January 01, 1970
Given \(H\leq G\) finite abelian groups, a transversal \(T\subseteq G\) for \(G/H\) has fixed size \(|G/H|\), but its ambient difference support \(D(T)=T-T\) can vary with the embedding of \(H\) in \(G\). We call \(\delta(G,H)=\min_T |D(T)|\) the transversal difference number of the pair \((G,H)\). This invariant is related to finite abelian factorisation, tiling complements, and small-sumset questions, and is motivated by recent work regarding ambient Galois labels in CRT transforms for cyclotomic-subfield homomorphic encryption. We prove various results regarding this invariant, including a general lower bound \(\delta(G,H)\geq 2|G/H|-m(G,H),\) where \(m(G,H)\) is the largest order of a subgroup of \(G\) disjoint from \(H\). The bound is sharp for cyclic quotients, and Kneser’s theorem gives a cross-transversal estimate leading to exact product families with one nonsplit cyclic coordinate and arbitrary split factors. These results isolate the first genuinely new residual obstruction, namely the same-prime square plane \[G=(\mathbb{Z}/p^2\mathbb{Z})^2,\qquad H=pG.\] For odd \(p\), this case is the technical core of the paper. Here transversals are graphs of functions \(\mathbb{F}_p^2\to \mathbb{F}_p^2\), and \(D(T)\) decomposes into carry-corrected finite-field derivative images. We conjecture that \[\delta(G,H)=(2p-1)^2\] for all odd primes \(p\), prove the unconditional lower bound \(3p^2-p-1\), and give small-prime, probabilistic, and fixed-polynomial evidence for the conjecture.
A quotient group \(G/H\) remembers the cosets of \(H\), but it forgets how a choice of coset representatives sits inside the ambient group. This ambient information can vary substantially from one section to another. Throughout, all groups are finite abelian and written additively. Given \(H\leq G\), a transversal \(T\subseteq G\) for \(G/H\) has cardinality \(|G/H|\). However, its difference support \(D(T)= T-T\) can varry substantially in size and a natural question is how small this set \(D(T)\) can be. We call \[\delta(G,H) = \min_T |D(T)|\] the transversal difference number of the pair \((G,H)\), where \(T\) ranges over all transversals for \(G/H\).
This invariant depends on the embedding \(H\leq G\), not only on the abstract quotient. For instance, \(\delta(C_2,0)=2\) and \(\delta(C_4,2C_4)=3\), although both quotients are isomorphic to \(C_2\). When trying to find \(\delta(G,H)\), transversals can easily be found, but proving that their pairwise differences are as small as possible in the ambient group can be challenging.
After translating \(T\) we may assume \(0\in T\); then \(T\) is a set-theoretic tiling complement to \(H\), so every element of \(G\) is written uniquely as \(h+t\) with \(h\in H\) and \(t\in T\). Thus \(\delta(G,H)\) is a support minimisation problem for finite abelian factorisations \(G=H\oplus T\), where \(\oplus\) denotes unique representation rather than an internal direct product. This places it near the classical theory of unique-representation factorisations and algebraic tilings [1]–[5]. The lower-bound methods are closest to small-sumset, critical-pair, and Kneser-type stabiliser arguments [6]–[11]. It is not a difference-set or a relative difference-set problem, where multiplicity conditions rather than support minimisation are central [12]–[14], nor a difference-basis problem, where the set itself may vary in size while its differences cover the ambient group [15]. Here \(T\) has fixed size and must be a transversal; only the number of distinct ambient differences is minimised.
Our motivation comes from the Galois-theoretic bookkeeping that appears in Peikert and Pepin’s Vive Galois! programme [16]. In exact fully homomorphic encryption one often wants many plaintext “slots” over a prescribed finite field or finite ring, so that one ciphertext carries a vector of plaintexts and homomorphic operations act componentwise on that vector. This is the SIMD idea of Smart and Vercauteren [17]. Cyclotomic rings provide fast arithmetic, many automorphisms, and good geometric properties, but their slot structure is arithmetically rigid: the residue degree of the chosen plaintext prime determines the slot type, and an unwanted extension degree gives fewer slots of a larger field than the application naturally wants. The point of the Peikert–Pepin construction is that cyclotomic subfields, and more generally abelian number rings, give a more flexible Galois-theoretic way to obtain the desired slot type and number of slots, while still supporting CRT bases, structured transforms, and packed bootstrapping.
A finite choice of representatives enters this construction quite explicitly. In a tower of abelian extensions \(M/L/K\), write \[G_{M/K}=\operatorname{Gal}(M/K), \qquad G_{M/L}=\operatorname{Gal}(M/L).\] Peikert and Pepin choose a transversal \(T\subseteq G_{M/K}\) for the quotient \(G_{M/K}/G_{M/L}\) in order to index a CRT basis and the corresponding top-down sparse CRT transform [16]*Def. 5.8. Their Lemma 5.13 [16]*Lemma 5.13 shows that, in the relevant automorphism expansion, a coefficient can be nonzero only for an ambient automorphism of the form \[\tau=t't^{-1} \qquad (t,t'\in T).\] Thus the possible ambient automorphism labels are contained in \(TT^{-1}\). In the additive notation of this paper, the same support is \(T-T\). Consequently, the purely finite-abelian problem of minimising \(|T-T|\) over quotient transversals is exactly the problem of minimising the representative-dependent ambient label support forced by this part of the Galois/CRT bookkeeping.
This is the precise sense in which a small value of \(\delta(G,H)\) can be useful for the FHE framework. In Peikert–Pepin’s homomorphic CRT transforms [16]*Secs. 3.2 and 4, structured linear maps are evaluated by applying automorphisms and taking linear combinations. Each distinct ambient automorphism label that is actually used is a potential automorphism operation, and in an FHE implementation such an operation typically comes with key-switching data and key-switching cost. The first and third phases of packed bootstrapping are precisely CRT transforms: they move noisy decryption coefficients into the SIMD slots, allow the slotwise nonlinear decoding step, and then move the result back. Hence, for an implementation model in which each distinct ambient automorphism label carries cost, a transversal with smaller \(TT^{-1}\) may reduce the candidate automorphism/key-switch label set and simplify the linear part of packed bootstrapping.
A word of caution about this motivation is worth mentioning. The quantity \(\delta(G,H)\) measures just the size of the support of ambient automorphism labels forced by a choice of representatives, and therefore it should not be mistaken for a statement about runtime, memory, key-switching cost, parameter selection, or security. Several effects sit between this invariant and any real implementation. The label set \(TT^{-1}\) records which automorphisms can occur, but some of the corresponding coefficients may vanish for reasons the transversal cannot see; the cost of applying an automorphism depends on the particular FHE platform; and the Ring-LWE hardness assumptions used in this line of work are unaffected by how small \(|T-T|\) is. Once these caveats are set aside, what remains is a clean, finite, and entirely algebraic question, and that is the one we pursue: how small can the difference support of a section of a finite abelian quotient be, what is its value across broad families, and where does the naive split-or-cyclic intuition fail?
