The simplex vertices sum to zero, so the first dd still span P/RP/R. Thus ⋂i=1dCA(gi)=CA(P)=Z(P)\bigcap_{i=1}^{d} C_A(g_i) = C_A(P) = Z(P). A coset of this intersection is determined by its cosets modulo the CA(gi)C_A(g_i), giving

∏i=1d[A:CA(gi)]≥[A:Z(P)]=2s.\prod_{i=1}^{d} [A : C_A(g_i)] \ge[A : Z(P)] = 2^s.

The first dd fibres therefore contribute at least ss vertices beyond their baseline widths, and the last fibre contributes at least one more. Hence

N≥(d+1)(m+1)+s+1=ν+dm+2.N \ge(d + 1)(m + 1) + s + 1 = \nu+ dm + 2.

Together with N≤2νN \le2\nu, this gives N≥2dm+4≥4(d+m−1)N \ge2dm + 4 \ge4(d + m - 1), using (d−2)(m−2)≥0(d - 2)(m - 2) \ge0. Put h=d+m−1h = d + m - 1 and κ=⌊(N−1)/2⌋\kappa= \lfloor(N - 1)/2 \rfloor. The clique bound gives κ≥2h−1\kappa\ge2h - 1. The cover estimate and (2d−1)/(d+1)≤2d−2(2^d - 1)/(d + 1) \le2^{d-2} give

a(P)≤2h−2(N+2)−1≤2h−1(κ+2)−1.a(P) \le2^{h-2}(N + 2) - 1 \le2^{h-1}(\kappa+ 2) - 1.

It remains to compare κ+2\kappa+ 2 with 2κ−h+12^{\kappa-h+1}. At κ=2h−1\kappa= 2h - 1, the required inequality is 2h+1≤2h2h + 1 \le2^h, which holds since h≥3h \ge3. Thereafter 2κ−h+12^{\kappa-h+1} doubles with each unit of κ\kappa, whereas κ+2\kappa+ 2 grows by one. Thus a(P)<B(N)a(P) < B(N).

An abelian noncentral radical

Assume now that all scalar radicals are abelian. If some scalar form has rank two, we have a(P)=Na(P) = N by Lemma 4.2. For N≥9N \ge9, this is strictly below T(N)T(N). If all scalar ranks are at least four, choose a noncentral radical of rank dd, if one exists, and use (4) with a(R)=I=1a(R) = I = 1. Write κ=⌊(N−1)/2⌋\kappa= \lfloor(N - 1)/2 \rfloor, so the fibre cost divided by B(N)B(N) is at most

2κ+2d+12d−κ.\frac{2\kappa+ 2}{d + 1} 2^{d-\kappa}.

For κ≥d+2\kappa\ge d + 2 the ratio decreases with κ\kappa, starting at (d+3)/(2(d+1))≤7/10(d + 3)/(2(d + 1)) \le7/10. When κ≤d+1\kappa\le d + 1, we have N≤2d+4N \le2d + 4. The averaging bound alone is insufficient here, since it does not exclude individual fibres of large width. We first show that every fibre has width at most two. Covering the fibres separately then suffices except at d=4d = 4 with N=9,10N = 9, 10, where we build a cover of the central quotient by common isotropic subspaces.

Lemma 5.1 (Width reduction in small scalar rank). Let RR be an abelian noncentral scalar radical in a finite 22-group PP, with d=dim⁡(P/R)≥4d = \dim(P/R) \ge4. Assume all nonzero scalar forms have rank at least four. If N=ω(P)≤2d+4N = \omega(P) \le2d + 4, then d∈{4,6}d \in\{4,6\}, N≥9N \ge9, and every fibre of P/RP/R has width at most two.

Proof. Step 1: cliques from two fibres. Write RR additively and set Dg=αg−1D_g = \alpha_g - 1, where αg\alpha_g is right conjugation on RR. Since RR is abelian, this action factors through the elementary abelian group V=P/RV = P/R. In particular the actions commute, and the notation DvD_v, for v∈Vv \in V, is unambiguous. The width of the fibre at vv is ∣Dv(R)∣|D_v(R)|, by Lemma 2.1. The commuting classes in gRgR are the cosets of Kg=ker⁡DgK_g = \ker D_g. For distinct fibres yRyR, zRzR, two elements yryr, zszs commute exactly when

Dz(r)−Dy(s)=−[y,z].(5)D_z(r) - D_y(s) = -[y,z]. \tag*{(5)}

Indeed, yrzs=yzrzsyrzs = yz r^zs and zsyr=yz[y,z]−1syrzsyr = yz[y,z]^{-1}s^yr; all the factors after yzyz belong to the abelian subgroup RR. Suppose both fibres have width two and their kernels differ. Start with a noncommuting pair. Vary rr by an element of Kz∖KyK_z \setminus K_y and ss by an element of Ky∖KzK_y \setminus K_z. Neither change affects the equation, and each gives a second class in its own fibre. This yields a four-point clique.

More generally, fibres with different actions give at least three clique points. A fibre of width at least four suffices on its own. If one fibre has width one and the other width two, choose both classes in the latter. Equation (5) lets us choose a point in the inactive fibre that commutes with neither. If both widths are two and the kernels agree, the two operators have distinct images of order two. The equation has at most one solution in the pair of class labels, so we can use both classes in one fibre and one class in the other.

Step 2: restricting the scalar dimension. Choose a direction x∈Vx \in V acting nontrivially. The d−1d - 1 points of a symplectic simplex in x⊥/⟨x⟩x^\perp/\langle x\rangle each lift to two directions yy, y+xy + x. Their actions differ, since αy+x=αyαx\alpha_{y+x} = \alpha_y\alpha_x and αx≠1\alpha_x \ne1. Each pair supplies at least three points by Step 1, and the scalar form separates different pairs. Thus 3(d−1)≤N≤2d+43(d - 1) \le N \le2d + 4, which leaves d=4,6d = 4,6 and N≥9N \ge9.

Step 3: linearizing the action. We next show that every fibre has width at most two. The first step is to make the action a vector space of maps. If the directions of width at most two lay in a hyperplane, the simplex in Lemma 4.3 would give at least 4d+14d + 1 vertices. This contradicts N≤2d+4N \le2d + 4, so those directions span V=P/RV = P/R. Their action operators have images of order at most two. Every other action preserves each such image because the actions commute. An image of order two has no nonidentity automorphism, so it is fixed pointwise. Thus DuDv=0D_uD_v = 0 for the spanning directions, and Du+v=Du+Dv+DuDvD_{u+v} = D_u + D_v + D_uD_v makes all operators sums of the spanning ones. Their images generate the central elementary abelian group [R,P][R,P], and R2≤Z(P)R^2 \le Z(P). The common kernel in RR of all the operators is CR(P)=Z(P)C_R(P) = Z(P), since the centre lies in every scalar radical. Thus R/Z(P)R/Z(P) and [R,P][R,P] are binary vector spaces. The assignment v↦Dvv \mapsto D_v is now linear, and its image D≤Hom⁡F2(R/Z(P),[R,P])\mathcal{D} \le\operatorname{Hom}_{\mathbb{F}_2}(R/Z(P),[R,P]) is spanned by maps of rank one. A fibre has width 2 iff DvD_v.

Step 4: excluding wider fibres except at one endpoint. If some direction xx has rank Dx≥2D_x \ge2, repeat the clique construction in x⊥/⟨x⟩x^\perp/\langle x\rangle. Each pair of lifts y,y+xy,y+x now supplies four vertices. Indeed, if both widths are at most two, the maps Dy,Dy+xD_y,D_{y+x} must have rank one and distinct kernels, since their sum has rank at least two. The two-fibre argument applies. Hence N≥4(d−1)N \ge4(d-1). In the range N≤2d+4N \le2d + 4, this leaves only d=4d = 4, N=12N = 12.

