betaFeedback

WorksMathematics

The Herzog–Schönheim Conjecture for Solvable Groups

Abstract

The Herzog–Schönheim conjecture states that if a group is partitioned into finitely many, and at least two, cosets of subgroups, then two of these subgroups have the same index. We prove the conjecture for all solvable groups, finite or infinite. Sun's bound on unions of cosets, applied in a quotient, is combined with divisibility restrictions on local indices relative to the subgroup generated by the pp-elements, where pp is the largest prime divisor of the order of a minimal counterexample. These restrictions are obtained by counting fixed points in a primitive solvable quotient. This suffices except at the prime 77, where the counting bound narrowly fails and an additional overlap argument is needed. The proof has been formalized in Lean~4 with mathlib.

Introduction

Erdős conjectured that in any partition of the integers into finitely many, and at least two, arithmetic progressions, the largest common difference occurs at least twice. This was proved independently by Mirsky and Newman and by Davenport and Rado; see the introduction of [9] for the history. Viewing arithmetic progressions as cosets of subgroups of Z\mathbb{Z}, Herzog and Schönheim proposed the following generalization [11].

Conjecture 1.1 (Herzog–Schönheim). Let GG be a group and let

G=⨆i=1kgiHi,Hi≤G,gi∈G,(1)G = \bigsqcup_{i=1}^{k} g_iH_i,\qquad H_i \leq G,\qquad g_i \in G, \tag*{(1)}

be a partition of GG into k≥2k \geq2 left cosets. Then [G:Hi]=[G:Hj][G : H_i] = [G : H_j] for some i≠ji \ne j.

We call the cosets giHig_iH_i the parts of the partition. Following [9], we say that GG is HSHS if every such partition of GG contains two cosets whose underlying subgroups have equal index. The conjecture is known for finite nilpotent groups [4] and pyramidal groups [5], for partitions by subnormal subgroups [18], under conditions on the prime divisors of ∣G∣|G| [10], for groups of order less than 1440 [13], and for simple and symmetric groups [9]. For arbitrary groups, Akman and Sissokho proved it for partitions into cosets of at most seven distinct subgroups [1]. They also study Steiner partitions, which use exactly one coset of each subgroup, when the subgroups commute pairwise; the case of five mutually commuting subgroups is treated in [3]; see also [2]. A preprint claiming the solvable case was withdrawn [6]; we do not use its argument. Our main result is the following.

Theorem 1.2. Every solvable group is HSHS.

The idea

For a finite group, a partition (1) satisfies ∑i1/[G:Hi]=1\sum_i 1/[G:H_i]=1. Garonzi and Margolis prove the conjecture for simple and symmetric groups by showing that the sum of 1/m1/m over the distinct subgroup indices mm of GG is less than 2 [9 Lemma 2.1]. Such a criterion depends only on the set of subgroup indices, and it cannot work for solvable groups: the sum is unbounded already on cyclic groups, as they note. Our argument also uses where the cosets lie. We compare the reciprocal sum of a family of parts with the size of the union of their images in a quotient G/FG/F, and bound that size from below by a theorem of Sun on unions of cosets.

For a finite group GG, put Π(G)=∏q∣∣G∣(1−q−1)−1\Pi(G)=\prod_{q\mid|G|}(1-q^{-1})^{-1}, with qq ranging over primes. Ginosar and Schnabel showed that Π(G)≤2\Pi(G)\le2 implies that GG is HS [10 Theorem A]. Our criterion (Proposition 5.1) takes pp to be the largest prime divisor of ∣G∣|G|. It combines the corresponding product over the primes below pp with a factor that depends on the multiplicative orders of the primes q<pq<p modulo pp and tends to zero with pp. That factor comes from the local indices, which are not visible to a criterion on $|G| alone.

How the proof works

Let GG be a finite solvable group of least order that is not HS, and let di=[G:Hi]d_i=[G:H_i]. No two did_i are coprime, because cosets of subgroups of coprime index always meet. Using the minimality of ∣G∣|G|, an elementary reduction (Lemma 2.3) also shows that no did_i equals 2 and that the largest prime pp dividing ∣G∣|G| is at least 5. The rest of the proof plays two facts against each other. The indices must be spread out, since they are distinct numbers whose reciprocals sum to 1. But the group structure forces them into a thin set.

Step 1: split each index at a normal subgroup (Sections 3 and 4). Let FF be the subgroup generated by the elements of pp-power order in GG. It is a nontrivial normal subgroup, and it has no proper normal subgroup of index prime to pp. Each index factors as

di=hiℓi,hi=[G:HiF],ℓi=[F:Hi∩F].d_i=h_i\ell_i,\qquad h_i=[G:H_iF],\qquad\ell_i=[F:H_i\cap F].

We call hih_i the quotient index and ℓi\ell_i the local index of the part. The part meets each FF-coset it touches in ∣F∣/ℓi|F|/\ell_i elements. A part with ℓi=1\ell_i=1 is a union of FF-cosets; we call it universal. If every part is universal, the partition descends to the smaller group G/FG/F with the same indices, contrary to minimality. So the set JJ of indices of the non-universal parts is nonempty.

Step 2: Sun’s bound applied in the quotient (Theorem 2.4 and Lemma 3.2). Record each positive integer nn by its exponent vector v(n)v(n), the list of exponents in its prime factorization, and give that vector the weight 1/n1/n. Let UU be the set of vectors that are coordinatewise at least v(hi)v(h_i) for some i∈Ji\in J; these are the vectors of the multiples of the quotient indices. Write W(U)W(U) for the total weight of UU. The non-universal parts together cover a union of FF-cosets, so their images cover rνr\nu of the r=[G:F]r=[G:F] cosets in G/FG/F, where ν=∑i∈J1/di\nu=\sum_{i\in J}1/d_i. Sun’s lower bound for the size of a union of cosets then gives W(U)≤ΠνW(U)\le\Pi\nu, where Π\Pi depends only on the primes involved.

Step 3: local indices are restricted (Lemma 4.2). For distinct primes p,qp,q, write ord⁡p(q)\operatorname{ord}_p(q) for the multiplicative order of qq modulo pp. If p∤ℓip\nmid\ell_i, a Sylow pp-subgroup of FF lies in a maximal subgroup MM of FF containing Hi∩FH_i\cap F. The primitive solvable quotient of FF acting on the cosets of MM has prime-power degree qcq^c, and the Sylow subgroup acts on its minimal normal subgroup nontrivially, because FF has no proper normal subgroup of index prime to pp. Counting fixed points shows that each ℓi>1\ell_i>1 is divisible by pp or by qord⁡p(q)q^{\operatorname{ord}_p(q)} for some prime q<pq<p.

Step 4: counting gives a contradiction for p≠7p\ne7 (Proposition 5.1 and Lemma 5.2). Since di=hiℓid_i=h_i\ell_i, we have v(di)=v(hi)+v(ℓi)v(d_i)=v(h_i)+v(\ell_i). By Step 3, every v(di)v(d_i) with i∈Ji\in J lies either in U+v(p)U+v(p) or in U+v(qord⁡p(q))U+v(q^{\operatorname{ord}_p(q)}) for some prime q<pq<p. Shifting a set by v(m)v(m) multiplies its weight by 1/m1/m. Accounting for their overlap, the union of these shifted sets has weight at most cpW(U)c_pW(U), where

cp=1/p+(1−1/p)∑q<pq−ord⁡p(q)c_p=1/p+(1-1/p)\sum_{q<p}q^{-\operatorname{ord}_p(q)}

tends to zero as pp grows. The v(di)v(d_i) are distinct and their weights add up to ν\nu, so ν≤cpW(U)≤cpΠν\nu\le c_pW(U)\le c_p\Pi\nu, which requires cpΠ≥1c_p\Pi\ge1. Here Π\Pi is taken over all primes at most pp. Exact rational estimates for the small primes, an interval argument up to 255, and a uniform elementary bound for p≥256p \ge256 show that cpΠ<1c_p\Pi< 1 for every prime p≥5p \ge5 except 7; the finite inequalities are also checked in the Lean development.

Step 5: the prime 7 (Section 6). At p=7p = 7 a local index may be divisible by 8=2ord⁡7(2)8 = 2^{\operatorname{ord}_7(2)}, and the shifted copy U+v(8)U + v(8) has the comparatively large weight W(U)/8W(U)/8. As a result c7Πc_7\Pi is about 1.1 with all four primes included, and the count of Step 4 falls short of a contradiction by about 10%. We retain information that the generic count discards: which minimal quotient-index vectors actually carry parts of local index exactly 7 or exactly 8, and which parts meet the same FF-coset as such a part (Lemma 6.1). A missing-point and overlap-charging argument turns this incidence information into incompatible bounds (Proposition 6.3).

We use Sun’s coset union bound and the standard structure theory of primitive solvable groups. The new ingredients are the combination of the local divisibility restriction with the quotient count, and the additional argument at the prime 7. Section 8 describes the Lean formalization [14].