Main results. The first part of the paper develops the theory in the finite abelian setting. Here the trivial lower bound \(\delta(G,H)\ge |G/H|\) is attained exactly when \(H\) has a subgroup complement. To go further, we isolate the singleton quotient directions of a transversal: their canonical lifts form a subgroup of \(G\) disjoint from \(H\), and quotienting by it makes the transversal primitive, leaving no nonzero singleton fibres. This reduction yields the uniform lower bound \(\delta(G,H)\ge 2|G/H|-m(G,H)\), where \(m(G,H)\) is the largest order of a subgroup of \(G\) disjoint from \(H\). For cyclic quotients the bound is sharp, and feeding Kneser’s theorem (Theorem 1) into a cross-transversal estimate pins down exact product families with one nonsplit cyclic coordinate and arbitrary split factors. This cyclic sharpness sits within the prime-cyclic critical-pair tradition initiated by Vosper [18], and more broadly within Kemperman’s analysis of small sumsets in abelian groups [7]. The precise statements are given in Sections 3 and 4, and there results therein should be viewed as both a general theory and as delimitation results: they identify the regimes in the ideas described above already determine the invariant. The main new difficulty begins when these mechanisms no longer control the support, namely in same-prime residual quotients with more than one nonsplit direction.
Section 5 marks this transition. A residual quotient that contains a \(C_2\times C_2\) plane already forces a strict gap from the trivial lower-bound.
For odd primes, however, the parity argument disappears, and the first case left open by all the preceding exact results is \[G_p=(\mathbb{Z}/p^2\mathbb{Z})^2,\qquad H_p=pG_p.\] This odd square-plane problem, treated from Section 6 onward, is the technical core of the paper. Here transversals become graphs of functions \(\mathbb{F}_p^2\to\mathbb{F}_p^2\), and the fibres of \(D(T)\) turn into carry-corrected finite-field derivative images. Read in these terms, the square-plane problem sits close to finite-field sum-product [19] and image-expansion [20] questions, while the fixed-polynomial evidence gathered in Section 7 lies nearest to value-set questions for polynomial maps [21]. We conjecture that \(\delta(G,H)=(2p-1)^2\) and prove the unconditional bound \(\delta(G,H) \geq 3p^2-p-1\). We also prove that random liftings meet the lower bound \((2p-1)^2\) with probability tending to one, and that no fixed integer-polynomial lifting can produce counterexamples once the prime is large enough. The final section gathers the primitive-quotient and square-modulus open problems that emerge along the way.
Throughout the paper all groups are finite abelian and are written additively. For subsets \(A,B\subseteq G\) we write \(A+B=\{a+b:a\in A,\;b\in B\}\), \(A-B=\{a-b:a\in A,\;b\in B\}\), and \(-A=\{-a:a\in A\}\). If \(K\leq G\) is a subgroup, then \(A+K\) denotes the union of the \(K\)-cosets meeting \(A\).
Let \(H\leq G\). The quotient map will be denoted \[\pi=\pi_{G,H}:G\longrightarrow G/H.\] When no confusion is possible we put \[Q=G/H,\qquad q=|Q|=|G/H|.\] A transversal for \(G/H\) is a subset \(T\subseteq G\) such that \(\pi|_T:T\to G/H\) is bijective. Equivalently, \(T\) contains exactly one element from each coset of \(H\) in \(G\).
For a transversal \(T\) we define its difference support \[D(T)=T-T.\] The invariant studied in this paper is the transversal difference number of the pair \((G,H)\): \[\label{eq:deltainv} \delta(G,H)=\min_T |D(T)|,\tag{1}\] where \(T\) ranges over all transversals for \(G/H\). Translating a transversal by an element of \(G\) does not change its difference support. Hence, whenever convenient, we shall assume that \(T\) is normalized, meaning that \(0\in T\).
A normalized transversal is the same thing as a set-theoretic complement to \(H\): every element \(g\in G\) has a unique expression \(g=h+t\), with \(h\in H\), and \(t\in T\). This is a unique-representation factorisation of the set \(G\); it does not mean that \(T\) is a subgroup unless this is explicitly stated. We use the terminology of set factorisations and complements in the standard sense of [5]; see also [1]–[3] for the classical factorisation and tiling background.
It will be useful to identify a transversal with its section of the quotient map. Thus, for a normalized transversal \(T\), we write \[\label{eq:sec} s=s_T:Q\longrightarrow G\tag{2}\] for the unique section such that \(s(0)=0\), \(\pi\circ s=\operatorname{id}_Q\), and \(T=s(Q)\). For a quotient direction \(a\in Q\) we define the corresponding difference fibre \[\label{eq:diff-fibre} \mathcal{D}_a(T) =D(T)\cap \pi^{-1}(a) =\{s(x+a)-s(x):x\in Q\}.\tag{3}\] The fibres \(\mathcal{D}_a(T)\) are nonempty and decompose the difference support into the following disjoint union \(D(T)=\bigsqcup_{a\in Q}\mathcal{D}_a(T).\) Note that in particular \(\mathcal{D}_0(T)=\{0\}\). We call \(a\in Q\) a singleton direction for \(T\) if \(|\mathcal{D}_a(T)|=1\). The set of singleton directions will be denoted \[\label{eq:singleton-dir} \Sigma(T)=\{a\in Q:|\mathcal{D}_a(T)|=1\}.\tag{4}\] If \(a\in \Sigma(T)\), the unique element of \(\mathcal{D}_a(T)\) will sometimes be written \(d_T(a)\); equivalently, \(s(x+a)-s(x)=d_T(a)\) for all \(x\in Q\). Therefore, \(\pi(d_T(a))=a\).
We write \[\widetilde{\Sigma}(T)=\{d_T(a):a\in\Sigma(T)\}\subseteq G\] for the corresponding lifted singleton set. Thus \(\Sigma(T)\) is a subset of the quotient \(G/H\), while \(\widetilde{\Sigma}(T)\) is a subset of the ambient group \(G\).
Definition 1. A normalized transversal \(T \subseteq G\) for \(G/H\) is called primitive if \(\Sigma(T)=\{0\}\). Equivalently, no nonzero quotient direction has a singleton difference fibre.
The following subgroup invariant will also appear in our work: \[\label{eq:defmGH} m(G,H)=\max\{|K|:K\leq G,\;K\cap H=\{0\}\}.\tag{5}\] Note that \(1\leq m(G,H)\leq |G/H|\) and the case in which the pair \(G\), \(H\) satisfies \(m(G,H)=1\) will be called residual. A subgroup \(K\leq G\) is a subgroup complement to \(H\) if \(K\cap H=\{0\}\) and \(G=H+K\). In particular, \(H\) has a subgroup complement in \(G\) if and only if \(m(G,H)=|G/H|\).
For later use we recall Kneser’s theorem (see [6] or [9]*Theorem 5.5) in the finite form needed below. If \(A\subseteq G\), its stabilizer is \[\operatorname{Stab}(A)=\{g\in G:A+g=A\}.\]
Theorem 1 (Kneser). Let \(A,B\) be nonempty subsets of a finite abelian group \(G\), and write \(P=\operatorname{Stab}(A+B)\). Then \[|A+B|\geq |A+P|+|B+P|-|P|.\]
Since \(A-B=A+(-B)\), the same estimate applies to difference sets \(A-B\) with \(P=\operatorname{Stab}(A-B)\). The reader interested in broader background on Kneser-type inequalities, critical pairs, and small-doubling methods in abelian groups, could consult [7], [9] or [10].
The first estimates come from decomposing the difference support over quotient directions. We notice that singleton fibres lift to a subgroup of \(G\) disjoint from \(H\) and quotienting by this subgroup makes the transversal primitive, gives the uniform lower bound in terms of \(m(G,H)\), and leads to an exact formula for cyclic quotients.
The first proposition below is the support-minimisation analogue of the basic exact-factorisation question: when does a set-theoretic complement become a subgroup complement? Classical factorisation results of Hajós [1] and Rédei [2] are related to this phenomenon.
Proposition 2. Let \(H\leq G\). Then \(\delta(G,H)=|G/H|\) if and only if \(H\) has a subgroup complement in \(G\).