Step 5: excluding the endpoint d=4d = 4, N=12N = 12. At this endpoint, we need two rank-one maps A1,A2A_1,A_2 with different images and different kernels. If no pair differed in both respects, the rank-one maps would share an image line or a kernel. Indeed, two with distinct images would have the same kernel, and every other rank-one map would have to share it. Their sums would then all have rank at most one, contrary to the existence of a map of rank at least two.

In the four-space VV, take the plane spanned by two directions inducing A1,A2A_1,A_2. Each translate of the plane contains a direction of width at least four. To see this, choose inputs u,vu,v with A1(u)=a≠0A_1(u) = a \ne0, A2(v)=b≠0A_2(v) = b \ne0 and A2(u)=A1(v)=0A_2(u) = A_1(v) = 0. The outputs a,ba,b are independent. In each translate, the two possible values at uu differ by aa, and the two at vv differ by bb. The choices are independent. If every pair were dependent, any nonzero value at vv would span a line containing both values at uu, and hence containing aa. Both values at vv would then lie on ⟨a⟩\langle a\rangle, contrary to their difference bb.

The four translates give at least four wide directions. A totally isotropic subspace in the symplectic four-space has at most three nonzero vectors. Two wide directions x,yx,y therefore pair nontrivially. Put U=⟨x,y⟩⊥U=\langle x,y\rangle^\perp. We complete x,yx,y to a simplex with the three directions x+y+ux+y+u for 0≠u∈U0\ne u\in U. They pair nontrivially with x,yx,y and with each other.

At least two of these last three directions act nontrivially. If two acted trivially, the kernel of the action map V→DV\to\mathcal{D} would have dimension at least two. Rank-nullity then gives dim⁡D≤2\dim\mathcal{D}\le2; equality holds since D\mathcal{D} contains A1,A2A_1,A_2. Its only rank-two map would be A1+A2A_1+A_2, so Dx=DyD_x=D_y. The actions at x+y+ux+y+u would then be DuD_u. Two trivial actions would force all of UU to act trivially. But then VV would act through just the one map DxD_x, contradicting dim⁡D=2\dim\mathcal{D}=2. The simplex therefore gives at least 4+4+2+2+1=134+4+2+2+1=13 clique vertices.

This excludes rank at least two throughout the range under consideration. Every fibre has width at most two.

Covering each fibre separately now costs at most 2d+1−12^{d+1}-1. For N≥2d+3N\ge2d+3, this is below B(N)≥2d+1B(N)\ge2^{d+1}. Only d=4d=4 and N=9,10N=9,10 remain. Here we replace the separate covers by subspaces isotropic for every commutator form.

Lemma 5.2 (Nine common isotropic subspaces). Let RR be an abelian noncentral scalar radical in a finite 22-group PP, with [P:R]=16[P:R]=16. Suppose every nonzero scalar form has rank at least four, every fibre has width at most two, and N=ω(P)≤10N=\omega(P)\le10. Then P/Z(P)≅F25P/Z(P)\cong\mathbb{F}_2^5 has a cover by nine common isotropic subspaces. Consequently a(P)=N=9<T(N)a(P)=N=9<T(N).

Proof. Step 1: the five-dimensional quotient. We first prove that [R:Z(P)]=2[R:Z(P)]=2 and that PP has class two with central squares. This reduces the problem to an alternating map on a five-dimensional binary space. We then construct the cover from a Kronecker block when there are two scalar forms, and from the incidence of common isotropic planes when there are three.

Put V=P/RV=P/R and retain the action notation Dg=αg−1D_g=\alpha_g-1 from the preceding proof. These operators form a space of maps from R/Z(P)R/Z(P) to [R,P][R,P], each of rank at most one. First we show [R:Z(P)]=2[R:Z(P)]=2; a larger quotient would give an eleven-point clique. If dim⁡(R/Z(P))≥2\dim(R/Z(P))\ge2, the kernels of the rank-one maps cannot all coincide, since they intersect in zero. The maps must therefore share an image line. Here we use the following rank-one observation: if a⊗fa\otimes f and b⊗gb\otimes g have different image lines and different kernels, their sum has rank two. Thus any two rank-one maps in D\mathcal{D} share an image or a kernel. Once two have different kernels, they share an image, and comparison with both shows that every other nonzero map has that same image. Their action functionals span the dual of R/Z(P)R/Z(P). Project these functionals onto two coordinates, giving a surjection λ:V→F22\lambda:V\to\mathbb{F}_2^2 with kernel JJ.

Choose x∉J∪J⊥x\notin J\cup J^\perp. Then λ(x)≠0\lambda(x)\ne0 and x⊥∩Jx^\perp\cap J has dimension one, so λ\lambda maps x⊥x^\perp onto F22\mathbb{F}_2^2. Write 00,10,01,1100,10,01,11 for the vectors of F22\mathbb{F}_2^2, and choose coordinates with λ(x)=10\lambda(x)=10. The second coordinate is then a nonzero functional on the symplectic plane x⊥/⟨x⟩x^\perp/\langle x\rangle. It is one at two of the three nonzero points and zero at the third. In each pair of lifts, adding xx toggles the first coordinate. The two pairs over the points with second coordinate one thus carry the labels 01,1101,11, and the remaining pair carries 00,1000,10.

In the first two pairs, the action functionals are nonzero and distinct because their projected labels are. Over F2\mathbb{F}_2 they therefore have distinct kernels. Each pair gives four clique points by the two-fibre argument in Lemma 5.1. In the last pair the actions differ, so the same lemma gives at least three points. The three nonzero points of a symplectic plane pair nontrivially, so the 4+4+34+4+3 points form a clique, contrary to N≤10N\le10.

Thus [R:Z(P)]=2[R:Z(P)]=2. We show that the central quotient is an elementary abelian space of dimension five, so we can look for a cover by common isotropic subspaces. The action of PP on the group R/Z(P)R/Z(P) of order two is trivial. If P′≰Z(P)P'\not\le Z(P), projection of the commutator onto R/Z(P)≅F2R/Z(P) \cong\mathbb{F}_2 gives a nonzero scalar form on P/RP/R. It is well defined because [R,P]≤Z(P)[R,P] \le Z(P). Moreover γ3(P)≤[R,P]≤Z(P)\gamma_3(P) \le[R,P] \le Z(P), so the projection kills γ3(P)\gamma_3(P): this is one of the scalar forms to which the rank hypothesis applies. On the four-space P/RP/R, it is therefore nondegenerate. Choose tt acting nontrivially on RR. Its perpendicular for this form contains two vectors with nonzero pairing, by rank at least four. For lifts x,yx,y of these vectors, conjugation by tt changes x,yx,y by central factors, so it fixes [x,y][x,y]. Yet [x,y][x,y] represents the nonzero class in R/Z(P)R/Z(P), whose representatives are moved by tt. This proves that PP has class two.

Its squares lie in RR. Since [x2,y]=[x,y2][x^2,y]=[x,y^2], the commutator [x2,y][x^2,y] is trivial if either square is central. If both are noncentral, the squares agree modulo Z(P)Z(P), so x2x^2 again commutes with yy. Thus every square is central, and [x,y]2=[x2,y]=1[x,y]^2=[x^2,y]=1 shows that P′P' is elementary abelian.

Step 2: the alternating map on the quotient. Write E=P/Z(P)≅F25E=P/Z(P)\cong\mathbb{F}_2^5, W=P′W=P' and let β\beta be the commutator map. Write Cβ(x)=ker⁡β(x,⋅)C_\beta(x)=\ker\beta(x,\cdot) for the common centralizer of xx in EE. The original scalar form vanishes on the nonzero space [R,P][R,P], so dim⁡W≥2\dim W\ge2. All nonzero scalar forms have rank four.