Preliminaries

Throughout, N={0,1,2,…}\mathbb{N} = \{0,1,2,\ldots\}. For a subgroup AA of a group GG, its normal core is core⁡G(A)=⋂g∈GgAg−1\operatorname{core}_G(A) = \bigcap_{g\in G} gAg^{-1}.

Coset intersections and reductions

The following standard fact is used repeatedly.

Lemma 2.1 ([17 Lemma 2.1 and Remark 2.2]). Let XX be a finite group and Y,Z≤XY,Z \le X with [X:Y][X:Y] and [X:Z][X:Z] coprime. Then YZ=XYZ = X, and every coset of YY meets every coset of ZZ.

Sun’s Lemma 2.1 states YZ=XYZ = X for subgroups of finite coprime index in any group. His Remark 2.2 states that if aY∩bZ=∅aY \cap bZ = \varnothing for some a,ba,b then the indices are not coprime. The same statement appears in [10 Corollary 2.1] and, in terms of harmonic tuples, in [13 Lemma 2.2].

Lemma 2.2. If every finite solvable group is HS, then every solvable group is HS. More precisely, for a partition (1) of a group GG the subgroup N=core⁡G(⋂iHi)N = \operatorname{core}_G\left(\bigcap_i H_i\right) is normal of finite index, and the cosets giHi/Ng_iH_i/N partition G/NG/N with [G/N:Hi/N]=[G:Hi][G/N:H_i/N] = [G:H_i].

Proof. By Korec and Znám [12 Theorem 1], every HiH_i has finite index in GG. Hence K=⋂iHiK = \bigcap_i H_i has finite index, with [G:K]≤∏i[G:Hi][G:K] \le\prod_i [G:H_i] [12 Lemma 1], and so does its normal core NN, the intersection of the finitely many conjugates of KK [12 Lemma 2]. Each part giHig_iH_i is a union of NN-cosets, so the images are pairwise disjoint and partition G/NG/N, with the same indices. This is [12 Theorem 4]. The group G/NG/N is solvable if GG is.

Lemma 2.3. Let GG be a finite solvable group of least order that is not HS, and fix a partition (1) with k≥2k \ge2 and distinct indices di=[G:Hi]d_i = [G:H_i]. Then

(a) ∑i1/di=1\sum_i 1/d_i = 1 and every di>1d_i > 1;

(b) no two of the did_i are coprime;

(c) no did_i equals 2;

(d) the largest prime divisor pp of ∣G∣|G| is at least 5.

Proof. (a) The parts have sizes ∣G∣/di|G|/d_i and are disjoint. A part of index 1 would be all of GG, but k≥2k \ge2. (The sum is [12 Theorem 2].)

(b) Distinct parts are disjoint, so this follows from Lemma 2.1. (This is also [12 Theorem 3].)

(c) Suppose d1=2d_1 = 2. The other parts partition the complementary coset xH1xH_1. Each gjHj⊆xH1g_jH_j \subseteq xH_1 forces Hj≤H1H_j \le H_1, and translating by x−1x^{-1} gives a partition of H1H_1 into cosets of the HjH_j, j≥2j \ge2, with distinct indices [H1:Hj]=dj/2[H_1:H_j] = d_j/2. If k≥3k \ge3 this partition has at least two parts,

contradicting the minimality of ∣G∣|G|. If k=2k=2 then g2H2=xH1g_2H_2=xH_1 and d2=2=d1d_2=2=d_1, contradicting distinctness. This is the reduction of [10 Lemma 2.3]; see also [13 Lemma 2.3].

(d) If p≤3p\le3, every index is of the form 2a3b2^a3^b. By (b), either all indices are even or all are divisible by 33: if an odd index occurs, it is a nontrivial power of 33 and every other index is divisible by 33. By (c), the index 22 is excluded. We have the following reciprocal sums:

∑a≥1, b≥0(a,b)≠(1,0)2−a3−b=11−1/3−12=1,\sum_{\substack{a\ge1,\ b\ge0\\(a,b)\ne(1,0)}}2^{-a}3^{-b}=\frac{1}{1-1/3}-\frac{1}{2}=1,
∑a≥0, b≥12−a3−b=11−1/2⋅1/31−1/3=1.\sum_{\substack{a\ge0,\ b\ge1}}2^{-a}3^{-b}=\frac{1}{1-1/2}\cdot\frac{1/3}{1-1/3}=1.

The did_i form a finite subset of one of these two infinite index sets, so ∑i1/di<1\sum_i1/d_i<1, contrary to (a).

Part (d) also follows from [10 Theorem B]: every finite group whose order has exactly two prime divisors is HS (a pp-group is nilpotent, so covered by [4]). The argument above keeps the proof self-contained. □

Sun’s coset union bound

A set A⊆ZA\subseteq\mathbb{Z} that is invariant under translation by some positive integer mm has density dens⁡(A)=m−1#(A∩{0,…,m−1})\operatorname{dens}(A)=m^{-1}\#(A\cap\{0,\ldots,m-1\}), which does not depend on the choice of mm. For example, dens⁡(hZ)=1/h\operatorname{dens}(h\mathbb{Z})=1/h.

Theorem 2.4 (Sun’s coset union bound). Let LL be a finite solvable group and let xjLjx_jL_j be a nonempty finite family of cosets of subgroups of LL. Then

∣L∣⋅dens⁡(⋃j[L:Lj]Z)≤∣⋃jxjLj∣.|L|\cdot\operatorname{dens}\left(\bigcup_j[L:L_j]\mathbb{Z}\right)\le\left|\bigcup_jx_jL_j\right|.

This is the case H=1H=1 of [18 Theorem 3.1]. Sun’s theorem concerns a subgroup HH of finite index and subgroups G1,…,GkG_1,\ldots,G_k containing HH, under either of two hypotheses; the second is a composition series from HH to the whole group with factors of prime order, which a finite solvable group has. With H=1H=1, Sun’s conclusion says that the union of the cosets has at least as many elements as there are n∈{0,…,∣L∣−1}n\in\{0,\ldots,|L|-1\} divisible by some [L:Lj][L:L_j], and this number is the left side above. Sun’s theorem has no hypothesis that the cosets cover the group, and we apply it to families whose union is a proper subset. In the nilpotent case the bound goes back to Berger, Felzenbaum and Fraenkel [4].

Exponent vectors and weights

Fix a finite set P\mathcal{P} of primes. For a positive integer nn whose prime divisors lie in P\mathcal{P}, let v(n)=(vq(n))q∈Pv(n)=(v_q(n))_{q\in\mathcal{P}} be its exponent vector. Order NP\mathbb{N}^{\mathcal{P}} coordinatewise. For a∈NPa\in\mathbb{N}^{\mathcal{P}} and E⊆NPE\subseteq\mathbb{N}^{\mathcal{P}}, put

↑a={b∈NP:a≤b},↑E=⋃a∈E↑a.\uparrow a=\{b\in\mathbb{N}^{\mathcal{P}}:a\le b\},\qquad\uparrow E=\bigcup_{a\in E}\uparrow a.

Thus ↑v(n)\uparrow v(n) is the set of vectors of the multiples of nn whose prime divisors lie in P\mathcal{P}. For q∈Pq\in\mathcal{P} let eq=v(q)e_q=v(q) be the unit vector at qq. A set E⊆NPE\subseteq\mathbb{N}^{\mathcal{P}} is upward closed if a∈Ea\in E and a≤ba\le b imply b∈Eb\in E; every set ↑E\uparrow E is upward closed. Define

w(a)=∏q∈Pq−aq,W(E)=∑a∈Ew(a),Π=∏q∈P(1−q−1)−1,w(a)=\prod_{q\in\mathcal{P}}q^{-a_q},\qquad W(E)=\sum_{a\in E}w(a),\qquad\Pi=\prod_{q\in\mathcal{P}}(1-q^{-1})^{-1},

so that w(v(n))=1/nw(v(n))=1/n and W(NP)=ΠW(\mathbb{N}^{\mathcal{P}})=\Pi. For a set EE and a vector bb write E+b={a+b:a∈E}E+b=\{a+b:a\in E\}. Then

W(E+b)=w(b)W(E),W(↑a)=Πw(a).(2)W(E+b)=w(b)W(E),\qquad W(\uparrow a)=\Pi w(a). \tag*{(2)}

Lemma 2.5. Let h1,…,hsh_1,\ldots,h_s be positive integers whose prime divisors lie in P\mathcal{P}, and put U=⋃j↑ν(hj)U=\bigcup_j \uparrow^{\nu(h_j)}. Then

W(U)=Π⋅dens⁡(⋃jhjZ).(3)W(U)=\Pi\cdot\operatorname{dens}\left(\bigcup_j h_j\mathbb{Z}\right). \tag*{(3)}

Proof. This is a reformulation of [18 Lemma 3.4], which Sun attributes to an earlier paper of his; we give the short proof for completeness. For a nonempty S⊆{1,…,s}S\subseteq\{1,\ldots,s\} let hS=lcm⁡j∈Shjh_S=\operatorname{lcm}_{j\in S}h_j. Then ⋂j∈S↑ν(hj)=↑ν(hS)\bigcap_{j\in S}\uparrow^{\nu(h_j)}=\uparrow^{\nu(h_S)} and ⋂j∈ShjZ=hSZ\bigcap_{j\in S}h_j\mathbb{Z}=h_S\mathbb{Z}. By inclusion–exclusion and (2), the two sides of (3) are ∑S≠∅(−1)∣S∣+1Π/hS\sum_{S\ne\varnothing}(-1)^{|S|+1}\Pi/h_S and Π∑S≠∅(−1)∣S∣+1/hS\Pi\sum_{S\ne\varnothing}(-1)^{|S|+1}/h_S.