Proof. Let \(q=|G/H|\). For every transversal \(T\) and every fixed \(t_0\in T\), the translate \(T-t_0\) has \(q\) elements and is contained in \(D(T)=T-T\). Hence \(|D(T)|\ge q\) for every \(T\), so \(\delta(G,H)\ge q\).
If \(K\leq G\) is a subgroup complement to \(H\), then \(K\) is a transversal for \(G/H\) and \(D(K)=K-K=K\). Thus \(\delta(G,H)\le |K|=q\), and equality follows.
Conversely, suppose \(\delta(G,H)=q\), and choose a transversal \(T\) with \(|D(T)|=q\). Translating \(T\) if necessary, assume \(0\in T\). Then \(T\subseteq D(T)\) because \(t=t-0\) for every \(t\in T\). Since the two sets have the same cardinality, \(T=D(T)\). Therefore \(T\) is closed under subtraction. As \(T\) is finite and contains \(0\), it is a subgroup of \(G\). Since \(T\) is also a transversal for \(G/H\), it is a subgroup complement to \(H\). ◻
We use the notation \(\Sigma(T)\), \(d_T(a)\), and \(\widetilde{\Sigma}(T)\) from Section 2. The next proposition explains the quotient reduction behind primitive transversals.
Proposition 3. Let \(H\leq G\), and let \(T\) be a normalized transversal for \(G/H\). Then:
\(\Sigma(T)\) is a subgroup of \(G/H\);
the map \[\Sigma(T)\longrightarrow G, \qquad a\longmapsto d_T(a),\] is an injective group homomorphism, its image \(\widetilde{\Sigma}(T)\) is a subgroup of \(G\), and \[\widetilde{\Sigma}(T)\cap H=\{0\}, \qquad |\widetilde{\Sigma}(T)|=|\Sigma(T)|;\]
one has \[T+\widetilde{\Sigma}(T)=T, \qquad D(T)+\widetilde{\Sigma}(T)=D(T),\] so \(D(T)\) is a union of cosets modulo \(\widetilde{\Sigma}(T)\);
if \[\overline{G}=G/\widetilde{\Sigma}(T), \qquad \overline{H}=(H+\widetilde{\Sigma}(T))/\widetilde{\Sigma}(T),\] and \(\overline{T}\) denotes the image of \(T\) in \(\overline{G}\), then \(\overline{T}\) is a transversal for \(\overline{G}/\overline{H}\) and \[|D(T)|=|\widetilde{\Sigma}(T)|\,|D(\overline{T})|;\]
the transversal \(\overline{T}\) has no nonzero singleton directions.
Proof. Put \(Q=G/H\) and write \(s=s_T\) for the normalized section attached to \(T\). The zero direction is singleton, since \(\mathcal{D}_0(T)=\{0\}\). Let \(a,b\in\Sigma(T)\). Then, for every \(x\in Q\), \[s(x+a)-s(x)=d_T(a), \qquad s(x+b)-s(x)=d_T(b).\] Hence \[\begin{align} s(x+a+b)-s(x) &=\bigl(s(x+a+b)-s(x+a)\bigr)+\bigl(s(x+a)-s(x)\bigr) \\ &=d_T(b)+d_T(a). \end{align}\] Thus \(a+b\) is a singleton direction and \[d_T(a+b)=d_T(a)+d_T(b).\] Since \(Q\) is finite, closure under addition and the presence of \(0\) imply that \(\Sigma(T)\) is a subgroup. The displayed identity also shows that \(a\mapsto d_T(a)\) is a group homomorphism. This proves (1) and the homomorphism assertion in (2).
For \(a\in\Sigma(T)\) we have \[\pi(d_T(a))=a,\] because \(d_T(a)=s(x+a)-s(x)\) for every \(x\in Q\). Hence the homomorphism \(a\mapsto d_T(a)\) is injective, and its image \(\widetilde{\Sigma}(T)\) is a subgroup of \(G\) with \(|\widetilde{\Sigma}(T)|=|\Sigma(T)|\). If \(d_T(a)\in H\), then \(a=\pi(d_T(a))=0\), hence \(d_T(a)=0\). Therefore \[\widetilde{\Sigma}(T)\cap H=\{0\}.\] This proves (2).
If \(\sigma=d_T(a)\in\widetilde{\Sigma}(T)\), then \[s(x+a)=s(x)+\sigma\] for every \(x\in Q\). As \(x\) varies over \(Q\), so does \(x+a\), and therefore translation by \(\sigma\) permutes the elements of \(T\). Thus \[T+\widetilde{\Sigma}(T)=T.\] Subtracting gives \[D(T)+\widetilde{\Sigma}(T)=D(T),\] so \(D(T)\) is a union of cosets modulo \(\widetilde{\Sigma}(T)\).
Let \(q_{\natural}:G\to\overline{G}\) be the quotient map. We prove (4). The image \(\overline{T}=q_{\natural}(T)\) plainly meets every \(\overline{H}\)-coset. For uniqueness, suppose that \(q_{\natural}(t_1)\) and \(q_{\natural}(t_2)\) lie in the same \(\overline{H}\)-coset. Then \[t_1-t_2\in H+\widetilde{\Sigma}(T).\] Write \(t_1-t_2=h+\sigma\), with \(h\in H\) and \(\sigma\in\widetilde{\Sigma}(T)\). Since \(T+\widetilde{\Sigma}(T)=T\), the element \(t_2+\sigma\) belongs to \(T\). But \[t_1-(t_2+\sigma)=h\in H.\] Since \(T\) is a transversal for \(G/H\), this forces \(t_1=t_2+\sigma\), hence \(q_{\natural}(t_1)=q_{\natural}(t_2)\). Thus \(\overline{T}\) is a transversal.
Moreover, \[q_{\natural}(D(T))=q_{\natural}(T-T)=\overline{T}-\overline{T}=D(\overline{T}).\] Since \(D(T)\) is a union of \(\widetilde{\Sigma}(T)\)-cosets, each element of \(D(\overline{T})\) has exactly \(|\widetilde{\Sigma}(T)|\) preimages in \(D(T)\). Hence \[|D(T)|=|\widetilde{\Sigma}(T)|\,|D(\overline{T})|.\]
It remains to show that \(\overline{T}\) is primitive. Under the natural identification \[\overline{G}/\overline{H} \simeq G/(H+\widetilde{\Sigma}(T)) \simeq (G/H)/\Sigma(T),\] suppose that a nonzero class \(\overline{a}\) has singleton fibre for \(\overline{T}\), and choose a representative \(a\in Q\). Then, for some \(d\in G\), \[\mathcal{D}_a(T)\subseteq d+\widetilde{\Sigma}(T).\] Indeed, the quotient fibre in direction \(\overline{a}\) is the image of the fibres \(\mathcal{D}_{a+k}(T)\), \(k\in\Sigma(T)\); if that image is a singleton, then each such fibre lies in one \(\widetilde{\Sigma}(T)\)-coset. For every \(k\in\Sigma(T)\), the fibre in direction \(a+k\) satisfies \[\begin{align} \mathcal{D}_{a+k}(T) &=\{s(x+a+k)-s(x):x\in Q\} \\ &=\{s(x+a)+d_T(k)-s(x):x\in Q\} \\ &=\mathcal{D}_a(T)+d_T(k). \end{align}\] Thus the union of the fibres above \(a+\Sigma(T)\) lies in the single coset \(d+\widetilde{\Sigma}(T)\), whose size is \(|\widetilde{\Sigma}(T)|=|\Sigma(T)|\). The \(|\Sigma(T)|\) fibres in this union are nonempty and disjoint because they lie over distinct quotient directions. Hence each is a singleton. In particular \(a\in\Sigma(T)\), so \(\overline{a}=0\), a contradiction. Thus \(\overline{T}\) has no nonzero singleton directions. ◻
Every transversal becomes primitive after quotienting by its lifted singleton subgroup. This is the structural reason for the lower bound below.