Step 3: two scalar forms. If dim⁡W=2\dim W=2, Scharlau’s decomposition [17] gives a single singular Kronecker block of dimension five. A singular block of dimension one would give a nonzero common radical. If a singular block had dimension three, the remaining dimension two could contain neither another singular block (there is no common radical) nor a regular block nondegenerate for all three nonzero combinations: the space of alternating forms on a two-space has dimension one. Some nonzero combination would consequently have rank at most two, contrary to the rank-four hypothesis. The only possible decomposition is therefore the single five-dimensional singular block. A singular Kronecker block of a prescribed odd dimension is unique up to simultaneous equivalence of the two alternating forms [17]. It is therefore enough to construct the cover in any model of this block. To exhibit such a model, take K=F8K=\mathbb{F}_8, a two-dimensional subspace U≤KU\le K, and a surjection π:K→F22\pi:K\to\mathbb{F}_2^2. On K⊕UK\oplus U the pair

β((x,s),(y,t))=π(xt−ys)\beta((x,s),(y,t))=\pi(xt-ys)

has zero common radical. Multiplication by s≠0s\ne0 is onto KK, and xUxU cannot lie in ker⁡π\ker\pi for x≠0x\ne0. Each nonzero scalarization has rank four because the pairing between KK and UU has rank two. It therefore models the same Kronecker block. The space K⊕0K\oplus0 and the eight graphs {(as,s):s∈U}\{(as,s):s\in U\}, a∈Ka\in K, partition the nonzero vectors into common isotropic subspaces.

Step 4: at most three scalar forms. Suppose dim⁡W≥3\dim W\ge3. We look for a three-space among the radical vectors of the scalar forms. There is one nonzero radical vector per nonzero scalar form, since every form has rank four.

If there were four independent forms, the fifteen nonzero combinations would have at most fifteen radical vectors. At a nonzero vector xx outside that set, the four functionals obtained by evaluation at xx are independent: a linear relation would put xx in the radical of a nonzero combination. Their common kernel has dimension 5−4=15-4=1 and contains xx, so it is exactly ⟨x⟩\langle x\rangle. Two distinct vectors outside the set therefore cannot commute even for these four forms. There are at least 31−15=1631-15=16 such vectors, contrary to N≤10N\le10. Thus dim⁡W=3\dim W=3.

Step 5: counting the common isotropic planes. Let SS be the set of radical vectors of the seven nonzero forms. Outside S∪{0}S\cup\{0\}, evaluation has rank three, so the common centralizer is a plane. Each such vector lies on just that one common isotropic plane: an isotropic plane containing xx must lie in Cβ(x)C_\beta(x), which already has dimension two. Take one vector outside SS from each plane that meets the complement of S∪{0}S\cup\{0\}. If two representatives commuted, they would span the unique isotropic plane through either of them. They would therefore come from the same plane. The selected representatives thus form a clique.

To see how many planes this leaves inside SS, count the planes through a nonzero vector xx. If nxn_x nonzero forms have radical ⟨x⟩\langle x\rangle, the annihilator of β(x,E)\beta(x,E) has size nx+1n_x+1. Rank-nullity gives ∣ker⁡β(x,⋅)∣=4(nx+1)|\ker\beta(x,\cdot)|=4(n_x+1). The planes through xx correspond to the nonzero vectors of ker⁡β(x,⋅)/⟨x⟩\ker\beta(x,\cdot)/\langle x\rangle, so there are

4(nx+1)2−1=1+2nx.\frac{4(n_x+1)}{2}-1=1+2n_x.

Each of the seven nonzero scalar forms has exactly one nonzero radical vector, so ∑x≠0nx=7\sum_{x \ne0} n_x = 7, even if several forms have the same radical. Summing over x≠0x \ne0 gives 31+2⋅7=4531 + 2 \cdot7 = 45 incidences. Each plane has three nonzero vectors, hence there are exactly fifteen common isotropic planes. The clique bound N≤10N \le10 leaves at least five planes with all their nonzero vectors in SS.

Step 6: forcing an isotropic three-space. These five planes force the three-space. Two must meet, since SS has at most seven points. If they meet at x≠0x \ne0, their span YY has dimension three and already contains five vectors of SS. A plane not contained in YY uses at least two vectors outside YY, and there are at most two available in SS. Those two vectors determine the plane, so at most one of the five planes lies outside YY. Among the at least four planes in YY, one avoids xx, since only three can pass through xx. It is isotropic, and xx commutes with all of YY by the construction of YY. Together this plane and xx span a common isotropic three-space.

Step 7: partitioning the remaining vectors. For 0≠y∈Y0 \ne y \in Y, we have Y≤Cβ(y)Y \le C_\beta(y), so y∈Sy \in S. These seven vectors exhaust SS. Since each accounts for one of the seven scalar forms, Cβ(y)=YC_\beta(y) = Y for 0≠y∈Y0 \ne y \in Y. Every vector outside YY lies on a unique common isotropic plane, and this plane meets YY only at zero. These planes partition the 24 vectors outside YY into eight triples, giving eight planes in addition to YY.

In either case the nine subspaces lift to nine abelian subgroups. Lemma 5.1 applies to the present hypotheses and gives N≥9N \ge9. Together with N≤a(P)≤9N \le a(P) \le9, this yields a(P)=N=9<T(N)a(P) = N = 9 < T(N).

Central radicals

If every scalar radical is central, then PP has class two, P/Z(P)P/Z(P) is elementary abelian of even dimension ν\nu, and every nonzero scalarization of the commutator is nondegenerate. If the commutator has one component, the spread gives a(P)=2ν/2+1a(P) = 2^{\nu/2} + 1, with N=ν+1N = \nu+ 1. For ν=2,4,6\nu= 2, 4, 6 these cover sizes are 3,5,93, 5, 9 and fit the stated profile. For ν≥8\nu\ge8, this group has the binary symplectic model and a(P)=T(N)a(P) = T(N).

With several forms, we again use one to separate fibres and a second to enlarge the clique inside them. Here we can triple each contribution by choosing a plane on which the forms behave differently. Take independent forms b,cb, c. There is a plane JJ isotropic for bb and nondegenerate for cc. To find it, write c(x,y)=b(Sx,y)c(x,y) = b(Sx,y), choose xx with x,Sxx, Sx independent, and choose y∈x⊥by \in x^{\perp_b} with c(x,y)≠0c(x,y) \ne0. Then J=⟨x,y⟩J = \langle x,y\rangle has the desired properties. In J⊥b/JJ^{\perp_b}/J, choose a binary symplectic clique of size ν−3\nu- 3, using the singleton {0}\{0\} if ν=4\nu= 4. Lift each vertex to a representative orthogonal to JJ for cc, which is possible because c∣Jc|_J is nondegenerate. Adding the three nonzero vectors of JJ to this representative gives three pairwise cc-nonorthogonal lifts. Lifts of different vertices are already separated by bb. Thus

N≥3(ν−3).(6)N \ge3(\nu- 3). \tag*{(6)}

When ν≥10\nu\ge10, this implies N≥2ν+1N \ge2\nu+ 1, so cyclic lifts suffice. In the smaller dimensions, we use subspaces isotropic for the full commutator. In dimensions four and six, spreads or common centralizers supply them. In dimension eight, we first cover disjoint common isotropic planes, then pair commuting points among the remaining vectors. The planes reduce the number of subgroups enough to obtain a strict bound.

Lemma 5.3 (Regular pencils in dimensions four and six). Let b,cb,c be independent alternating forms on a 2j2j-dimensional space VV over Fp\mathbb{F}_p, with j=2j=2 or 33. If every nonzero form in their span is nondegenerate, then VV can be identified with Fpj2\mathbb{F}_{p^j}^2 so that its Fpj\mathbb{F}_{p^j}-lines are isotropic for both forms.

Proof. In Scharlau’s regular decomposition [17], a block with elementary divisor fkf^k has dimension 2kdeg⁡f2k\deg f. Here there are no singular blocks or linear elementary divisors, since every nonzero combination is nondegenerate. In particular each irreducible ff has degree at least two. The sum of the quantities kdeg⁡fk\deg f over the regular blocks is jj. Since j=2j = 2 or 33, two blocks would already contribute at least four, and k≥2k \ge2 in one block would also contribute at least four. Thus there is exactly one block, with k=1k = 1 and deg⁡f=j\deg f = j. On this block the two forms are scalarizations of the determinant on Fpj\mathbb{F}_{p^j}. The Fpj\mathbb{F}_{p^j}-lines are therefore isotropic for both.