Quotient and local indices

In this section GG is a finite solvable group with a partition (1), FF is a normal subgroup of GG, and r=[G:F]r=[G:F]. We take P\mathcal{P} to contain every prime divisor of ∣G∣|G|.

Definition 3.1. For each part giHig_iH_i put

di=[G:Hi],hi=[G:HiF],ℓi=[F:Hi∩F].d_i=[G:H_i],\qquad h_i=[G:H_iF],\qquad\ell_i=[F:H_i\cap F].

so that di=hiℓid_i=h_i\ell_i. We call hih_i the quotient index and ℓi\ell_i the local index of the part. The part is universal if ℓi=1\ell_i=1, that is, if F≤HiF\leq H_i. Let J={i:ℓi>1}J=\{i:\ell_i>1\} be the set of indices of the non-universal parts, let Xi⊆G/FX_i\subseteq G/F be the image of giHig_iH_i, and put

ν=∑i∈J1di,U=⋃i∈J↑ν(hi).\nu=\sum_{i\in J}\frac{1}{d_i},\qquad U=\bigcup_{i\in J}\uparrow^{\nu(h_i)}.

The image XiX_i is a coset of HiF/FH_iF/F and has r/hir/h_i elements. If giHig_iH_i meets an FF-coset yFyF in a point y′y', then giHi∩yF=y′(Hi∩F)g_iH_i\cap yF=y'(H_i\cap F) has ∣F∣/ℓi|F|/\ell_i elements.

Lemma 3.2. For every I⊆JI\subseteq J,

W(⋃i∈I↑ν(hi))≤Πr∣⋃i∈IXi∣.W\left(\bigcup_{i\in I}\uparrow^{\nu(h_i)}\right)\leq\frac{\Pi}{r}\left|\bigcup_{i\in I}X_i\right|.

Moreover ∣⋃i∈JXi∣=rν\left|\bigcup_{i\in J}X_i\right|=r\nu, so

(4)

W(U)≤Πν.(4)W(U)\leq\Pi\nu. \tag*{(4)}

Proof. The first inequality is trivial for I=∅I=\varnothing. Otherwise apply Theorem 2.4 in the solvable group G/FG/F to the cosets XiX_i, i∈Ii\in I, whose subgroups HiF/FH_iF/F have index hih_i, and then Lemma 2.5.

An FF-coset met by a non-universal part cannot meet a universal part, since a universal part contains every FF-coset it meets. Hence the union of the non-universal parts is exactly the inverse image of ⋃i∈JXi\bigcup_{i\in J}X_i. Its size is ∣G∣ν|G|\nu, which gives ∣⋃i∈JXi∣=rν\left|\bigcup_{i\in J}X_i\right|=r\nu.

The subgroup F and local divisibility

An element of a finite group is a pp-element if its order is a power of pp. The subgroup generated by the pp-elements is also generated by the Sylow pp-subgroups. For primes p≠qp\ne q, ord⁡p(q)\operatorname{ord}_p(q) denotes the multiplicative order of qq modulo pp.

Lemma 4.1. Let GG be a finite group and pp a prime dividing ∣G∣|G|, and let FF be the subgroup generated by the pp-elements of GG. Then FF is a nontrivial normal subgroup of GG, and every normal subgroup of FF of index prime to pp is FF itself.

Proof. FF is characteristic in GG, and it contains a Sylow pp-subgroup of GG, which is nontrivial. Let N◃FN \triangleleft F with p∤[F:N]p \nmid[F:N]. A pp-subgroup TT of FF maps to a pp-subgroup of F/NF/N, whose order is prime to pp. Thus T≤NT \leq N. Each Sylow pp-subgroup of GG consists of pp-elements, so it is a pp-subgroup of FF and lies in NN. These subgroups generate FF; hence N=FN=F. □

Lemma 4.1 does not require solvability. Only its property concerning normal subgroups of index prime to pp is used in the next lemma.

Lemma 4.2 (Cyclotomic divisibility). Let FF be a finite solvable group and pp a prime such that every normal subgroup of FF of index prime to pp is FF itself. If Y<FY<F and p∤[F:Y]p\nmid[F:Y], then some prime q≠pq\ne p dividing ∣F∣|F| satisfies

qord⁡p(q)∣[F:Y].q^{\operatorname{ord}_p(q)} \mid[F:Y].

Proof. Let PP be a Sylow pp-subgroup of FF. The PP-orbits on F/YF/Y have pp-power sizes summing to [F:Y][F:Y], so PP fixes some coset xYxY, that is, P≤xYx−1P\leq xYx^{-1}. Replacing YY by this conjugate does not change the index, so we may assume P≤YP\leq Y. Choose a maximal subgroup MM of FF containing YY, let R=core⁡F(M)R=\operatorname{core}_F(M), and let V/RV/R be a minimal normal subgroup of the solvable group F/RF/R. Then V/RV/R is elementary abelian of order qcq^c for a prime qq, and c≥1c\geq1.

We have V≰MV\nleq M, since otherwise the normal subgroup VV would lie in RR. Hence VM=FVM=F by maximality. The subgroup V∩MV\cap M contains RR, so it is normal in VV, as V/RV/R is abelian; it is also normalized by MM, hence normal in VM=FVM=F. As RR is the largest normal subgroup of FF contained in MM, this gives V∩M≤RV\cap M\leq R, and therefore V∩M=RV\cap M=R. So [F:M]=[V:V∩M]=qc[F:M]=[V:V\cap M]=q^c. As [F:M][F:M] divides [F:Y][F:Y], which is prime to pp, we get q≠pq\ne p; and qq divides ∣F∣|F|.

Since VM=FVM=F and V∩M=RV\cap M=R, the map φ:V/R→F/M\varphi:V/R\to F/M, vR↦vMvR\mapsto vM, is a bijection. For x∈P≤Mx\in P\leq M it satisfies

φ(xvx−1R)=(xvx−1)M=xvM=xφ(vR).\varphi(xvx^{-1}R)=(xvx^{-1})M=xvM=x\varphi(vR).

Thus it carries conjugation by PP on V/RV/R to left multiplication by PP on F/MF/M. If PP fixed every point of F/MF/M, then P≤RP\leq R; but R◃FR\triangleleft F contains a Sylow pp-subgroup, so [F:R][F:R] is prime to pp and R=FR=F by hypothesis, contrary to R≤M<FR\leq M<F. Hence the fixed subgroup CV/R(P)C_{V/R}(P) has order qbq^b with b<cb<c. The non-singleton PP-orbits on V/RV/R have size divisible by pp, so qc≡qb(modp)q^c\equiv q^b\pmod p, hence qc−b≡1(modp)q^{c-b}\equiv1\pmod p and ord⁡p(q)≤c−b≤c\operatorname{ord}_p(q)\leq c-b\leq c. Therefore qord⁡p(q)∣qc=[F:M]q^{\operatorname{ord}_p(q)}\mid q^c=[F:M], which divides [F:Y][F:Y]. □

The argument is the usual one for solvable primitive groups, where a minimal normal subgroup of the primitive quotient F/RF/R acts regularly; compare [16 Theorem 7.2.6]. Counting fixed points avoids a separate appeal to the orders of general linear groups.

Largest prime other than 7

For a partition and normal subgroup FF, we use the notation of Definition 3.1: di=hiℓid_i=h_i\ell_i, J={i:ℓi>1}J=\{i:\ell_i>1\}, ν=∑i∈J1/di\nu=\sum_{i\in J}1/d_i, and U=⋃i∈JtV(hi)U=\bigcup_{i\in J}{}^tV(h_i). The weight WW and the product Π\Pi are as in Section 2, for the specified prime set P\mathcal{P}.

For a prime pp put

Πp=∏q<p(1−q−1)−1,Sp=∑q<pq−ord⁡p(q),cp=1p+(1−1p)Sp.\Pi_p=\prod_{q<p}(1-q^{-1})^{-1},\qquad S_p=\sum_{q<p}q^{-\operatorname{ord}_p(q)},\qquad c_p=\frac{1}{p}+\left(1-\frac{1}{p}\right)S_p.

where here and below sums and products indexed by qq run over primes.