Proposition 4. For every \(H\leq G\) one has \[\delta(G,H)\geq 2|G/H|-m(G,H).\] More precisely, for every normalized transversal \(T\) one has \[|D(T)|\geq 2|G/H|-|\Sigma(T)|.\]
Proof. Let \(T\) be a normalized transversal for \(G/H\). Put \(Q=G/H\), \(q=|Q|\), and \(k=|\Sigma(T)|\). By the fibre decomposition, \[|D(T)| =\sum_{a\in Q}|\mathcal{D}_a(T)| \geq k+2(q-k) =2q-k,\] because the \(k\) singleton fibres have size \(1\) and all other fibres are nonempty of size at least \(2\). This proves the more precise bound. By Proposition 3, the lifted singleton subgroup \(\widetilde{\Sigma}(T)\) is disjoint from \(H\), so \(k=|\widetilde{\Sigma}(T)|\leq m(G,H)\). Thus \[|D(T)|\geq 2q-m(G,H).\] Taking the minimum over all transversals proves the proposition. ◻
Remark 5. Looking at the proof above, if \(T\) is a normalized transversal, it is easy to see that the equality \[|D(T)|=2|G/H|-m(G,H)\] holds if and only if \[|\Sigma(T)|=m(G,H)\] and every non-singleton quotient-direction fibre \(\mathcal{D}_a(T)\) has size exactly \(2\).
The bound of Proposition 4 is sharp for all cyclic quotients.
Theorem 6. Let \(H\leq G\), and suppose that \(G/H\) is cyclic of order \(q\). Then \[\delta(G,H)=2q-m(G,H).\]
Proof. The lower bound is Proposition 4. We first handle the residual case \(m(G,H)=1\). Choose \(g\in G\) whose image generates \(G/H\). If \(q=1\), there is nothing to prove. For \(q>1\), the integer \(q\) divides \(\operatorname{ord}(g)\). The equality \(\operatorname{ord}(g)=q\) would make \(\langle g\rangle\) a nontrivial subgroup disjoint from \(H\), contradicting residuality. Hence \(\operatorname{ord}(g)\geq 2q.\)
Now observe that \(T=\{0,g,2g,\ldots,(q-1)g\}\) is a transversal for \(G/H\), and \(D(T)=\{kg:-(q-1)\leq k\leq q-1\}\). These \(2q-1\) elements are distinct because any two exponents in the displayed range differ by an integer of absolute value at most \(2q-2\), which is strictly smaller than \(\operatorname{ord}(g)\). Thus \(\delta(G,H)=2q-1\) in the residual cyclic case.
For the general case, let \(K\leq G\) be a subgroup disjoint from \(H\) with \(|K|=m(G,H)\), and write \(m=|K|\). Since \(G/H\) is cyclic, the image of \(K\) in \(G/H\) is the unique subgroup of order \(m\). Hence \[(G/K)/((H+K)/K)\simeq G/(H+K)\] is cyclic of order \(s=q/m\).
The pair \((G/K,\,(H+K)/K)\) is residual. Indeed, if \(L/K\leq G/K\) were nontrivial and disjoint from \((H+K)/K\), then \(L\cap(H+K)=K\). In particular, \(L\cap H=\{0\}\), because \(K\cap H=\{0\}\). But then \(L\) would be a subgroup of \(G\) disjoint from \(H\) and strictly larger than \(K\), contradicting the maximality of \(K\).
By the residual case just proved, there is a transversal \(\overline{B}\subseteq G/K\) for \[(G/K)/((H+K)/K)\] such that \[|\overline{B}-\overline{B}|=2s-1.\] Choose representatives \(B\subseteq G\) for \(\overline{B}\), and define \[T=B+K.\] Since \(B\) represents the quotient \(G/(H+K)\) and \(K\) maps isomorphically onto its image in \(G/H\), the set \(T\) is a transversal for \(G/H\). Moreover, \[D(T)=(B-B)+K.\] Modulo \(K\), this support is exactly \(\overline{B}-\overline{B}\), so it is a union of \(2s-1\) distinct \(K\)-cosets. Therefore \[|D(T)|=|K|(2s-1)=m\left(2\frac{q}{m}-1\right)=2q-m.\] Together with the lower bound, this proves the formula. ◻
For a cyclic ambient group with a prescribed subgroup, the formula becomes completely explicit. In what follows, we write \(C_n\) for the cyclic subgroup of order \(n\), for any \(n \geq 1\).
Corollary 1. Let \(h,q\geq 1\), let \(G=C_{hq}\), and let \(H\leq G\) be the subgroup of order \(h\). If \(c\) is the largest divisor of \(q\) that is coprime to \(h\), then \[\delta(G,H)=2q-c.\]
Proof. In a cyclic group there is a unique subgroup of each order. A subgroup of order \(k\) intersects \(H\) trivially if and only if \(\gcd(k,h)=1\). Since such a subgroup must inject into \(G/H\), its order must also divide \(q\). Therefore \(m(G,H)\) is exactly the largest divisor \(c\) of \(q\) that is coprime to \(h\). The claim follows from Theorem 6. ◻
Example 1 (Mixed-prime square-modulus products). Let \(p_1,\ldots,p_s\) be distinct primes, let \(n=\prod_i p_i\), and put \[G=\prod_{i=1}^s C_{p_i^2}, \qquad H=\prod_{i=1}^s p_iC_{p_i^2}.\] Then \[\delta(G,H)=2n-1.\] Indeed, the Chinese remainder theorem identifies \((G,H)\) with \((C_{n^2},\,nC_{n^2}).\) Here \(|H|=n\) and \(|G/H|=n\). In Corollary 1, with \(h=q=n\), the largest divisor of \(q\) coprime to \(h\) is \(1\). Hence \(\delta(G,H)=2n-1\). Thus square-modulus products with distinct prime factors behave cyclically. The first obstruction not covered by the cyclic analysis is therefore not a product over distinct primes, but a same-prime \(p\)-primary quotient of rank at least two. This is the source of the square-plane problem studied later in the manuscript.
Such CRT and product constructions are frequent in the algebraic-tiling and finite-abelian factorisation literature [3]–[5].
The aim of this section is to study the behaviour of the invariant \(\delta\) with respect to subgroup chains and direct products. The technique we use for this study involves slicing a transversal over an intermediate quotient. Kneser’s theorem then upgrades the elementary nonzero-fibre count and gives exact product families with one nonsplit cyclic coordinate and arbitrary split factors. The resulting estimates are cross-transversal bounds based on Kneser’s theorem, and should be read against the structural additive theory of small sumsets in finite abelian groups [7], [9], [10].
We start with some elementary chain bounds: a multiplicative construction and two lower bounds coming from the zero and nonzero quotient fibres.