Lemma 5.4 (Small binary central quotients). Let PP have class two and elementary abelian central quotient of dimension ν∈{4,6,8}\nu\in\{4,6,8\}. Suppose that there are at least two independent scalar commutator forms, and that every nonzero form is nondegenerate. Then a(P)≤T(N)a(P) \le T(N), where N=ω(P)N = \omega(P). For N≥9N \ge9 the inequality is strict.

Proof. Step 1: the common-centralizer dimension. Write V=P/Z(P)V = P/Z(P), let β\beta be the commutator map on VV, and put N=ω(P)N = \omega(P) and e=dim⁡F2P′e = \dim_{\mathbb{F}_2} P'. The Pfaffian bound [10 Lemma 3] gives e≤ν/2e \le\nu/2.

For every z≠0z \ne0, the evaluation map β(z,⋅):V→P′\beta(z,\mathord\cdot): V \to P' is onto: otherwise a nonzero functional annihilating its image would give a scalar form with zz in its radical. Hence every common centralizer has dimension ν−e\nu- e.

Step 2: dimension four. If ν=4\nu= 4, then e=2e = 2, and the regular pencil gives five blocks by Lemma 5.3. The scalar clique has size five, so a(P)=N=5a(P) = N = 5.

Step 3: dimension six, a common spread or a third form. If ν=6\nu= 6, then e=2e = 2 or 33 and (6) gives N≥9N \ge9. With two forms, the regular pencil gives nine blocks. With three forms and N≥11N \ge11, first apply the regular pencil lemma to two forms, giving nine common isotropic three-spaces. On each three-space the restriction of the third alternating form has rank zero or two. In rank zero keep the whole space. In rank two its radical is a line, and the inverse images of the three lines of the two-dimensional quotient are isotropic planes covering the three-space. Thus each of the nine blocks costs at most three. This costs at most 27<B(N)27 < B(N).

Step 4: dimension six, the range N≤10N \le10. With three forms, put Yz=ker⁡β(z,⋅)Y_z = \ker\beta(z,\mathord\cdot). Each YzY_z has dimension three by the evaluation calculation above. We show that all these spaces are common isotropic, unless there is an eleven-point clique. If all YzY_z are common isotropic, and 0≠t∈Yz0 \ne t \in Y_z, then Yz⊆YtY_z \subseteq Y_t. Both have dimension three, so Yz=YtY_z = Y_t. Thus two such spaces meeting off zero coincide. They partition the 63 nonzero vectors into blocks of seven, giving nine blocks.

Otherwise choose zz with Y=YzY = Y_z nonisotropic and a scalar form cc nonzero on YY. In the symplectic space z⊥c/⟨z⟩z^{\perp_c}/\langle z\rangle, the plane J=Y/⟨z⟩J = Y/\langle z\rangle is nondegenerate. Translate the three nonzero vectors of JJ by a nonzero u∈J⊥u \in J^\perp and add the two nonzero vectors of J⊥J^\perp other than uu. The five vectors lie outside JJ. Within each set the pairings are unchanged, and across the sets they are the nonzero pairings with uu. Lift each vector to its two translates by zz. The form cc separates lifts of different vectors. If vv is one of the chosen lifts, its image lies outside JJ, so v∉Yv \notin Y and β(v,z)≠0\beta(v,z) \ne0. Hence vv and v+zv + z fail to commute with one another, and both fail to commute with zz. The five doubled directions together with zz therefore give eleven vertices, a contradiction.

Step 5: dimension eight. Here N≥15N \ge15 by (6). If N≥17N \ge17, cyclic lifts suffice since z(P)=256≤B(N)z(P) = 256 \le B(N). For N=15,16N = 15,16, pairing commuting points could leave up to NN singletons. We first choose NN disjoint common isotropic planes. Each covers three points in one subgroup, offsetting the cost of one singleton.

Every nonzero vector lies on the same positive number DD of common isotropic planes. Indeed, β(x,⋅)\beta(x,\mathord\cdot) is onto P′P' for x≠0x \ne0, so those planes are the lines in ker⁡β(x,⋅)/⟨x⟩\ker\beta(x,\mathord\cdot)/\langle x\rangle, of dimension 7−e≥37-e \ge3. There are 255D/3=85D255D/3 = 85D planes in total, and each chosen plane meets at most 3D3D planes off zero, including itself. After k<N≤16k < N \le16 choices, at most 3kD≤45D<85D3kD \le45D < 85D planes are forbidden. A greedy choice therefore gives NN planes with pairwise disjoint sets of nonzero vectors. Cover their 3N3N points in NN subgroups, then pair commuting points among the rest until no pair remains. The unpaired points form a clique and hence number at most NN. The total cost is at most

N+⌊255−3N+N2⌋=127<B(N).N + \left\lfloor\frac{255 - 3N + N}{2} \right\rfloor= 127 < B(N).

Products of Sylow subgroups

An abelian group needs one subgroup. If only one Sylow subgroup is nonabelian, the abelian factors preserve aa, ω\omega and the isoclinism class. Otherwise split GG into two nonabelian direct factors. Their clique numbers i,j≥3i,j \ge3 satisfy ij≤Nij \le N, and multiplying the two covers gives an upper bound for the product. We use T(k)≤2(k+1)/2T(k) \le2^{(k+1)/2} for k≥3k \ge3, direct from the formula for TT (for 3≤k≤63 \le k \le6, use k2≤2k+1k^2 \le2^{k+1}). Unless i=j=3i=j=3, we have (i−1)(j−1)≥6(i-1)(j-1) \ge6, so

T(i)T(j)≤2(i+j+2)/2≤2(ij−3)/2≤B(ij)/2.T(i)T(j) \le2^{(i+j+2)/2} \le2^{(ij-3)/2} \le B(ij)/\sqrt{2}.

For i=j=3i=j=3 the cost is nine. Induction on the number of nonabelian factors gives a(G)≤B(N)<T(N)a(G) \le B(N) < T(N), since BB is nondecreasing.

Abelian covers of nonnilpotent groups

Proposition 6.1 (Nonnilpotent groups). If GG is nonnilpotent and N=ω(G)N=\omega(G), then a(G)≤T(N)a(G) \le T(N), with strict inequality for N≥9N \ge9.

Case map. We argue by strong induction on NN. The nilpotent case is already available, as is the profile for smaller clique numbers.

Group or quotientClique rangeCovering argument
Any nonnilpotent groupN≥45N \ge45Centralizer descent and induction on the smaller clique number.
Nonsolvable groupN≤44N \le44A central extension of A5A_5 of index at most two.
Affine, cyclic complement; HH abelians=1s=1; or s≥2s \ge2Cover the nilpotent translation preimage; otherwise the quadratic bound.
Affine, cyclic complement; HH nonabelianN≥(m+1)rN \ge(m+1)r, r≥3r \ge3Lemma 6.3 and the weighted fibre estimate.
Affine, nonabelian complement21≤N≤4421 \le N \le44Affine degree nine and a cover from the widths of translation conjugates.

Table 1.

Here mm is the affine degree, ss a complement-fibre width, and rr a selected fibre width, as defined below.

Centralizer descent for large clique numbers

The selection argument is the one in [13 Proposition 4.3]; we give the proof because it uses no central-index bound. We use the classical Baer–Suzuki theorem [9 Theorem 3.8.2]: if XX is a pp-subgroup and ⟨X,Xg⟩\langle X,X^g\rangle is a pp-group for every g∈Gg \in G, then X≤Op(G)X \le O_p(G). Its nilpotent form says that if every two conjugates of a cyclic subgroup generate a nilpotent group, its normal closure is nilpotent. Indeed, apply the theorem to the pp-part of a generator for each prime pp; each such part lies in Op(G)O_p(G), so the generator lies in F(G)F(G). See also [11 pp. 281–282] for this formulation; the stronger extensions proved there using the classification of finite simple groups are not needed.