Proposition 5.1. Let GG be a finite solvable group with a partition (1) whose indices are distinct, let p≥5p\geq5 be the largest prime divisor of ∣G∣|G|, and let FF be the subgroup generated by the pp-elements

of GG; use the notation of Definition 3.1 for this FF. If some part is not universal relative to FF, then

1≤Πp(Sp+1p−1).(5)1 \le\Pi_p\left(S_p+\frac{1}{p-1}\right). \tag*{(5)}

Proof. Take P\mathcal{P} to be the set of primes at most pp, so that Π=pp−1Πp\Pi=\frac{p}{p-1}\Pi_p and cpΠ=Πp(Sp+1p−1)c_p\Pi=\Pi_p\left(S_p+\frac{1}{p-1}\right). Write fq=ord⁡p(q)f_q=\operatorname{ord}_p(q). For i∈Ji\in J the local index ℓi\ell_i is the index of a proper subgroup Hi∩FH_i\cap F of FF. By Lemmas 4.1 and 4.2, either p∣ℓip\mid\ell_i or qfq∣ℓiq^{f_q}\mid\ell_i for a prime q≠pq\ne p dividing ∣G∣|G|, and such a qq is less than pp. So every vector v(di)v(d_i) with i∈Ji\in J lies in

(U+ep)∪⋃q<p(U+fqeq).(U+e_p)\cup\bigcup_{q<p}(U+f_qe_q).

The vectors v(di)v(d_i), i∈Ji\in J, are distinct, and their total weight is ν\nu. For each q<pq<p, the intersection of U+epU+e_p and U+fqeqU+f_qe_q contains U+ep+fqeqU+e_p+f_qe_q, since UU is upward closed. Counting U+epU+e_p first and then the portions of the other sets outside it, and using (2), gives

ν≤W(U+ep)+∑q<pW((U+fqeq)∖(U+ep))≤[1p+∑q<p(q−fq−p−1q−fq)]W(U)=cpW(U).\begin{aligned} \nu\le W(U+e_p)+\sum_{q<p}W\bigl((U+f_qe_q)\setminus(U+e_p)\bigr) \\ &\le\left[\frac{1}{p}+\sum_{q<p}\left(q^{-f_q}-p^{-1}q^{-f_q}\right)\right]W(U)=c_pW(U). \end{aligned}

By Lemma 3.2, which applies Sun’s bound in G/FG/F, the right side is at most cpΠν=Πp(Sp+1p−1)νc_p\Pi\nu=\Pi_p\left(S_p+\frac{1}{p-1}\right)\nu. As ν>0\nu>0, the assertion follows. □

Lemma 5.2. For every prime p≥5p\ge5,

Πp(Sp+1p−1)<1⟺p≠7.\Pi_p\left(S_p+\frac{1}{p-1}\right)<1\Longleftrightarrow p\ne7.

Proof. At 5 the expression is 421/432421/432; at 7 it is

154(18+1729+115625+16)≈1.099>1.\frac{15}{4}\left(\frac{1}{8}+\frac{1}{729}+\frac{1}{15625}+\frac{1}{6}\right)\approx1.099>1.

For the other primes up to 31 the following upper bounds suffice. Each bound on SpS_p is obtained by finding the least f≥1f\ge1 with qf≡1(modp)q^f\equiv1\pmod p for each q<pq<p and adding q−fq^{-f}. For example, modulo 13 the orders of 2, 3, 5, 7, 11 are 12, 3, 4, 12, 12, so S13=3−3+5−4+2−12+7−12+11−12<389/10000S_{13}=3^{-3}+5^{-4}+2^{-12}+7^{-12}+11^{-12}<389/10000; modulo 29 the two largest terms are 17−417^{-4} and 7−77^{-7} (the orders of 17 and 7 are 4 and 7), and the other terms are below 10−810^{-8} in total, giving S29<1/70000S_{29}<1/70000.

ppupper bound for Πp\Pi_pupper bound for SpS_p
1135/835/81/1801/180
1377/1677/16389/10000389/10000
1721/421/41/2501/250
1928/528/51/2701/270
2359/1059/101/20001/2000
2931/531/51/700001/70000
31317/50317/501/32+1/125+1/70001/32+1/125+1/7000

Table 1.

Let q<pq<p be prime, with p≥5p\ge5. Since qord⁡p(q)≡1(modp)q^{\operatorname{ord}_p(q)}\equiv1\pmod p and qord⁡p(q)>1q^{\operatorname{ord}_p(q)}>1, we have qord⁡p(q)≥p+1q^{\operatorname{ord}_p(q)}\ge p+1. Moreover ord⁡p(q)≥3\operatorname{ord}_p(q)\ge3: order 1 would give p∣q−1p\mid q-1, and order 2 would give p∣q+1p \mid q+1, hence q=p−1q=p-1, which is even and at least 4. So

Sp≤∑q<pmin⁡{q−3,(p+1)−1}.(6)S_p \le\sum_{q<p}\min\{q^{-3},(p+1)^{-1}\}. \tag*{(6)}

For 37≤p≤25137\le p\le251, monotonicity of Πp\Pi_p and (6) bound the expression in the lemma by

Π251(∑q<251min⁡{q−3,38−1}+136)<924410000.(7)\Pi_{251}\left(\sum_{q<251}\min\{q^{-3},38^{-1}\}+\frac{1}{36}\right)<\frac{9244}{10000}. \tag*{(7)}

by summing the rational terms. No prime lies strictly between 251 and 256.

The range p≥256p\ge256. We first bound Πp\Pi_p. Every prime greater than 7 is of the form 6m−16m-1 or 6m+16m+1 with m≥2m\ge2. Put L=⌈p/6⌉L=\lceil p/6\rceil. All primes 7<q<p7<q<p occur among these pairs for 2≤m≤L2\le m\le L. Including both entries of every pair, even when an entry is composite or at least pp, only increases the product, since each factor is greater than 1. Hence

Πp≤358∏m=2L(6m−1)(6m+1)(6m−2)6m.\Pi_p\le\frac{35}{8}\prod_{m=2}^{L}\frac{(6m-1)(6m+1)}{(6m-2)6m}.

For m≥2m\ge2 the pair factor satisfies

((6m−1)(6m+1)(6m−2)6m)3≤12m+512m−7.\left(\frac{(6m-1)(6m+1)}{(6m-2)6m}\right)^3\le\frac{12m+5}{12m-7}.

To verify this inequality, write the right side minus the left side over the positive denominator (12m−7)((6m−2)6m)3(12m-7)((6m-2)6m)^3. Its numerator is

(12m+5)((6m−2)6m)3−(12m−7)((6m−1)(6m+1))3.(12m+5)((6m-2)6m)^3-(12m-7)((6m-1)(6m+1))^3.

Writing M=m−1≥1M=m-1\ge1, expansion gives

29808M4+109296M3+149796M2+90948M+20633>0.29808M^4+109296M^3+149796M^2+90948M+20633>0.

The resulting product telescopes, because 12m−7=12(m−1)+512m-7=12(m-1)+5:

∏m=2L12m+512m−7=2917⋅4129⋯12L+512L−7=12L+517.\prod_{m=2}^{L}\frac{12m+5}{12m-7} =\frac{29}{17}\cdot\frac{41}{29}\cdots\frac{12L+5}{12L-7} =\frac{12L+5}{17}.

Since L≤p/6+1L\le p/6+1, we obtain

Πp3≤(358)312L+517≤(358)32p+1717.(8)\Pi_p^3\le\left(\frac{35}{8}\right)^3\frac{12L+5}{17} \le\left(\frac{35}{8}\right)^3\frac{2p+17}{17}. \tag*{(8)}

Next let tt be the integer with t3≤p<(t+1)3t^3\le p<(t+1)^3. As p≥256>63p\ge256>6^3, we have t≥6t\ge6. In (6), bound each of the at most tt terms with q≤tq\le t by (p+1)−1(p+1)^{-1}, and the remaining terms by q−3q^{-3}. Extending the latter sum to all integers greater than tt gives

Sp≤tp+1+∑n=t+1∞n−3.S_p\le\frac{t}{p+1}+\sum_{n=t+1}^{\infty}n^{-3}.

For n≥2n\ge2,

12((n−1)−2−n−2)−n−3=3n−22n3(n−1)2>0.\frac{1}{2}\left((n-1)^{-2}-n^{-2}\right)-n^{-3} =\frac{3n-2}{2n^3(n-1)^2}>0.

Consequently, for N>tN>t,

∑n=t+1Nn−3≤12∑n=t+1N((n−1)−2−n−2)=12t2−12N2.\sum_{n=t+1}^{N}n^{-3}\le\frac{1}{2}\sum_{n=t+1}^{N}\left((n-1)^{-2}-n^{-2}\right) =\frac{1}{2t^2}-\frac{1}{2N^2}.