Proposition 7. Let \(H\leq K\leq G\), and write \[d_0=\delta(K,H), \qquad d_1=\delta(G,K), \qquad r=|K/H|, \qquad q=|G/K|.\] Then \[\max\{d_0+d_1-1,\;d_0+(q-1)r\} \leq \delta(G,H)\leq d_0d_1.\]
Proof. Choose a transversal \(V\subseteq K\) for \(K/H\) with \(|D(V)|=d_0\), and choose a transversal \(U\subseteq G\) for \(G/K\) with \(|D(U)|=d_1\). Then \[T=U+V\] is a transversal for \(G/H\). Moreover \[D(T)=(U+V)-(U+V)\subseteq D(U)+D(V).\] Thus \[\delta(G,H)\leq |D(T)|\leq |D(U)|\,|D(V)|=d_1d_0.\]
For the first lower bound, let \(T\) be a transversal for \(G/H\), and write \(T_y=T\cap \pi_{G,K}^{-1}(y)\) for its slice over \(y\in G/K\). Choose one element of each nonempty slice; these elements form a transversal \(U_T\) for \(G/K\), so \(|D(U_T)|\ge d_1\). Each translated slice is a transversal for \(K/H\), so the zero \(G/K\)-fibre of \(D(T)\) contains at least \(d_0\) elements. Since \(D(U_T)\) meets that zero fibre only in \(0\), \[|D(T)|\geq d_0+d_1-1.\]
For the second lower bound, one translated slice gives at least \(d_0\) elements in the zero fibre. In every nonzero \(G/K\)-direction, the cross-difference between two corresponding slices maps onto all of \(K/H\): the two slices each meet every \(H\)-coset inside their \(K\)-coset. Hence that fibre has at least \(r=|K/H|\) elements, and \[|D(T)|\geq d_0+(q-1)r.\] Taking the minimum over all \(T\) proves the proposition. ◻
The proof above isolates the point where information is lost: in each nonzero \(G/K\)-direction it uses only that a cross-difference of two slices surjects onto \(K/H\). Kneser’s theorem gives a uniform replacement for this crude fibre count.
Theorem 8. Let \(H\leq G\), put \(q=|G/H|\), and let \(A,B\subseteq G\) be transversals for \(G/H\). Then \[|A-B|\geq 2|G/H|-m(G,H).\]
Proof. Let \(P=\operatorname{Stab}(A-B)\), and let \(\pi:G\to G/H\) be the quotient map. Denote by \(s=|P\cap H|\), and by \(u=|\pi(P)|\). The exact sequence \[0\longrightarrow P\cap H\longrightarrow P\longrightarrow \pi(P) \longrightarrow 0\] gives \(|P|=su\). By Kneser’s theorem, applied to \(A+(-B)\), \[|A-B|\geq |A+P|+|B+P|-|P|.\] Since \(A\) contains one element in each \(H\)-coset, the set \(A+P\) contains, in each \(H\)-coset, at least one coset of \(P\cap H\). Hence \[|A+P|\geq sq.\] The same argument gives \(|B+P|\geq sq\) and therefore \(|A-B|\geq s(2q-u)\).
If \(s=1\), then \(P\cap H=\{0\}\), so \(P\) is a subgroup of \(G\) disjoint from \(H\). Thus \[u=|P|\leq m(G,H),\] and hence \[|A-B|\geq 2q-u\geq 2q-m(G,H).\] If \(s>1\), then \(u\leq q\), and so \[|A-B|\geq s(2q-u)\geq sq\geq 2q.\] Since \(m(G,H)\geq 1\), this implies \[|A-B|\geq 2q\geq 2q-m(G,H)\] and the theorem follows. ◻
Applying the theorem to pairs of slices in the chain \(H\leq K\leq G\) replaces the elementary contribution \(|K/H|\) in every nonzero \(G/K\)-fibre by \(2|K/H|-m(K,H)\).
Corollary 2. Let \(H\leq K\leq G\), and put \[\begin{align} d_0&=\delta(K,H), & d_1&=\delta(G,K), r&=|K/H|, & q&=|G/K|, & m_0&=m(K,H). \end{align}\] Then \[\delta(G,H) \geq \max\{d_0+d_1-1,\;d_0+(q-1)(2r-m_0)\}.\]
Proof. The bound \(d_0+d_1-1\) is Proposition 7. Let \(T\) be a transversal for \(G/H\) and slice it above the cosets of \(K\). The zero \(G/K\)-fibre contains the difference set of one translated slice and contributes at least \(d_0\) elements.
For a nonzero direction in \(G/K\), the corresponding fibre contains a translated cross-difference between two slices. After translation into \(K\), these are transversals for \(K/H\), so Theorem 8 gives at least \[2r-m_0\] elements. Summing over the \(q-1\) nonzero directions in \(G/K\) gives \[|D(T)|\geq d_0+(q-1)(2r-m_0).\] Taking the minimum over \(T\) proves the claim. ◻
Remark 9. Passing to quotients contained in \(H\) can only identify differences. If \(N\leq H\leq G\), then \[\delta(G/N,H/N)\leq \delta(G,H).\] Indeed, the image of a transversal for \(G/H\) under the quotient map \(G\to G/N\) is a transversal for \((G/N)/(H/N)\), and quotienting can only identify differences.
Proposition 3 gives a more precise form of quotienting for the particular subgroup \(\widetilde{\Sigma}(T)\) attached to a transversal \(T\): in that case the image transversal is primitive and the difference support satisfies the exact multiplicative identity \[|D(T)|=|\widetilde{\Sigma}(T)|\,|D(\overline{T})|.\]
Applying the chain estimates to the two natural coordinate chains gives the following elementary product bounds. Let \(H_i\leq G_i\), put \[q_i=|G_i/H_i|, \qquad d_i=\delta(G_i,H_i) \qquad (i=1,2).\] Then \[\label{eq:bound-dir-products} \max\{d_1+d_2-1,\;d_1+(q_2-1)q_1,d_2+(q_1-1)q_2\}\leq \delta(G_1\times G_2,H_1\times H_2)\leq d_1d_2.\tag{6}\] The product of optimal transversals gives the upper bound: if \(T_i\subseteq G_i\) is a transversal for \(G_i/H_i\) with \(|D(T_i)|=d_i\), then \(T_1\times T_2\) is a transversal for \[(G_1\times G_2)/(H_1\times H_2),\] and \[D(T_1\times T_2)=D(T_1)\times D(T_2).\] Thus the difference support has size \(d_1d_2\).
For the lower bounds, apply Proposition 7 to the chain \[H_1\times H_2\leq G_1\times H_2\leq G_1\times G_2\] and then to the symmetric chain \[H_1\times H_2\leq H_1\times G_2\leq G_1\times G_2.\] The identities \[\delta(G_1\times H_2,H_1\times H_2)=\delta(G_1,H_1), \qquad \delta(G_1\times G_2,G_1\times H_2)=\delta(G_2,H_2)\] follow from projection and the fixed-coordinate construction. Substituting them in the two chains gives the displayed estimates.
The same sharpening applies to products through the two coordinate chains. With \(m_i=m(G_i,H_i)\), Corollary 2 gives \[\begin{align} \delta(G_1\times G_2,H_1\times H_2) \geq \max\{&d_1+d_2-1,\;d_1+(q_2-1)(2q_1-m_1),\\ &d_2+(q_1-1)(2q_2-m_2)\}. \end{align}\]
Remark 10. Since \(m_i\leq q_i\), the Kneser-sharpened product terms dominate the corresponding elementary product terms above. The improvement is strict in the term where the factor providing \(m_i\) is nonsplit, meaning \(m_i<q_i\), and the other quotient is nontrivial.
The Cartesian-product construction need not be optimal.
Example 2 (Products need not be multiplicative). The cyclic formula gives \[\delta(C_4,2C_4)=3, \qquad \delta(C_9,3C_9)=5.\] However \[(C_4\times C_9,\;2C_4\times 3C_9) \simeq (C_{36},\;6C_{36}).\] By Corollary 1, \[\delta(C_{36},6C_{36})=11.\] Thus the optimal value for the product pair is \(11\), not \(15\).