We first record two elementary facts used in the descent. A nonnilpotent group has clique number at least four. Indeed, if ω(U)≤3\omega(U) \le3, (3) with the whole group as its coset gives z(U)≤6z(U) \le6. A central quotient cannot be nontrivial cyclic, so its only nontrivial possibilities are C22C_2^2 and S3S_3. The latter has a four-point clique (the three transpositions and a 3-cycle), which lifts to UU. The former makes UU nilpotent of class at most two.

Also, if KK and LL centralize one another, then

ω(KL)≥ω(K)ω(L/Z(L)).(7)\omega(KL) \ge\omega(K)\omega(L/Z(L)). \tag*{(7)}

To see this, multiply every vertex of a clique in KK by every chosen lift of a clique in L/Z(L)L/Z(L). For distinct quotient vertices, a commutator from LL is noncentral in LL, whereas any cancelling commutator from KK would lie in K∩L≤Z(L)K \cap L \le Z(L). For equal quotient vertices, noncommutation is inherited from KK. The products are distinct for the same reason.

Lemma 6.2 (Centralizer descent). Let GG be a finite nonnilpotent group and put N=ω(G)N = \omega(G). There is a nonnilpotent subgroup L≤GL \leq G whose centralizer K=CG(L)K = C_G(L) satisfies

[G:K]≤N(N−1),4ω(K)≤N.[G : K] \leq N(N-1), \qquad4\omega(K) \leq N.

Proof. Put F=F(G)F = F(G), S=G∖FS = G \setminus F, and f=[G:F]f = [G : F]. Take a maximum clique DD in SS, of size t≤Nt \leq N. Its centralizers cover SS. For C=CG(x)C = C_G(x) with x∈Dx \in D, we have [C:C∩F]≤f[C : C \cap F] \leq f, so ∣C∩S∣≤(1−1/f)∣C∣|C \cap S| \leq(1 - 1/f)|C|. Since ∣S∣=(1−1/f)∣G∣|S| = (1 - 1/f)|G|, summing gives

∑x∈D1[G:CG(x)]≥1.\sum_{x \in D} \frac{1}{[G : C_G(x)]} \geq1.

Some x∈Dx \in D therefore has r=[G:CG(x)]≤t≤Nr = [G : C_G(x)] \leq t \leq N. By Baer–Suzuki, xx has a conjugate yy such that L=⟨x,y⟩L = \langle x, y \rangle is nonnilpotent. The CG(x)C_G(x)-orbit of yy avoids xx, so it has size at most r−1r - 1. Its stabilizer is K=CG(x,y)K = C_G(x, y), giving [G:K]≤r(r−1)[G : K] \leq r(r - 1). The quotient L/Z(L)L/Z(L) is nonnilpotent, since a central extension of a nilpotent group is nilpotent. It has a clique of size four. Now (7) gives 4ω(K)≤N4\omega(K) \leq N.

The numerical cost of descent. Let N≥45N \geq45 and put j=⌊N/4⌋j = \lfloor N/4 \rfloor. For the centralizer in Lemma 6.2, induction gives a(K)≤T(j)a(K) \leq T(j), and (2) gives

a(G)≤T(j){1+(N(N−1)−1)(N−1)}<N3T(j).a(G) \leq T(j)\{1 + (N(N - 1) - 1)(N - 1)\} < N^3T(j).

For 45≤N≤4845 \leq N \leq48 this bound is at most 483⋅33<222=B(45)48^3 \cdot33 < 2^{22} = B(45). Increasing NN by four quadruples B(N)B(N), while T(j)T(j) grows by at most a factor two and N3N^3 grows by a factor at most (49/45)3<2(49/45)^3 < 2. Thus a(G)<B(N)<T(N)a(G) < B(N) < T(N) for N≥45N \geq45.

Nonsolvable groups with NN at most 44

Let GG be nonsolvable with N≤44N \leq44, and let SS be its solvable radical. Write G′′=(G′)′G'' = (G')'. In alternatives (1) and (2) of [2 Theorem 1.1], G=Z(G)KG = Z(G)K with K≅PSL⁡2(p)K \cong\operatorname{PSL}_2(p) or SL⁡2(p)\operatorname{SL}_2(p), where p∈{5,7}p \in\{5, 7\}. Here G′′=KG'' = K, since KK is perfect, and S=Z(G)S = Z(G), since G/Z(G)G/Z(G) is nonabelian simple. Taking a=1a = 1 in these two cases, all four alternatives give G=G′′⟨a⟩SG = G''\langle a\rangle S with a2∈Z(G)a^2 \in Z(G) and G′′G'' isomorphic to A5A_5, SL⁡2(5)\operatorname{SL}_2(5), PSL⁡2(7)\operatorname{PSL}_2(7) or SL⁡2(7)\operatorname{SL}_2(7). The last two are excluded by ω(G′′/Z(G′′))≤N\omega(G''/Z(G'')) \leq N and ω(PSL⁡2(7))=57\omega(\operatorname{PSL}_2(7)) = 57 [1 Lemma 4.4]. Thus G′′/Z(G′′)≅A5G''/Z(G'') \cong A_5.

The nonidentity elements of A5A_5 have as centralizers its five Klein four subgroups, ten subgroups of order three and six subgroups of order five. Two commuting nonidentity elements have the same centralizer, since these centralizers are abelian. Choosing one element from each gives ω(A5)≥21\omega(A_5) \geq21.

The intersection G′′∩SG'' \cap S is solvable and normal in G′′G'', so its image in A5A_5 is trivial. Conversely Z(G′′)Z(G'') is abelian and normal in GG, hence G′′∩S=Z(G′′)G'' \cap S = Z(G''). In particular [G′′,S]≤Z(G′′)[G'', S] \leq Z(G''). For each s∈Ss \in S, the map x↦x−1xsx \mapsto x^{-1}x^s from G′′G'' to its centre is a homomorphism. It is trivial by perfectness, so [G′′,S]=1[G'', S] = 1. By (7), 21ω(S)≤ω(G′′S)≤4421\omega(S) \leq\omega(G''S) \leq44, forcing SS to be abelian.

Put M=G′′SM = G''S. Then S≤Z(M)S \leq Z(M) and M/S≅A5M/S \cong A_5. The 15+10+6=3115 + 10 + 6 = 31 nontrivial cyclic subgroups of A5A_5 have abelian preimages in MM and cover it. Since a2∈Z(G)≤Sa^2 \in Z(G) \leq S, we have [G:M]≤2[G : M] \leq2. By (2),

a(G)≤31N<B(N)(N≥21).a(G) \leq31N < B(N) \qquad(N \geq21).

Indeed 31⋅22<210=B(21)=B(22)31 \cdot22 < 2^{10} = B(21) = B(22), and in steps of two the left side grows by less than a factor two while BB doubles.

The affine quotient

Let GG be solvable and nonnilpotent with N≤44N \leq44. Choose a nonnilpotent quotient Q=G/HQ = G/H of smallest order. Every proper quotient of QQ is nilpotent. Take a nonnormal maximal subgroup $L of QQ. It has trivial core, since a nontrivial core would give a smaller quotient with a nonnormal maximal subgroup, hence a smaller nonnilpotent quotient. The primitive solvable structure theorem [6, Theorem A.15.6} gives Q=V⋊LQ=V\rtimes L, with ∣V∣=m=pd|V|=m=p^d and a faithful irreducible action of LL on VV. We view (v,a)∈Q(v,a)\in Q as the affine map x↦v+a(x)x\mapsto v+a(x). Thus VV consists of translations and LL of maps fixing the origin. The inverse images in GG of affine maps are the fibres whose widths we will add.