Letting N→∞N \to\infty and using p+1≥t3p+1 \ge t^3 yields

Sp≤tp+1+12t2≤1t2+12t2.S_p \le\frac{t}{p+1}+\frac{1}{2t^2}\le\frac{1}{t^2}+\frac{1}{2t^2}.

Also p−1≥t3−1≥6t2−1≥5t2p-1\ge t^3-1\ge6t^2-1\ge5t^2, so

Sp+1p−1≤1t2+12t2+15t2=1710t2.S_p+\frac{1}{p-1}\le\frac{1}{t^2}+\frac{1}{2t^2}+\frac{1}{5t^2}=\frac{17}{10t^2}.

Since p≥256p\ge256, we have 2p+17≤(529/256)p2p+17\le(529/256)p. Thus (8) gives

Πp3≤(358)35294352p<(358)35294352(t+1)3.\Pi_p^3\le\left(\frac{35}{8}\right)^3\frac{529}{4352}p<\left(\frac{35}{8}\right)^3\frac{529}{4352}(t+1)^3.

Finally, (t+1)/t2=1/t+1/t2≤7/36(t+1)/t^2=1/t+1/t^2\le7/36 for t≥6t\ge6, and hence

(Πp(Sp+1p−1))3<(358⋅1710⋅t+1t2)35294352≤(833576)35294352<(32)35294352<12.\left(\Pi_p\left(S_p+\frac{1}{p-1}\right)\right)^3< \left(\frac{35}{8}\cdot\frac{17}{10}\cdot\frac{t+1}{t^2}\right)^3\frac{529}{4352} \le\left(\frac{833}{576}\right)^3\frac{529}{4352} <\left(\frac{3}{2}\right)^3\frac{529}{4352}<\frac{1}{2}.

Taking cube roots proves the required bound for p≥256p\ge256.

All the finite inequalities in this proof are also checked by exact rational arithmetic in the formalization. For the row p=31p=31, the orders ord⁡31(q)\operatorname{ord}_{31}(q) for q=2,3,5,7,11,13,17,19,23,29q=2,3,5,7,11,13,17,19,23,29 are 5,30,3,15,30,30,30,15,10,105,30,3,15,30,30,30,15,10,10. Remark 6.4 shows that the failure at p=7p=7 cannot be removed merely by sharpening the uniform weight bound for the same divisibility shells.

The Python script solvable_paper_numerics.py, provided as an ancillary file with the arXiv version of this paper, checks the small-prime values, every bound in the table, the interval estimate (7), and the polynomial and telescoping identities in the argument for p≥256p\ge256, using exact arithmetic. It requires only the Python standard library and can be run with

python solvable_paper_numerics.py

The prime 7

The generic count of Proposition 5.1 uses only the divisibility of local indices. At p=7p=7 it fails: 8=2ord⁡7(2)8=2^{\operatorname{ord}_{7}(2)} is one of the relevant cyclotomic divisors, the shifted set U+v(8)U+v(8) has weight W(U)/8W(U)/8, and by Lemma 5.2 the right side of (5) is about 1.1. We therefore use two kinds of information that the generic count ignores. The first is which minimal heads (defined below) have no part of local index exactly 7 or exactly 8; these give the missing points. The second is which parts meet the same FF-coset as a part of local index 7 or 8 (Lemma 6.1). Together these yield incompatible lower and upper bounds involving the weights of the missing points.

In this section GG, the partition, FF, JJ, UU and ν\nu are as in Proposition 5.1 with p=7p=7, and P={2,3,5,7}\mathcal{P}=\{2,3,5,7\}. In particular,

r=[G:F],di=hiℓi,J={i:ℓi>1},ν=∑i∈J1di,U=⋃i∈J↑v(hi).r=[G:F],\qquad d_i=h_i\ell_i,\qquad J=\{i:\ell_i>1\},\qquad\nu=\sum_{i\in J}\frac{1}{d_i},\qquad U=\bigcup_{i\in J}\uparrow v(h_i).

We write

Π=76Π7=358,α=e7=v(7),β=3e2=v(8),ε=3−6+5−6.(9)\Pi=\frac{7}{6}\Pi_7=\frac{35}{8},\qquad\alpha=e_7=v(7),\qquad\beta=3e_2=v(8),\qquad\varepsilon=3^{-6}+5^{-6}. \tag*{(9)}

Since ord⁡7(2)=3\operatorname{ord}_7(2)=3 and ord⁡7(3)=ord⁡7(5)=6\operatorname{ord}_7(3)=\operatorname{ord}_7(5)=6, Lemma 4.2 shows that the local index of every non-universal part is divisible by at least one of 77, 88, 363^6, 565^6.

After Lemma 6.1, which needs only the notation introduced so far, we introduce the minimal heads H\mathcal{H}, the sets AA and BB, the missing points Z7Z_7 and Z8Z_8, and the sets DD and TT; these are used throughout the rest of the section.

Lemma 6.1. Let two distinct parts meet the same FF-coset. Then their local indices are not coprime. If one of them has local index 7, the other has local index divisible by 7. If one of them has local index 8, the other has local index divisible by 7 or by 8.

Proof. Both parts are non-universal, since a universal part contains every FF-coset it meets. If giHig_iH_i meets yFyF in a point y′y', then giHi∩yF=y′(Hi∩F)g_iH_i \cap yF = y'(H_i \cap F). So two parts i,ji,j meeting yFyF give disjoint pieces y′(Hi∩F)y'(H_i \cap F) and y′′(Hj∩F)y''(H_j \cap F). Translating by y′−1{y'}^{-1}, we obtain disjoint cosets Y0Y_0 and uYuY in FF, where Y0=Hi∩FY_0 = H_i \cap F, Y=Hj∩FY = H_j \cap F, and u=y′−1y′′∈Fu = {y'}^{-1}y'' \in F because y′y' and y′′y'' lie in the same left coset yFyF. The indices are the local indices ℓi,ℓj\ell_i,\ell_j. By Lemma 2.1, ℓi\ell_i and ℓj\ell_j are not coprime. This proves the first assertion, and the second follows because 7 is prime.

For the third, suppose [F:Y0]=8[F:Y_0]=8. If 7∣[F:Y]7 \mid[F:Y] there is nothing to prove. Otherwise YY contains a Sylow 7-subgroup PP of FF. The action of PP on the eight cosets of Y0Y_0 is nontrivial. Indeed, if PP lay in the core of Y0Y_0, that normal subgroup of FF would have index prime to 7, so it would be FF by Lemma 4.1, contrary to [F:Y0]=8[F:Y_0]=8. A nontrivial 7-group action on eight points has one orbit of size 7 and one fixed point x0x_0. Since P≤YP \leq Y, the YY-orbit of x0x_0 and its complement are PP-invariant, so the complement has size 0 or 7. In the first case YY is transitive on F/Y0F/Y_0, so F=YY0F=YY_0. Then uY∩Y0≠∅uY \cap Y_0 \ne\varnothing, since u−1∈F=YY0u^{-1} \in F=YY_0, contrary to the two disjoint pieces. In the second case YY fixes x0x_0 and is contained in its stabilizer, a conjugate of Y0Y_0. Thus 8∣[F:Y]8 \mid[F:Y]. □

We call v(hi)v(h_i) the head of part ii. Let H\mathcal{H} be the set of minimal elements of {v(hi):i∈J}\{v(h_i):i\in J\}, so that H\mathcal{H} is a finite antichain and U=↑HU=\mathord{\uparrow}\mathcal{H}. We call the elements of H\mathcal{H} the minimal heads. Let A⊆HA \subseteq\mathcal{H} consist of those η\eta for which some part has head η\eta and local index exactly 7, and let B⊆HB \subseteq\mathcal{H} consist of those η\eta for which some part has head η\eta and local index exactly 8. Set

Z7={η+α:η∈H∖A},Z8={η+β:η∈H∖B},z7=W(Z7),z8=W(Z8).Z_7=\{\eta+\alpha:\eta\in\mathcal{H}\setminus A\},\qquad Z_8=\{\eta+\beta:\eta\in\mathcal{H}\setminus B\},\qquad z_7=W(Z_7),\qquad z_8=W(Z_8).

We call the elements of Z7∪Z8Z_7\cup Z_8 the missing points. A missing point may still be the vector of some other part if it lies in both U+αU+\alpha and U+βU+\beta. To handle this we use the overlap set

D=((U+α)∩(U+β))∖(U+α+β),D=((U+\alpha)\cap(U+\beta))\setminus(U+\alpha+\beta),

and we put

T=(U+α)∪(U+β).T=(U+\alpha)\cup(U+\beta).