The cross-transversal estimate is exact after adjoining a split factor, provided the base pair already attains the lower bound.
Corollary 3. Let \(H_0\leq G_0\), let \(L\) be a finite abelian group, for which we write \(q_0=|G_0/H_0|\), and \(m_0=m(G_0,H_0)\). Then \[\delta(G_0\times L,H_0\times\{0\}) \geq |L|(2q_0-m_0).\] If \(\delta(G_0,H_0)=2q_0-m_0\), the inequality above holds with equality.
Proof. Let \(T\) be a transversal for \[(G_0\times L)/(H_0\times\{0\}).\] For each \(\ell\in L\), define \[A_\ell=\{g\in G_0:(g,\ell)\in T\}.\] Because quotient cosets are indexed by \((G_0/H_0)\times L\), each \(A_\ell\) is a transversal for \(G_0/H_0\).
For each \(\ell\in L\), the \(L\)-coordinate \(\ell\) fibre of \(D(T)\) contains \[(A_\ell-A_0)\times\{\ell\}.\] By Theorem 8, \[|A_\ell-A_0|\geq 2q_0-m_0.\] These fibres are disjoint for different \(\ell\), so \[|D(T)|\geq |L|(2q_0-m_0).\] Taking the minimum over \(T\) proves the lower bound.
For the upper bound under the displayed hypothesis, choose a transversal \(T_0\subseteq G_0\) for \(G_0/H_0\) with \[|D(T_0)|=2q_0-m_0.\] Then \(T_0\times L\) is a transversal for \((G_0\times L)/(H_0\times\{0\})\), and \[D(T_0\times L)=D(T_0)\times L.\] Thus \[|D(T_0\times L)|=|L|\,|D(T_0)|=|L|(2q_0-m_0),\] which proves equality. ◻
Together with the cyclic quotient formula, this gives the main exact product family of the section.
Corollary 4. Let \(p\) be prime, let \(1\leq b<a\), and let \(L\) be a finite abelian group. Then \[\delta(C_{p^a}\times L,\;p^b C_{p^a}\times\{0\}) = |L|(2p^b-1).\]
Proof. For the base pair \[G_0=C_{p^a}, \qquad H_0=p^bC_{p^a},\] the quotient has size \(p^b\). Since \(1\leq b<a\), the subgroup \(H_0\) is nonzero, and every nontrivial subgroup of the cyclic \(p\)-group \(C_{p^a}\) meets \(H_0\) nontrivially. Therefore \[m(G_0,H_0)=1.\] The cyclic quotient formula, Theorem 6, gives \[\delta(G_0,H_0)=2p^b-1=2|G_0/H_0|-m(G_0,H_0).\] Corollary 3 now gives the desired equality after adjoining the split factor \(L\). ◻
Remark 11. Corollary 4 is not a product multiplicativity statement: the added factor is split, with subgroup \(p^bC_{p^a}\times\{0\}\). It does not cover a second nonsplit \(p\)-primary coordinate, which is why \(((\mathbb{Z}/p^2\mathbb{Z})^2,\;p(\mathbb{Z}/p^2\mathbb{Z})^2)\) is the first genuinely new same-prime case.
The cyclic quotient formula shows that the fibre lower bound is sharp for every cyclic quotient. Noncyclic residual quotients behave differently: the first obstruction is already visible in a quotient plane of order four, and it separates the cyclic theory from the higher-rank phenomena studied later.
The two-torsion obstruction gives a strict improvement whenever the quotient contains a two-dimensional \(2\)-torsion plane.
Theorem 12. Let \(H\leq G\) be residual, so that \(m(G,H)=1\). Suppose that \(G/H\) contains a subgroup isomorphic to \(C_2\times C_2\). Then every transversal \(T\) for \(G/H\) satisfies \[|D(T)|\geq 2|G/H|+1.\] In particular, the residual fibre lower bound \(2|G/H|-1\) is not sharp for such a pair.
Proof. Let \(Q=G/H\), let \(q=|Q|\), and normalize \(T\), with associated section \(s:Q\to G\). Since the pair is residual, Proposition 3 implies that no nonzero quotient direction can be singleton. Therefore \[|\mathcal{D}_a(T)|\geq 2 \qquad(a\in Q,\;a\neq 0).\]
For directions of order two there is a stronger parity fact. If \(a\in Q\) has order two, then \(a=-a\), hence \[\mathcal{D}_a(T)=-\mathcal{D}_a(T).\] No element of this fibre can be fixed by negation: if \(d=-d\), then \(2d=0\) and \(\pi(d)=a\neq 0\), so \(\langle d\rangle\) is a nontrivial subgroup disjoint from \(H\). Thus every order-two direction has an even fibre of size at least \(2\).
Choose a subgroup \[P=\{0,a,b,c\}\leq Q, \qquad c=a+b,\] with \(P\simeq C_2\times C_2\). We claim that the three nonzero fibres over \(a,b,c\) cannot all have size \(2\). Suppose, to the contrary, that they do. Put \[A=s(a),\qquad B=s(b),\qquad C=s(c).\] Since each of the three fibres is stable under negation and has two elements, we have \[C-B=\varepsilon A, \qquad C-A=\eta B, \qquad B-A=\theta C\] for some signs \(\varepsilon,\eta,\theta\in\{\pm1\}\). The first two identities give \[B+\varepsilon A=A+\eta B.\] If \(\varepsilon=1\) and \(\eta=-1\), this gives \(2B=0\). If \(\varepsilon=-1\) and \(\eta=1\), it gives \(2A=0\). If \(\varepsilon=-1\) and \(\eta=-1\), it gives \(2(A-B)=0\). Finally, if \(\varepsilon=\eta=1\), then \(C=A+B\); using \(B-A=\theta C\) gives either \(2A=0\) or \(2B=0\). In all cases, one of \(A\), \(B\), or \(A-B\) has order two. Its image in \(Q\) is respectively \(a\), \(b\), or \(a+b\), hence nonzero. The subgroup it generates is therefore nontrivial and disjoint from \(H\), again contradicting residuality. Thus at least one of the three fibres over \(a,b,c\) has size at least \(4\).
Counting quotient-direction fibres now gives \[\begin{align} |D(T)| &=\sum_{u\in Q}|\mathcal{D}_u(T)| \\ &\geq 1+\bigl(2+2+4\bigr)+2(q-4) \\ &=2q+1. \end{align}\] This proves the theorem. ◻
For square-modulus \(2\)-groups this gives an exact value in rank two. The follwing corollary is immediate.
Corollary 5. Let \(r\geq 2\), let \[G=(\mathbb{Z}/4\mathbb{Z})^r, \qquad H=2G.\] Then \[\delta(G,H)\geq 2^{r+1}+1.\] For \(r=2\), the equality \(\delta(G,H)=9\) is realized by the transversal \(T=\{0,1 \}^2\).
The preceding obstruction is specific to quotient directions of order two. For odd primes there is no analogous negation parity argument. The same square-modulus construction is residual for every prime, and the coordinate box gives the basic upper bound.
Proposition 13. Let \(p\) be a prime and let \[G=(\mathbb{Z}/p^2\mathbb{Z})^r, \qquad H=pG.\] Then \(m(G,H)=1\). Moreover the coordinate box gives \[\delta(G,H)\leq (2p-1)^r.\]
Proof. Let \(0\neq g\in G\). If \(g\) has order \(p\), then \(g\in pG=H\). If \(g\) has order \(p^2\), then \(pg\) is a nonzero element of \(\langle g\rangle\cap H\). Hence every nontrivial subgroup of \(G\) meets \(H\) nontrivially, and therefore \(m(G,H)=1\).