Since LL acts faithfully and QQ is nonnilpotent, we have m≥3m\ge3. The complement L≅Q/VL\cong Q/V is nilpotent by minimality, and its largest normal pp-subgroup Op(L)O_p(L) is trivial. Thus p∤∣L∣p\nmid|L|. A clique of QQ lifts to a clique of GG. Any nonidentity central element a∈La\in L acts without nonzero fixed points, since its fixed space is LL-invariant and the action is faithful. The mm elements with linear part aa form a clique, to which any nonzero translation can be adjoined. Hence m+1≤Nm+1\le N.

When the complement is abelian. The algebra generated by LL is a field by irreducibility, so LL is cyclic of order e>1e>1 dividing m−1m-1. If the linear part aa is nonidentity, then 1−a1-a is invertible, so the affine map has a unique fixed point. It therefore belongs to exactly one of the mm point stabilizers, which are the conjugates of LL. An affine map commuting with it must fix that point. Thus nonidentity maps from different stabilizers do not commute, and a nonzero translation commutes with none of them. This lets us add clique sizes across the corresponding fibres.

Let KK be the preimage of the translations. Write uu for the largest width of a nonzero translation fibre and ss for the largest width of a nonidentity complement fibre. By conjugacy a fibre of width ss occurs over each point stabilizer. Choose one there and a translation fibre of width uu. The cliques in these fibres join by Lemma 2.2(ii). Writing v=msv=ms for the contribution from the complements, we get u+v≤Nu+v\le N.

By Lemma 2.2(i), select one translation fibre per line and one generator per nontrivial cyclic subgroup of each complement. There are c=(m−1)/(p−1)c=(m-1)/(p-1) translation fibres and ℓ\ell fibres per complement. In particular ℓ≤⌊e/2⌋\ell\le\lfloor e/2\rfloor, the number of inversion orbits on nonidentity elements. For an abelian subgroup of index II in HH, the quotient estimate gives

a(G)≤a(H)+I(cu+ℓv).(8)a(G)\le a(H)+I(cu+\ell v). \tag*{(8)}

If mm is odd, then p≥3p\ge3 and c,ℓ≤(m−1)/2c,\ell\le(m-1)/2. If mm is even, then c≤m−1c\le m-1 and ℓ≤(m−2)/2\ell\le(m-2)/2.

Abelian kernel and cyclic complement

Write κ=⌊(N−1)/2⌋\kappa=\lfloor(N-1)/2\rfloor, so N≤2κ+2N\le2\kappa+2. For κ≥4\kappa\ge4 we have κ2≤2κ=B(N)\kappa^2\le2^\kappa=B(N), since the square starts at 42=244^2=2^4 and grows by less than a factor two in each step.

Suppose first that s=1s=1. For every nonidentity complement element, every lift centralizes HH by the abelian coset formula (1). The conjugates of the complement generate QQ. Indeed, for a nonidentity complement element aa, the map 1−a1-a is invertible on the translations, and its values arise as commutators with aa. Because HH is itself abelian, it follows that H≤Z(G)H\le Z(G). Thus KK is nilpotent, as it is a central extension of an abelian group, and the preimage of each cyclic complement is abelian.

A maximum clique in KK can be chosen in K∖HK\setminus H. If KK is nonabelian, its maximum cliques avoid the centre. If KK is abelian, choose one lift of a nonzero translation. It follows that m+ω(K)≤Nm+\omega(K)\le N. The cover of KK together with the mm abelian complement preimages costs m+T(k)m+T(k), where k=ω(K)k=\omega(K). If N≤8N\le8, then k≤5k\le5 and the cost is at most N≤T(N)N\le T(N). For N≥9N\ge9 and k≤8k\le8, use T(k)≤k+3T(k)\le k+3 to bound the cost by N+3≤κ2N+3\le\kappa^2. If k≥9k\ge9, the m≥3m\ge3 complements leave k≤N−3k\le N-3, so B(k)≤B(N)/2B(k)\le B(N)/2. The cost is at most N−8+B(N)/2≤B(N)N-8+B(N)/2\le B(N), since N−8≤2κ−6≤2κ−1N-8\le2\kappa-6\le2^{\kappa-1}. Thus the bound is strict for N≥9N\ge9.

Now suppose s≥2s\ge2. The complement fibres contribute at least 2m2m clique vertices and the translation fibres contribute at least one, so κ≥m\kappa\ge m. In the fibre cover (8), with a(H)=I=1a(H)=I=1, those contributions give a cost 1+c+2mℓ1+c+2m\ell. Each further vertex costs at most D=max⁡(c,ℓ)D=\max(c,\ell) subgroups. Since u+v≤N≤2κ+2u+v\le N\le2\kappa+2, there are at most 1+2(κ−m)1+2(\kappa-m) further vertices. Hence

a(G)≤1+c+2mℓ+D+2(m−1)(κ−m).a(G)\le1+c+2m\ell+D+2(m-1)(\kappa-m).

The bounds on c,ℓc,\ell give 1+c+2mℓ+D≤m21+c+2m\ell+D\le m^2: for odd mm use c,D,ℓ≤(m−1)/2c,D,\ell\le(m-1)/2, and for even mm use c,D≤m−1c,D\le m-1 and 2ℓ≤m−22\ell\le m-2. Consequently

a(G)≤m2+2(m−1)(κ−m)≤κ2.a(G)\le m^2+2(m-1)(\kappa-m)\le\kappa^2.

This proves a(G)≤B(N)<T(N)a(G)\le B(N)<T(N) when κ≥4\kappa\ge4. The only smaller value is κ=3\kappa=3, where N=7,8N=7,8 and the same bound gives a(G)≤9<T(N)a(G)\le9<T(N).

Nonabelian kernel and cyclic complement

A narrow fibre supplies a large abelian subgroup of the kernel and controls the cost of its cosets.

Lemma 6.3 (Covering the kernel from one fibre). Let H◃GH\triangleleft G be nonabelian and let gHgH have width rr. Assume a(C)≤T(r)a(C)\le T(r) for each C≤HC\le H with ω(C)≤r\omega(C)\le r. There is an abelian subgroup A≤HA\le H of index II, centralized by an element of gHgH, with

a(H)≤rI≤316B(4r),I≤2r.a(H)\le rI\le\frac{3}{16}B(4r),\qquad I\le2^r.

Proof. As in § 2, if x∈gHx\in gH centralizes AA then the II cosets of AA in HH cost at most rIrI in all. We find AA inside a centralizer in the cover of a maximum clique in gHgH. The centralizers in HH have intersection D≤Z(H)D\le Z(H) and satisfy [H:D]≤r![H:D]\le r! by (3).

The uniform construction. For r≥4r\ge4, if all centralizers have index at least rr, their cosets partition the fibre into commuting sets by Lemma 2.3. Each centralizer is then abelian of index rr. Otherwise choose C=CH(x)C=C_H(x) of index i≤r−1i\le r-1. It satisfies ω(C)≤r\omega(C)\le r by Lemma 2.4. If CC is nonabelian, a cover by at most T(r)T(r) abelian subgroups contains one of index at most T(r)−1T(r)-1: if the largest has order h0h_0, their union has size at most 1+T(r)(h0−1)<T(r)h01+T(r)(h_0-1)<T(r)h_0. If CC is abelian, take CC itself. For r≥5r\ge5, this gives I≤(r−1)(T(r)−1)I\le(r-1)(T(r)-1).

Small widths. Only widths three and four need sharper information about the central quotient. For r=3r=3, the noncyclic quotient H/Z(H)H/Z(H) has order four or six. In the latter case it is S3S_3, whose automorphisms are inner. As HH is normal, one can choose a representative of gHgH that centralizes H/Z(H)H/Z(H) in G/Z(H)G/Z(H). Translating the three transpositions and a 3-cycle by this representative gives a four-point clique in the image of gHgH, a contradiction. Thus [H:Z(H)]=4[H:Z(H)]=4 and D=Z(H)D=Z(H). Some centralizer has index at most three by Lemma 2.4. Its index divides four and is at least two, hence equals two. Its quotient by Z(H)Z(H) is cyclic, so this centralizer is abelian.