Inclusion–exclusion and (2) give

W(T)+W(D)=(17+18−156)W(U)=14W(U).(10)W(T)+W(D)=\left(\frac{1}{7}+\frac{1}{8}-\frac{1}{56}\right)W(U)=\frac{1}{4}W(U). \tag*{(10)}

Lemma 6.2 (Missing points in the overlap). With the notation above:

(1) 74W(Z7∩D)+W((Z8∖Z7)∩D)≤W(D);\frac{7}{4}W(Z_7\cap D)+W((Z_8\setminus Z_7)\cap D)\leq W(D);

(2) Z7∩Z8⊆D;Z_7\cap Z_8\subseteq D;

(3) if i∈Ji\in J and 7∣ℓi7\mid\ell_i or 8∣ℓi8\mid\ell_i, then v(di)v(d_i) lies in neither Z7∖DZ_7\setminus D nor Z8∖DZ_8\setminus D.

Proof. We use repeatedly that for η∈H\eta\in\mathcal{H} and x∈Ux\in U, x≤ηx\leq\eta implies x=ηx=\eta: some η′∈H\eta'\in\mathcal{H} has η′≤x≤η\eta'\leq x\leq\eta, and η′=η\eta'=\eta because H\mathcal{H} is an antichain. In particular, for η∈H\eta\in\mathcal{H} neither η+α\eta+\alpha nor η+β\eta+\beta lies in U+α+βU+\alpha+\beta, since η+α=x+α+β\eta+\alpha=x+\alpha+\beta or η+β=x+α+β\eta+\beta=x+\alpha+\beta with x∈Ux\in U would give x<ηx<\eta.

(1) For each η+α∈Z7∩D\eta+\alpha\in Z_7\cap D and m∈{0,1,2}m\in\{0,1,2\}, consider cm(η)=η+α+me2c_m(\eta)=\eta+\alpha+m e_2. Each cm(η)c_m(\eta) lies in U+αU+\alpha, because η+me2∈U\eta+m e_2\in U, and in U+βU+\beta, because η+α∈D⊆U+β\eta+\alpha\in D\subseteq U+\beta and U+βU+\beta is upward closed. None lies in U+α+β=U+α+3e2U+\alpha+\beta=U+\alpha+3e_2: an equality cm(η)=x+α+3e2c_m(\eta)=x+\alpha+3e_2 with x∈Ux\in U gives x=η−(3−m)e2<ηx=\eta-(3-m)e_2<\eta. Hence c0(η),c1(η),c2(η)∈Dc_0(\eta),c_1(\eta),c_2(\eta)\in D, and their weights add to (1+1/2+1/4)w(η+α)=74w(η+α)(1+1/2+1/4)w(\eta+\alpha)=\frac{7}{4}w(\eta+\alpha).

These triples are disjoint. If cm(η)=cm′(η′)c_m(\eta)=c_{m'}(\eta') then η−η′=(m′−m)e2\eta-\eta'=(m'-m)e_2, so η,η′∈H\eta,\eta'\in\mathcal{H} are comparable; the antichain property gives η=η′\eta=\eta' and then m=m′m=m'. They also avoid (Z8∖Z7)∩D(Z_8\setminus Z_7)\cap D. The point c0(η)=η+αc_0(\eta)=\eta+\alpha lies in Z7Z_7. For m∈{1,2}m\in\{1,2\} suppose cm(η)=η′+βc_m(\eta)=\eta'+\beta with η′∈H\eta'\in\mathcal{H}, so η′=η+α−(3−m)e2\eta'=\eta+\alpha-(3-m)e_2. Since η+α∈U+β\eta+\alpha\in U+\beta, some η′′∈H\eta''\in\mathcal{H} satisfies η′′≤η+α−3e2<η′\eta''\leq\eta+\alpha-3e_2<\eta', contradicting the antichain property. The disjoint triples and (Z8∖Z7)∩D(Z_8\setminus Z_7)\cap D all lie in DD, which proves (1).

(2) A point of Z7∩Z8Z_7\cap Z_8 lies in (U+α)∩(U+β)(U+\alpha)\cap(U+\beta), and by the first paragraph it does not lie in U+α+βU+\alpha+\beta.

(3) Put x=v(hi)∈Ux=v(h_i)\in U and l=v(ℓi)l=v(\ell_i), so that v(di)=x+lv(d_i)=x+l. Suppose first that v(di)∈Z7∖Dv(d_i)\in Z_7\setminus D, so that x+l=η+αx+l=\eta+\alpha for some η∈H∖A\eta\in\mathcal{H}\setminus A. If 7∣ℓi7\mid\ell_i, then l≥αl\geq\alpha, hence x≤ηx\leq\eta. Thus x=ηx=\eta and l=αl=\alpha, so part ii has head η\eta and local index 77, contrary to η∉A\eta\notin A. If 8∣ℓi8\mid\ell_i, then l≥βl\geq\beta, so x+l∈U+βx+l\in U+\beta. It also lies in U+αU+\alpha, and by the first paragraph it does not lie in U+α+βU+\alpha+\beta. Hence x+l∈Dx+l\in D, again a contradiction.

Now suppose that v(di)∈Z8∖Dv(d_i)\in Z_8\setminus D, so that x+l=η+βx+l=\eta+\beta for some η∈H∖B\eta\in\mathcal{H}\setminus B. If 8∣ℓi8\mid\ell_i, then l≥βl\geq\beta, hence x≤ηx\leq\eta. Thus x=ηx=\eta and l=βl=\beta, so part ii has head η\eta and local index 88, contrary to η∉B\eta\notin B. If 7∣ℓi7\mid\ell_i, then l≥αl\geq\alpha, so x+l∈U+αx+l\in U+\alpha. It also lies in U+βU+\beta, and by the first paragraph it does not lie in U+α+βU+\alpha+\beta. Hence x+l∈Dx+l\in D, a contradiction. □

Proposition 6.3. Let GG be a finite solvable group with a partition (1) whose indices are distinct, suppose that 77 is the largest prime divisor of ∣G∣|G|, and let FF be the subgroup generated by the 77-elements of GG. Then every part is universal relative to FF.

Proof. Suppose, for a contradiction, that J≠∅J\neq\varnothing. Then U≠∅U\neq\varnothing. Put

λ=W(U)Π>0,z7,out=W(Z7∖D),z8,out=W(Z8∖D).\lambda=\frac{W(U)}{\Pi}>0,\qquad z_{7,\mathrm{out}}=W(Z_7\setminus D),\qquad z_{8,\mathrm{out}}=W(Z_8\setminus D).

By (4), λ≤ν\lambda\leq\nu.

Step 1: a lower bound for the missing points. Let Ω7⊆G/F\Omega_7\subseteq G/F be the union of the images XiX_i of the parts with local index exactly 77 and head in AA, and define Ω8\Omega_8 likewise using local index 88 and BB. Put ρ7=∣Ω7∣/r\rho_7=|\Omega_7|/r and ρ8=∣Ω8∣/r\rho_8=|\Omega_8|/r. Membership of a head in AA or BB asserts the existence of a part of local index 77 or 88 with that head; we select only parts with the specified local index. By Lemma 3.2, W(↑A)≤Πρ7W(\uparrow A)\leq\Pi\rho_7 and W(↑B)≤Πρ8W(\uparrow B)\leq\Pi\rho_8. Parts of local index 77 or 88 whose heads are not minimal are not needed here. Since U=↑A∪⋃η∈H∖A↑ηU=\uparrow A\cup\bigcup_{\eta\in\mathcal{H}\setminus A}\uparrow\eta and W(↑η)=Πw(η)=7Πw(η+α)W(\uparrow\eta)=\Pi w(\eta)=7\Pi w(\eta+\alpha) by (2),

W(U)≤Πρ7+7Π∑η∈H∖Aw(η+α)=Π(ρ7+7z7).W(U)\leq\Pi\rho_7+7\Pi\sum_{\eta\in\mathcal{H}\setminus A}w(\eta+\alpha)=\Pi(\rho_7+7z_7).

With w(η)=8w(η+β)w(\eta)=8w(\eta+\beta) the same argument for BB gives

λ≤ρ7+7z7,λ≤ρ8+8z8.(11)\lambda\leq\rho_7+7z_7,\qquad\lambda\leq\rho_8+8z_8. \tag*{(11)}

The sets Ω7\Omega_7 and Ω8\Omega_8 are disjoint: an FF-coset in both would be met by parts of local indices 77 and 88, contrary to Lemma 6.1. Let IΩI_\Omega be the set of parts meeting the inverse image of Ω7∪Ω8\Omega_7\cup\Omega_8 in GG, and let VΩ={v(di):i∈IΩ}V_\Omega=\{v(d_i):i\in I_\Omega\}. These parts are non-universal, because each such FF-coset is met by a non-universal part. Their union contains the inverse image of Ω7∪Ω8\Omega_7\cup\Omega_8. For the quotient map π:G→G/F\pi:G\to G/F, this inverse image has proportion

∣π−1(Ω7∪Ω8)∣∣G∣=∣F∣(∣Ω7∣+∣Ω8∣)∣F∣r=ρ7+ρ8.\frac{|\pi^{-1}(\Omega_7\cup\Omega_8)|}{|G|} =\frac{|F|(|\Omega_7|+|\Omega_8|)}{|F|r} =\rho_7+\rho_8.