The coordinate box \[B_p^r=\{0,1,\ldots,p-1\}^r\subseteq(\mathbb{Z}/p^2\mathbb{Z})^r\] is a transversal for \(G/H\). Its difference support is \[D(B_p^r)=\{-(p-1),\ldots,p-1\}^r,\] which has size \((2p-1)^r\). This proves the upper bound. ◻
Thus, for odd \(p\), the pair \[\bigl((\mathbb{Z}/p^2\mathbb{Z})^2, p(\mathbb{Z}/p^2\mathbb{Z})^2\bigr)\] is the first same-prime residual case not covered by the cyclic formula, the two-torsion obstruction, or the one-nonsplit-coordinate product theorem. The rest of the paper rewrites its transversals as graphs \(\mathbb{F}_p^2\to\mathbb{F}_p^2\) and studies the resulting carry-corrected finite-field difference images.
We now specialize to the first unresolved same-prime residual case. Throughout this section \(p\) is an odd prime, \[G_p=(\mathbb{Z}/p^2\mathbb{Z})^2, \qquad H_p=pG_p.\] By Proposition 13, the pair is residual and the coordinate box gives \[\delta(G_p,H_p)\leq (2p-1)^2.\] We rewrite its transversals in finite-field terms and prove a uniform lower bound towards the box value. The resulting corrected-derivative images are finite-field image sets; these are adjacent to the sum-product/expander literature [19], [20] and, for graph/function language to relative difference sets and planar functions works such as [14].
We identify the quotient \(G_p/H_p\) with \(\mathbb{F}_p^2\). For \(x\in\mathbb{F}_p^2\), let \([x]\in\{0,1,\ldots,p-1\}^2\) denote the coordinatewise standard lift. Every normalized transversal for \(G_p/H_p\) has a unique form \[T_f=\{[x]+p[f(x)]:x\in\mathbb{F}_p^2\},\] where \(f:\mathbb{F}_p^2\to\mathbb{F}_p^2\) satisfies \(f(0)=0\). This normalization is harmless for difference supports: translating \(T_f\) by an element of \(pG_p\) subtracts a constant from \(f\).
For \(u,x\in\mathbb{F}_p^2\), define the carry vector \[c_u(x)=\frac{[x+u]-[x]-[u]}{p}\in\mathbb{F}_p^2.\] Here the numerator is computed in \(\mathbb{Z}^2\) using standard representatives; its coordinates are either \(0\) or \(-p\), so \(c_u(x)\) has coordinates \(0\) or \(-1\) modulo \(p\). We then define the carry-corrected derivative \[d_u f(x)=f(x+u)-f(x)+c_u(x),\] and its image \[A_u(f)=\{d_u f(x):x\in\mathbb{F}_p^2\}\subseteq\mathbb{F}_p^2.\]
The finite-field parametrisation gives an exact fibre decomposition of the difference support.
Lemma 1. For every \(f:\mathbb{F}_p^2\to\mathbb{F}_p^2\) one has \[|T_f-T_f|=\sum_{u\in\mathbb{F}_p^2}|A_u(f)|.\] Moreover \(A_0(f)=\{0\}\).
Proof. The difference between the representative above \(x+u\) and the representative above \(x\) is \[\begin{align} {[x+u]}+p[f(x+u)]-[x]-p[f(x)] &= [u]+p\bigl(f(x+u)-f(x)+c_u(x)\bigr) \\ &= [u]+p\,d_u f(x) \end{align}\] in \((\mathbb{Z}/p^2\mathbb{Z})^2\). In the quotient fibre \(u\in G_p/H_p\simeq\mathbb{F}_p^2\), the possible \(p\)-parts are exactly \(A_u(f)\). The quotient fibres are disjoint, so summing over all \(u\) gives the formula. For \(u=0\) one has \(c_0(x)=0\) and \(d_0f(x)=0\), hence \(A_0(f)=\{0\}\). ◻
The coordinate box corresponds to \(f=0\), and its support has size \((2p-1)^2\). This motivates the central conjecture.
Conjecture 14. For every odd prime \(p\), \[\delta\bigl((\mathbb{Z}/p^2\mathbb{Z})^2, p(\mathbb{Z}/p^2\mathbb{Z})^2\bigr)=(2p-1)^2.\] Equivalently, for every function \(f:\mathbb{F}_p^2\to\mathbb{F}_p^2\), \[\sum_{u\in\mathbb{F}_p^2}|A_u(f)|\geq (2p-1)^2.\]
The general residual fibre lower bound gives only \[|T_f-T_f|\geq 2p^2-1.\] Thus the conjecture asks for an additional \[(2p-1)^2-(2p^2-1)=2(p-1)^2\] forced differences coming from the rank-two same-prime geometry.
The deterministic input is a pair of elementary identities using only that the corrected derivatives come from a global quotient section.
Lemma 2. Let \(p\) be an odd prime and let \(f:\mathbb{F}_p^2\to\mathbb{F}_p^2\). For all \(u,v,x\in\mathbb{F}_p^2\) one has the curl identity \[d_u f(x+v)-d_u f(x)=d_v f(x+u)-d_v f(x).\] Moreover, for every nonzero \(u\in\mathbb{F}_p^2\) and every affine \(u\)-line, \[x+\langle u\rangle=\{x+iu:0\leq i\leq p-1\},\] one has the cycle identity \[\sum_{i=0}^{p-1} d_u f(x+iu)=-u.\] Consequently:
\(|A_u(f)|\geq 2\) for every \(u\neq 0\);
if \(|A_u(f)|=2\), then \(A_u(f)\) is contained in an affine line parallel to \(\langle u\rangle\).
Proof. Expanding the definition of \(d_u f\) gives \[\begin{align} d_u f(x+v)-d_u f(x) &=[f(x+u+v)-f(x+v)]-[f(x+u)-f(x)] \\ &\quad +c_u(x+v)-c_u(x), \end{align}\] and the analogous expression with \(u\) and \(v\) interchanged. The \(f\)-terms are identical. The carry terms are also identical, since \[c_u(x+v)-c_u(x) =\frac{[x+u+v]-[x+v]-[x+u]+[x]}{p} =c_v(x+u)-c_v(x).\] This proves the curl identity.
For the cycle identity, sum in \(G_p\) the \(p\) successive differences from the representative above \(x+iu\) to the representative above \(x+(i+1)u\): \[[u]+p\,d_u f(x+iu), \qquad 0\leq i\leq p-1.\] These differences go once around the affine cycle, so their sum is zero in \(G_p\): \[p[u]+p\sum_{i=0}^{p-1}d_u f(x+iu)=0\pmod {p^2}.\] Dividing by \(p\) modulo \(p\) gives \[u+\sum_{i=0}^{p-1}d_u f(x+iu)=0\] in \(\mathbb{F}_p^2\).
If \(A_u(f)=\{a\}\) for some \(u\neq 0\), then the cycle identity gives \(pa=0=-u\), a contradiction. Hence \(|A_u(f)|\geq 2\) for every nonzero \(u\).
Finally suppose \(A_u(f)=\{a,b\}\). On a fixed affine \(u\)-line, let \(n\) be the number of points at which \(d_u f\) takes the value \(a\). Then the other \(p-n\) points take the value \(b\), and the cycle identity gives \[na+(p-n)b=n(a-b)=-u.\] The cases \(n=0\) and \(n=p\) would make the left-hand side zero, so \(n\) is nonzero modulo \(p\). Thus \[a-b=-n^{-1}u\in\langle u\rangle.\] Therefore \(A_u(f)\) lies in an affine line parallel to \(\langle u\rangle\). ◻
Two independent quotient directions with minimal derivative images already force the conjectural box lower bound.