For r=4r=4, the construction above gives I≤6I\le6 if i≤2i\le2, and I=4I=4 in the partition case. If i=3i=3, we need an abelian subgroup of index at most two in CC. The same centralizer intersection gives z(C)≤[C:D]≤4!/3=8z(C)\le[C:D]\le4!/3=8. If C/Z(C)C/Z(C) has an element of order at least three, its cyclic subgroup has index at most two, and its preimage is abelian. Otherwise the quotient is elementary abelian of order at most eight. A line suffices when its order is two or four, and CC itself suffices if it is abelian. At order eight, two distinct nonzero classes must have commuting lifts, since ω(C)≤4\omega(C)\le4. They span a plane with abelian preimage of index two. In all cases I≤6I\le6.

The resulting bounds. For r=3,4r=3,4, the index bounds give rI≤(3/16)B(4r)rI\le(3/16)B(4r). For r≥5r\ge5, the profile gives T(r)−1≤2r−3T(r)-1\le2^{r-3}, so

rIB(4r)≤r(r−1)2−r−2≤532.\frac{rI}{B(4r)}\le r(r-1)2^{-r-2}\le\frac{5}{32}.

The last expression decreases from r=5r=5 since (r+1)/(2(r−1))<1(r+1)/(2(r-1))<1. The same index bounds give I≤2rI\le2^r: for r≥5r\ge5, both r−1r-1 and T(r)−1T(r)-1 are at most 2r/22^{r/2}. ◻

Take r=min⁡(s,u)r=\min(s,u). Both widths are at least three, and N≥v+u≥(m+1)rN \ge v+u \ge(m+1)r ensures r<Nr<N. The induction hypothesis gives the cover bound required in Lemma 6.3. With its subgroup AA, the fibre estimate (8) becomes

a(G)≤I(r+cu+ℓv).a(G) \le I(r+cu+\ell v).

The translation fibres contribute at least rr clique vertices, and the complements at least mrmr. At these widths, the cover bound is Ir(1+c+mℓ)Ir(1+c+m\ell), including the IrIr subgroups for the kernel. The bounds on c,ℓc,\ell give 1+c+mℓ≤(m2+1)/21+c+m\ell\le(m^2+1)/2 for either parity of mm. Put M=(m2+1)/2M=(m^2+1)/2.

For fixed mm, each vertex beyond (m+1)r(m+1)r increases the cover bound by at most I(m−1)I(m-1), since c,ℓ≤m−1c,\ell\le m-1. As rM≥3(m−1)rM \ge3(m-1), the linear bound

IrM+I(m−1)(N−(m+1)r)IrM+I(m-1)(N-(m+1)r)

grows by a factor at most 4/34/3 per extra vertex. The lower bound 2N/2−12^{N/2-1} for B(N)B(N) grows faster, by a factor 2\sqrt{2}. Comparing at the starting size (m+1)r(m+1)r therefore gives

a(G)B(N)≤rIB(4r)M2−(m−3)r/2.\frac{a(G)}{B(N)} \le\frac{rI}{B(4r)}M2^{-(m-3)r/2}.

Finally, increasing mm by one adds rr vertices to the starting clique size. The factor MM increases by a factor of at most two, whereas the exponential denominator increases by 2r/2>22^{r/2}>2. Thus the last bound is largest at m=3m=3, where M=5M=5, and is at most 15/1615/16 by Lemma 6.3.

Nonabelian complement

Choose 1≠x∈Z(L)1\ne x\in Z(L). Its fixed-point-free action on VV gives CQ(x)=LC_Q(x)=L. We enlarge a clique in this centralizer through the translation fibres.

Take a maximum clique DD in LL. For each c∈Dc\in D, use the two complement elements c,xcc,xc. Distinct pairs are completely noncommuting modulo VV. Coprimeness gives

V=Uc⊕Wc,Uc=ker⁡(1−c),Wc=(1−c)V.V=U_c\oplus W_c,\qquad U_c=\ker(1-c),\qquad W_c=(1-c)V.

Fixing the UcU_c component and varying the WcW_c component gives a clique of size ∣Wc∣|W_c| over cc. When Uc≠0U_c\ne0, choose its component nonzero. Commutation with an element (w,e)(w,e), where ee commutes with cc, requires

(1−e)v=(1−c)w.(1-e)v=(1-c)w.

Projection onto UcU_c shows that this clique is completely noncommuting with both fibres over xcxc and over xx: both act without nonzero fixed vectors on UcU_c.

The fixed spaces of cc and xcxc intersect trivially. If both are nonzero, the two cliques have sizes with product at least mm, so together give at least 2m2\sqrt{m} vertices. If one is zero, its full fibre gives m≥2mm\ge2\sqrt{m} vertices. Use such a pair for mm vertices if it exists. Otherwise adjoin the full fibre over xx. Each of at least ∣D∣−1|D|-1 further pairs contributes 2m2\sqrt{m}. Thus

ω(Q)≥m+2(ω(L)−1)m≥q0:=m+4m.\omega(Q)\ge m+2(\omega(L)-1)\sqrt{m}\ge q_0:=m+4\sqrt{m}.

A nonabelian LL requires d≥2d\ge2. It has a nonabelian Sylow subgroup, so its order is divisible by the cube of a prime. Since p∤∣L∣p\nmid|L|, its order divides ∏i=1d(pi−1)\prod_{i=1}^{d}(p^i-1), the part of ∣GL⁡d(p)∣|\operatorname{GL}_d(p)| coprime to pp. For m=4,8,16m=4,8,16 this product is 33, 2121 and 32⋅5⋅73^2\cdot5\cdot7, respectively. None has a prime cube divisor. The remaining possibilities are m=9m=9 and m≥25m\ge25. The latter gives N≥q0≥45N\ge q_0\ge45, already treated in § 6.1. Thus m=9m=9, N≥21N\ge21, and LL is a 22-group of order at most 1616, since ∣GL⁡2(3)∣=48|\operatorname{GL}_2(3)|=48 and 3∤∣L∣3\nmid|L|.

The kernel HH is abelian. Otherwise every fibre has width at least three by (3), and a clique of size at least 2121 in QQ would give N≥63N\ge63. Write w(q)w(q) for the width of the fibre over qq. For q=(v,c)∈Q∖Vq=(v,c)\in Q\setminus V, its conjugates by translations are the ∣Wc∣|W_c| elements (v+t,c)(v+t,c) with t∈Wct\in W_c. Distinct such elements do not commute, since Uc∩Wc=0U_c \cap W_c = 0. Their fibres have the same width, so ∣Wc∣w(q)≤N|W_c|w(q) \le N. Summing over the m=∣Uc∣∣Wc∣m = |U_c||W_c| elements with linear part cc, and then over c≠1c \ne1, gives

∑q∈Q∖Vw(q)≤N∑c∈L∖{1}∣Uc∣.\sum_{q \in Q \setminus V} w(q) \le N \sum_{c \in L \setminus\{1\}} |U_c|.

If c≠1c \ne1 and Uc≠0U_c \ne0, its eigenvalues are 1,−11,-1, since the action is coprime and the field is F3\mathbb{F}_3. Thus ∣Uc∣=3|U_c| = 3 and det⁡(c)=−1\det(c) = -1. There are at most ∣L∣/2|L|/2 such elements by the determinant homomorphism. For the other c≠1c \ne1 we have ∣Uc∣=1|U_c| = 1, so the sum is at most 2∣L∣−1≤312|L| - 1 \le31. Add one fibre for each of the four lines in VV and include HH itself. By Lemma 2.2,

a(G)≤1+4N+31N<B(N)(21≤N≤44),a(G) \le1 + 4N + 31N < B(N) \qquad(21 \le N \le44),

since 1+35⋅22<210=B(21)1 + 35 \cdot22 < 2^{10} = B(21) and 1+35⋅44<211=B(23)1 + 35 \cdot44 < 2^{11} = B(23). This completes the proof of Proposition 6.1.