As the did_i are distinct,

ρ7+ρ8≤∑i∈IΩ1di=W(VΩ).\rho_7+\rho_8 \le\sum_{i\in I_\Omega}\frac{1}{d_i}=W(V_\Omega).

By Lemma 6.1, the local index of each part in IΩI_\Omega is divisible by 7 or 8. Hence VΩ⊆TV_\Omega\subseteq T, and Lemma 6.2(3) shows that VΩV_\Omega does not meet Z7∖D⊆TZ_7\setminus D\subseteq T. Therefore, using (10),

ρ7+ρ8+z7,out≤W(T)=14W(U)−W(D)=3532λ−W(D).(12)\rho_7+\rho_8+z_{7,\mathrm{out}}\le W(T)=\frac{1}{4}W(U)-W(D)=\frac{35}{32}\lambda-W(D). \tag*{(12)}

Adding the two inequalities (11) and using (12),

2λ≤ρ7+ρ8+7z7+8z8≤3532λ−W(D)−z7,out+7z7+8z8.2\lambda\le\rho_7+\rho_8+7z_7+8z_8\le\frac{35}{32}\lambda-W(D)-z_{7,\mathrm{out}}+7z_7+8z_8.

Since z7=W(Z7∩D)+z7,out≤W(D)+z7,outz_7=W(Z_7\cap D)+z_{7,\mathrm{out}}\le W(D)+z_{7,\mathrm{out}}, this gives 2932λ≤6z7+8z8\frac{29}{32}\lambda\le6z_7+8z_8, that is,

34z7+z8≥29256λ.(13)\frac{3}{4}z_7+z_8\ge\frac{29}{256}\lambda. \tag*{(13)}

Step 2: an upper bound for the missing points. Divide the non-universal parts by local divisibility: J78={i∈J:7∣ℓi or 8∣ℓi}J_{78}=\{i\in J:7\mid\ell_i\text{ or }8\mid\ell_i\}, J35=J∖J78J_{35}=J\setminus J_{78}, and let V78V_{78} and V35V_{35} be the corresponding sets of vectors v(di)v(d_i). They are disjoint because the did_i are distinct, and ν=W(V78)+W(V35)\nu=W(V_{78})+W(V_{35}). We have V78⊆TV_{78}\subseteq T, and by Lemma 6.2(3) the set V78V_{78} meets neither Z7∖DZ_7\setminus D nor Z8∖DZ_8\setminus D. These two sets lie in TT and are disjoint by Lemma 6.2(2). Hence W(V78)+z7,out+z8,out≤W(T)W(V_{78})+z_{7,\mathrm{out}}+z_{8,\mathrm{out}}\le W(T). For i∈J35i\in J_{35}, the local index is divisible by 363^6 or 565^6, so V35⊆(U+6e3)∪(U+6e5)V_{35}\subseteq(U+6e_3)\cup(U+6e_5) and W(V35)≤εW(U)W(V_{35})\le\varepsilon W(U). Adding and using (10),

ν+z7,out+z8,out≤W(T)+εW(U)=14W(U)−W(D)+εW(U).\nu+z_{7,\mathrm{out}}+z_{8,\mathrm{out}}\le W(T)+\varepsilon W(U)=\frac{1}{4}W(U)-W(D)+\varepsilon W(U).

Since ν≥λ\nu\ge\lambda by Lemma 3.2 and W(U)=358λW(U)=\frac{35}{8}\lambda, this gives

W(D)+z7,out+z8,out≤(332+358ε)λ.(14)W(D)+z_{7,\mathrm{out}}+z_{8,\mathrm{out}}\le\left(\frac{3}{32}+\frac{35}{8}\varepsilon\right)\lambda. \tag*{(14)}

The coefficient 34=68\frac{3}{4}=\frac{6}{8} in (13) comes from 7z7−z77z_7-z_7 and 8z88z_8. A point of Z7∩DZ_7\cap D may also lie in Z8Z_8, so below it contributes up to 34+1=74\frac{3}{4}+1=\frac{7}{4} times its weight, which is exactly the factor in Lemma 6.2(1).

Step 3: conclusion. Put γ=W(Z7∩D)\gamma=W(Z_7\cap D) and δ=W((Z8∖Z7)∩D)\delta=W((Z_8\setminus Z_7)\cap D). Then z7=z7,out+γz_7=z_{7,\mathrm{out}}+\gamma. By Lemma 6.2(2), Z7∩Z8⊆Z7∩DZ_7\cap Z_8\subseteq Z_7\cap D, so splitting Z8Z_8 into its parts outside DD, in D∖Z7D\setminus Z_7, and in Z7Z_7 gives z8≤z8,out+δ+γz_8\le z_{8,\mathrm{out}}+\delta+\gamma. By Lemma 6.2(1),

34z7+z8≤34z7,out+z8,out+74γ+δ≤34z7,out+z8,out+W(D)≤z7,out+z8,out+W(D).\frac{3}{4}z_7+z_8\le\frac{3}{4}z_{7,\mathrm{out}}+z_{8,\mathrm{out}}+\frac{7}{4}\gamma+\delta\le\frac{3}{4}z_{7,\mathrm{out}}+z_{8,\mathrm{out}}+W(D)\le z_{7,\mathrm{out}}+z_{8,\mathrm{out}}+W(D).

Combining this with (13) and (14) gives

29256≤332+358(1729+115625),\frac{29}{256}\le\frac{3}{32}+\frac{35}{8}\left(\frac{1}{729}+\frac{1}{15625}\right),

which is false: the left side exceeds the right side by 7727329/5832000007727329/583200000. □

Remark 6.4. The shortfall is not an artifact of the union bound in the proof of Proposition 5.1. Let P={2,3,5,7}P=\{2,3,5,7\}, list the shifts α,β,6e3,6e5\alpha,\beta,6e_3,6e_5 as a1,…,a4a_1,\ldots,a_4, and let U⊆NPU\subseteq\mathbb{N}^P be nonempty and upward closed. Put Vm=U∖⋃j<m(U+aj)V_m=U\setminus\bigcup_{j<m}(U+a_j). Since UU is upward closed, Vm∩(U+am)⊆Vm+amV_m\cap(U+a_m)\subseteq V_m+a_m, so W(Vm+1)≥(1−w(am))W(Vm)W(V_{m+1})\ge(1-w(a_m))W(V_m). Iterating, and using U+aj⊆UU+a_j\subseteq U,

W(⋃m(U+am))≤cW(U),c=1−∏m(1−w(am)).W\left(\bigcup_m(U+a_m)\right)\le cW(U),\qquad c=1-\prod_m(1-w(a_m)).

For U=NPU = \mathbb{N}^{\mathcal{P}} the four shifts lie on distinct coordinates, so U∖⋃m(U+am)U \setminus\bigcup_m(U+a_m) is a product and equality holds. Thus cc is the best uniform constant, and still 358c≈1.098>1\frac{35}{8}c \approx1.098 > 1.

Proof of the main theorem

Proof of Theorem 1.2. By Lemma 2.2 it suffices to treat finite groups. Suppose that GG is a finite solvable group of least order that is not HS, and fix a partition with distinct indices. By Lemma 2.3, the largest prime divisor pp of ∣G∣|G| is at least 5. Let FF be the subgroup generated by the pp-elements of GG. By Lemma 4.1, it is a nontrivial normal subgroup. If every part is universal relative to FF, the partition descends to a partition of the smaller solvable group G/FG/F with the same indices, contradicting minimality. Otherwise Proposition 5.1 and Lemma 5.2 exclude p≠7p \ne7, and Proposition 6.3 excludes p=7p = 7. □

Remark 7.1. Solvability is used in two places: in Theorem 2.4, applied to the quotient G/FG/F, and in Lemma 4.2, where a minimal normal subgroup of the primitive quotient F/RF/R is elementary abelian, so that the degree [F:M][F : M] is a prime power. Lemma 4.1 holds for every finite group. The counting in Sections 5 and 6 uses only the conclusions of these results. Sun’s theorem also holds without solvability for cosets of subnormal subgroups [18 Theorem 3.1], but the images HiF/FH_iF/F need not be subnormal, and for nonsolvable groups primitive degrees need not be prime powers, so Lemma 4.2 would also need replacement.