Attaining the profile

Proof of Theorem 1.1. Propositions 4.1 and 6.1 give a(G)≤T(ω(G))a(G) \le T(\omega(G)). Since TT is nondecreasing, h(n)≤T(n)h(n) \le T(n). The groups in § 3.1 attain these covering bounds, so h(n)=T(n)h(n) = T(n).

Fix n≥9n \ge9 and suppose a(G)=T(n)a(G) = T(n) with N=ω(G)≤nN = \omega(G) \le n. Then T(N)=T(n)T(N) = T(n), so N≥9N \ge9 and ⌊(N−1)/2⌋=⌊(n−1)/2⌋\lfloor(N - 1)/2 \rfloor= \lfloor(n - 1)/2 \rfloor. By Propositions 4.1 and 6.1, GG has the binary symplectic model of dimension N−1N - 1. Thus NN is odd and this dimension is 2⌊(n−1)/2⌋2\lfloor(n - 1)/2 \rfloor. Conversely, the extraspecial model in this dimension attains the bound by the symplectic spread.

The central-index theorem [13 Theorems 1.1 and 1.2(v)] gives ζ(n)=(n−1)2\zeta(n) = (n - 1)^2 for 3≤n≤63 \le n \le6, ζ(7)=ζ(8)=81\zeta(7) = \zeta(8) = 81, and ζ(n)=4⌊(n−1)/2⌋\zeta(n) = 4^{\lfloor(n-1)/2 \rfloor} for n≥9n \ge9. Thus T(n)=1+ζ(n)T(n) = 1 + \sqrt{\zeta(n)} for n≥3n \ge3. For n≥9n \ge9, its equality case is the same binary symplectic model just obtained for the covering problem.

Extremal profiles and equality for individual groups. The identity h(n)=1+ζ(n)h(n) = 1 + \sqrt{\zeta(n)} compares two maxima under the same clique bound. The equality cases show what that bound selects. For q=2rq = 2^r, the determinant on Fq\mathbb{F}_q and its binary trace both give central index q2q^2 and covering number q+1q + 1. The same field lines give an optimal cover in each model. Their clique numbers, however, are q+1q + 1 and 2r+12^r + 1, respectively. Thus the clique constraint can distinguish commutator maps that give the same central index and the same covering number.

For an individual group satisfying a(G)=1+z(G)a(G) = 1 + \sqrt{z(G)}, let s(G)s(G) be the largest order of the image of an abelian subgroup in G/Z(G)G/Z(G). In the tight case, s(G)=z(G)s(G) = \sqrt{z(G)}, and every optimal cover induces a partition of the nonidentity elements of the central quotient. In the loose case, s(G)>z(G)s(G) > \sqrt{z(G)}, and every optimal cover has two members whose images share a nonidentity element. The loose case occurs even for groups of class two with elementary abelian central quotient. The open question is which incidence properties of the abelian images determine whether exactly 1+z(G)1 + \sqrt{z(G)} members are needed in this regime.

Statements and declarations

Funding. The author received no specific funding for this work.

Competing interests. The author declares no competing interests.

Data availability. No datasets were generated or analysed during this study.

Author contributions. The sole author carried out the research and wrote the manuscript.

Declaration of generative AI and AI-assisted technologies

OpenAI ChatGPT assisted with mathematical discussion, reference checks and LaTeX editing.

The author takes responsibility for the mathematical arguments, references and final text.

References

  1. [1]A. Abdollahi, S. Akbari and H. R. Maimani, Non-commuting graph of a group, J. Algebra 298 (2006), 468–492. doi:10.1016/j.jalgebra.2006.02.015.DOI
  2. [2]A. Abdollahi, A. Azad, A. Mohammadi Hassanabadi and M. Zarrin, On the clique numbers of non-commuting graphs of certain groups, Algebra Colloq. 17 (2010), 611–620. doi:10.1142/S1005386710000581. arXiv:0903.0692.DOI
  3. [3]E. A. Bertram, Some applications of graph theory to finite groups, Discrete Math. 44 (1983), 31–43. doi:10.1016/0012-365X(83)90004-3.DOI
  4. [4]M. Ceria, J. De Beule, F. Pavese and V. Smaldore, On large partial ovoids of symplectic and Hermitian polar spaces, J. Combin. Des. 31 (2023), no. 1, 5–22. doi:10.1002/jcd.21864. arXiv:2203.04553.DOI
  5. [5]A. Y. M. Chin, On non-commuting sets in an extraspecial p-group, J. Group Theory 8 (2005), 189–194. doi:10.1515/jgth.2005.8.2.189.DOI
  6. [7]P. Erdős, Some unsolved problems, in B. Bollobás and A. Thomason (eds.), Combinatorics, Geometry and Probability: A Tribute to Paul Erdős, Cambridge University Press, Cambridge, 1997, pp. 1–10. doi:10.1017/CBO9780511662034.004.DOI
  7. [8]V. Faber, R. Laver and R. McKenzie, Coverings of groups by abelian subgroups, Canad. J. Math. 30 (1978), 933–945. doi:10.4153/CJM-1978-081-1.DOI
  8. [9]D. Gorenstein, Finite Groups, Harper & Row, New York, 1968.
  9. [10]R. Gow and R. Quinlan, On the vanishing of subspaces of alternating bilinear forms, Linear Multilinear Algebra 54 (2006), 415–428. doi:10.1080/03081080500361264.DOI
  10. [11]R. M. Guralnick and G. R. Robinson, On extensions of the Baer–Suzuki theorem, Israel J. Math. 82 (1993), 281–297. doi:10.1007/BF02808114.
  11. [12]A. Klein and K. Metsch, New results on covers and partial spreads of polar spaces, Innov. Incidence Geom. 1 (2005), 19–34. doi:10.2140/iig.2005.1.19. Corrections, Innov. Incidence Geom. 11 (2010), 237–240.
  12. [13]G. Lecomte, The exact central-index profile of groups with bounded noncommutativity, preprint, 2026. arXiv:2611.00943.
  13. [14]D. R. Mason, On coverings of a finite group by abelian subgroups, Math. Proc. Cambridge Philos. Soc. 83 (1978), 205–209. doi:10.1017/S0305004100054463.
  14. [15]B. H. Neumann, A problem of Paul Erdős on groups, J. Austral. Math. Soc. Ser. A 21 (1976), 467–472. doi:10.1017/S1446788700019303.
  15. [16]L. Pyber, The number of pairwise non-commuting elements and the index of the centre in a finite group, J. London Math. Soc. (2) 35 (1987), 287–295. doi:10.1112/jlms/s2-35.2.287.
  16. [17]R. Scharlau, Paare alternierender Formen, Math. Z. 147 (1976), 13–19. doi:10.1007/BF01214270.DOI
  17. [18]G. Tallini, Fibrations through lines in a nonsingular quadric Q₄,q of PG(4, q), Atti Accad. Peloritana Pericolanti Cl. Sci. Fis. Mat. Natur. 66 (1988), 127–146 (1989).
  18. [19]G. Tallini, Blocking sets with respect to planes in PG(3, q) and maximal spreads of a nonsingular quadric in PG(4, q), Mitt. Math. Sem. Giessen 201 (1991), 141–147.
  19. [20]M. J. Tomkinson, Groups covered by finitely many cosets or subgroups, Comm. Algebra 15 (1987), 845–859. doi:10.1080/00927878708823445.
  20. [22]M. Zarrin, On non-commuting sets and centralisers in infinite groups, Bull. Aust. Math. Soc. 93 (2016), 42–46. doi:10.1017/S0004972715000696.

Paper details

vlnc:W_m534axy7
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 7

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

OpenAI ChatGPT

View statement
Looking for related papers…
Contents