Remark 7.2. The only direct uses of minimality of GG are to exclude index 2 and to descend to G/FG/F. In the setting of the proof minimality gives more when F<GF < G. The parts meeting a non-universal FF-coset yFyF number at least two, and after translation by an element of yFyF their pieces y′(Hi∩F)y'(H_i \cap F) form a partition of FF into left cosets. As FF is a smaller solvable group, it is HS, so there exist two distinct parts meeting yFyF with equal local index. We do not use this, because the proof cannot assume F<GF < G, including when p=7p = 7. For example, in C23⋊C7C_2^3 \rtimes C_7 the subgroup generated by the 7-elements is the whole group when C7C_7 acts nontrivially and irreducibly on C23C_2^3. This semidirect product is unique up to isomorphism, since GL⁡(3,2)\operatorname{GL}(3,2) has a unique conjugacy class of cyclic subgroups of order 7. This group is HS, so it is not a counterexample, but largest prime 7 and solvability do not by themselves force F<GF < G.

Formal proof

The accompanying Lean development [14] proves Theorem 1.2 as

Erdos274.herzog_schonheim.variants.solvable.

The proof revision is d701decf37c2. It is pinned to Lean v4.35.0-rc2 and the mathlib tag v4.35.0-rc2, a release candidate, by the files lean-toolchain and lakefile.toml. The repository consists of exactly the import closure of the main theorem. From its root, build it with

lake exe cache get
lake build

In a Lean file importing the module Erdos274.KnownCases.Solvable, the command

#print axioms Erdos274.herzog_schonheim.variants.solvable

reports precisely

[propExt, Classical.choice, Quot.sound]

Thus the elaborated theorem has no further axiomatic dependencies, including sorryAx. This check concerns the formal theorem; the interpretation of its hypotheses and its correspondence with the paper remain mathematical questions. In mathematical notation, the formal theorem reads as follows. Let GG be a solvable group with more than one element, and let bb be a finite type with at least two elements. Let (Hi)i∈ι(H_i)_{i \in\iota} be subgroups and (gi)i∈ι(g_i)_{i \in\iota} elements of GG such that the left cosets giHig_iH_i are nonempty, pairwise disjoint, and have union GG. Then [G:Hi]=[G:Hj][G : H_i] = [G : H_j] for some i≠ji \ne j. No finiteness hypothesis is placed on GG, and the cosets are left cosets, as in (1). The index is mathlib’s Subgroup.index, which is 0 for a subgroup of infinite index; in a finite partition every index is finite (Lemma 2.2). The hypothesis ∣G∣>1|G| > 1 follows from the others and is kept to match the upstream statement. The exact covering structure is copied field for field from the Erdős problem 274 file of the Formal Conjectures repository [7, 8], and the file Challenge.lean of the repository states the theorem using only mathlib and that structure, so the formal statement can be audited without reading the proof.

The formalization follows Sections 3–6. Its treatment of the inputs from Section 2 differs in form only. Lean proves Sun’s bound itself, in a more general form for orbits, and states it as ∑d∈Dφ(d)≤∣⋃jxjLj∣\sum_{d\in D}\varphi(d) \le\left|\bigcup_j x_jL_j\right|, where DD is the set of divisors of the orders ∣Lj∣|L_j|. Counting elements of a cyclic group of order ∣L∣|L| by their orders shows that the left side equals the left side of Theorem 2.4. Lemma 2.5 is proved there by decomposing UU into fibres over the divisors of a common multiple, rather than by inclusion–exclusion. Likewise, to see that every part of a partition of a possibly infinite group has finite index, Lean uses mathlib’s version of B. H. Neumann’s lemma [15 Lemma 4.1], in place of [12 Theorem 1].

Acknowledgments

The author thanks Leo Margolis for a careful reading of an earlier version, many helpful comments, and his encouragement.

Under the author’s direction, OpenAI GPT-5.6 Sol contributed the comparison of reciprocal sums with unions of quotient cosets (Section 3), the cyclotomic divisibility argument, and the argument for the prime 7, including the overlap set DD and Lemma 6.2. Anthropic Claude Opus 5 developed and formalized the early solvable track and the enlargement to all non-universal parts. Claude Opus 5.5 audited and formalized the final argument and completed the theorem for arbitrary solvable groups; its simplifications include localizing at the subgroup generated by the pp-elements, the fixed-point proofs of the local divisibility restrictions, including the maximal-subgroup proof of Lemma 4.2, the bound ord⁡p(q)≥3\operatorname{ord}_p(q) \ge3, and the elementary estimates in Section 5. OpenAI GPT-5.6 Sol, reviewing the paper, proposed the present form of the large-prime estimate in the proof of Lemma 5.2. Claude Fable 5 contributed earlier known-case formalizations and conditional reductions. Reviews by DeepSeek (version not recorded) and Z.ai GLM 5.2 clarified the earlier cyclotomic argument. OpenAI GPT-6 in Codex agreed with the exposition and attribution audit. The commit history of the repository [14] records the AI contributions to the formalization through its Co-Authored-By trailers.

The author thanks the Lean and mathlib communities for the libraries supporting the formal proof. The author is responsible for the mathematical claims, their attribution, and the correspondence between the paper and the formal statements.

References

References

  1. [1]F. Akman and P. Sissokho, Transversal coset partitions of groups, Beitr. Algebra Geom. 66 (2025), 417–441, doi:10.1007/s13366-024-00748-9.DOI
  2. [2]F. Akman and P. Sissokho, Steiner coset partitions of groups, Canad. Math. Bull. 68 (2025), 1359–1373, doi:10.4153/S0008439525100787.DOI
  3. [3]F. Akman and P. Sissokho, Steiner coset partitions for five mutually commuting subgroups, Acta Math. Hungar. (2026), published online, doi:10.1007/s10474-026-01635-6.DOI
  4. [4]M. A. Berger, A. Felzenbaum, and A. S. Fraenkel, The Herzog–Schönheim conjecture for finite nilpotent groups, Canad. Math. Bull. 29 (1986), no. 3, 329–333.DOI
  5. [5]M. A. Berger, A. Felzenbaum, and A. S. Fraenkel, Remark on the multiplicity of a partition of a group into cosets, Fund. Math. 128 (1987), 139–144.
  6. [6]M. C. Burkhart, The Herzog–Schönheim conjecture for solvable groups, arXiv:1901.10131 (2019), withdrawn in version 2, March 1, 2019.arxiv.org/abs/1901.10131
  7. [7]The Formal Conjectures Authors, The Formal Conjectures Repository, 2025, https://github.com/google-deepmind/formal-conjectures.
  8. [8]M. Firsching, P. Lezeau, S. Mercuri, M. Z. Horváth, Y. Dillies, C. Sönne, E. Wieser, F. Zhang, T. Hubert, B. Agüera y Arcas, and P. Kohli, Formal Conjectures: An open and evolving benchmark for verified discovery in mathematics, 2026, arXiv:2605.13171v1.arxiv.org/abs/2605.13171
  9. [9]M. Garonzi and L. Margolis, The Herzog–Schönheim conjecture for simple and symmetric groups, Algebra & Number Theory, to appear; arXiv:2509.25118v2.arxiv.org/abs/2509.25118
  10. [10]Y. Ginosar and O. Schnabel, Prime factorization conditions providing multiplicities in coset partitions of groups, J. Comb. Number Theory 3 (2011), no. 2, 75–86.
  11. [11]M. Herzog and J. Schönheim, Research problem no. 9, Canad. Math. Bull. 17 (1974), 150.
  12. [12]I. Korec and Š. Znám, On disjoint covering of groups by their cosets, Math. Slovaca 27 (1977), no. 1, 3–7.
  13. [13]L. Margolis and O. Schnabel, The Herzog–Schönheim conjecture for small groups and harmonic subgroups, Beitr. Algebra Geom. 60 (2019), no. 3, 399–418; arXiv:1803.03569.arxiv.org/abs/1803.03569
  14. [14]M. Menon, The Herzog–Schönheim conjecture for solvable groups in Lean 4, source repository, https://github.com/Jostamon/erdos274-hs-solvable, commit d701decf37c225a3e5326fade4bd92121b1f93d4 (October 2, 2026).
  15. [15]B. H. Neumann, Groups covered by permutable subsets, J. London Math. Soc. 29 (1954), 236–248, doi:10.1112/jlms/s1-29.2.236.DOI
  16. [16]D. J. S. Robinson, A Course in the Theory of Groups, second ed., Graduate Texts in Mathematics, vol. 80, Springer, New York, 1996.
  17. [17]Z.-W. Sun, Finite covers of groups by cosets or subgroups, Internat. J. Math. 17 (2006), no. 9, 1047–1064, arXiv:math/0501451.DOI
  18. [18]Z.-W. Sun, On the Herzog–Schönheim conjecture for uniform covers of groups, J. Algebra 273 (2004), 153–175, arXiv:math/0306099.DOI

Paper details

vlnc:W_6drpvr78
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Mathematics

GPT-5.6 Sol (bare), Claude Opus 5 (bare), Claude Opus 5.5, Claude Fable 5 (bare), DeepSeek (multiple versions), Z.ai GLM 5.2

View statement
Looking for related papers…
Contents