Lemma 3. Let \(p\) be an odd prime and let \(f:\mathbb{F}_p^2\to\mathbb{F}_p^2\). Suppose there are two linearly independent directions \(u,v\in\mathbb{F}_p^2\) such that \[|A_u(f)|=|A_v(f)|=2.\] Then \[|T_f-T_f|\geq (2p-1)^2.\]
Proof. By Lemma 2, the two images \(A_u(f)\) and \(A_v(f)\) are respectively contained in affine lines parallel to \(\langle u\rangle\) and \(\langle v\rangle\). Therefore \[d_u f(x+v)-d_u f(x)\in\langle u\rangle, \qquad d_v f(x+u)-d_v f(x)\in\langle v\rangle.\] By the curl identity these two quantities are equal. Since \(u\) and \(v\) are independent, \(\langle u\rangle\cap\langle v\rangle=\{0\}\), so both quantities vanish. Hence \(d_u f\) is invariant in the \(v\)-direction and \(d_v f\) is invariant in the \(u\)-direction.
Every nonzero \(w\in\mathbb{F}_p^2\) can be written uniquely as \[w=\lambda u+\mu v, \qquad \lambda,\mu\in\mathbb{F}_p.\] If exactly one of \(\lambda,\mu\) is nonzero, then Lemma 2 gives \(|A_w(f)|\geq 2\).
Assume now that \(\lambda\mu\neq 0\), and choose representatives \(\lambda,\mu\in\{1,\ldots,p-1\}\). Following first \(\lambda\) steps in the \(u\)-direction and then \(\mu\) steps in the \(v\)-direction, define \[B_\lambda(x)=\sum_{i=0}^{\lambda-1}d_u f(x+iu), \qquad C_\mu(x)=\sum_{j=0}^{\mu-1}d_v f(x+\lambda u+jv).\] The direct corrected derivative \(d_w f(x)\) differs from \(B_\lambda(x)+C_\mu(x)\) by a constant depending only on \(\lambda,\mu,u,v\) and on our standard-lift carry convention, not on \(x\). This is just the telescoping of the \(f\)-increments along the chosen path; the discrepancy between the path carry and the standard carry for \(w\) is independent of the starting point. Hence \(A_w(f)\) contains a translate of the sum of the two images of \(B_\lambda\) and \(C_\mu\). Since \(d_u f\) is invariant in the \(v\)-direction, \(B_\lambda\) depends only on the \(u\)-coordinate of \(x\) and its image lies in an affine line parallel to \(\langle u\rangle\). Moreover \(B_\lambda\) differs from the direct derivative \(d_{\lambda u}f\) by a constant, so \[|\operatorname{im} B_\lambda|=|A_{\lambda u}(f)|\geq 2.\] Similarly, because \(d_v f\) is invariant in the \(u\)-direction, \(C_\mu\) depends only on the \(v\)-coordinate, has image contained in an affine line parallel to \(\langle v\rangle\), and \[|\operatorname{im} C_\mu|=|A_{\mu v}(f)|\geq 2.\] As the \(u\)- and \(v\)-coordinates of \(x\) vary independently, the pair \((B_\lambda(x),C_\mu(x))\) runs over \(\operatorname{im}B_\lambda\times\operatorname{im}C_\mu\). Since the two image directions are parallel to the independent lines \(\langle u\rangle\) and \(\langle v\rangle\), addition is injective on their product. Therefore \[|A_w(f)|\geq |\operatorname{im} B_\lambda|\,|\operatorname{im} C_\mu| \geq 4\] for every mixed direction \(w=\lambda u+\mu v\) with \(\lambda\mu\neq 0\).
There are \(p-1\) nonzero multiples of \(u\), \(p-1\) nonzero multiples of \(v\), and \((p-1)^2\) mixed directions. Using Lemma 1, we get \[\begin{align} |T_f-T_f| &=1+\sum_{w\neq 0}|A_w(f)| \\ &\geq 1+2(p-1)+2(p-1)+4(p-1)^2 \\ &=(2p-1)^2. \end{align}\] ◻
This dichotomy gives an unconditional lower bound for every odd prime.
Theorem 15. For every odd prime \(p\), \[\delta\bigl((\mathbb{Z}/p^2\mathbb{Z})^2, p(\mathbb{Z}/p^2\mathbb{Z})^2\bigr) \geq 3p^2-p-1.\] Equivalently, every transversal \(T\) for \(G_p/H_p\) satisfies \[|D(T)|\geq 3p^2-p-1.\]
Proof. Normalize the transversal and write it as \(T=T_f\). Let \[S=\{u\in\mathbb{F}_p^2\setminus\{0\}: |A_u(f)|=2\}.\] If \(S\) contains two independent directions, then Lemma 3 gives the stronger bound \[|T_f-T_f|\geq (2p-1)^2.\] Since \[(2p-1)^2-(3p^2-p-1)=(p-1)(p-2)\geq 0\] for \(p\geq 3\), the desired estimate follows in this case.
It remains to consider the case where \(S\) contains no two independent directions. Then \(S\) is contained in a single one-dimensional subspace of \(\mathbb{F}_p^2\), so \[|S|\leq p-1.\] By Lemma 2, every nonzero direction contributes at least \(2\), and every nonzero direction outside \(S\) contributes at least \(3\). Together with \(A_0(f)=\{0\}\), this gives \[\begin{align} |T_f-T_f| &=1+\sum_{u\neq 0}|A_u(f)| \\ &\geq 1+2|S|+3\bigl(p^2-1-|S|\bigr) \\ &=3p^2-2-|S| \\ &\geq 3p^2-p-1. \end{align}\] This proves the theorem. ◻
The remaining gap to the conjectural box value is \((p-1)(p-2)\); Section 7 gives asymptotic evidence for Conjecture 14.
We verified the conjecture for \(p=3\) and \(p=5\) by deterministic Python computations using the graph parametrisation and fibre decomposition above. For \(p=5\) the search also uses the cycle and curl constraints from this section and the action of \(\operatorname{GL}_2(\mathbb{F}_5)\). Since \(T-T\) is negation-invariant and \(0\) is the only self-inverse element in an odd-order group, \(|T-T|\) is odd. Thus it suffices to exclude the largest odd support below the box value: the computations rule out \(|T_f-T_f|\le 23\) for \(p=3\) and \(|T_f-T_f|\le 79\) for \(p=5\). The coordinate boxes attain \(25\) and \(81\), so \[\delta\bigl((\mathbb{Z}/9\mathbb{Z})^2,3(\mathbb{Z}/9\mathbb{Z})^2\bigr)=25, \qquad \delta\bigl((\mathbb{Z}/25\mathbb{Z})^2,5(\mathbb{Z}/25\mathbb{Z})^2\bigr)=81.\]
The Python scripts used and the results could be consulted on our online companion github.com/georgeturcasubb/transversals.
We could not complete a verification of the conjecture for \(p=7\) due to the high computational complexity. However, in these case we conducted many unsuccessful searches for transversals that could give rise to counter-examples.
The exact values in Subsection 6.4 cover the first two odd primes. We add two asymptotic pieces of evidence: a uniformly random lifting satisfies the box lower bound with probability tending to one, and no fixed integer-polynomial rule produces counterexamples for all sufficiently large primes.
Call a nonzero direction \(u=(u_1,u_2)\in\mathbb{F}_p^2\) mixed if \(u_1u_2\neq 0\). For a function \(f:\mathbb{F}_p^2\to\mathbb{F}_p^2\), define \[B(f)=\{u\in\mathbb{F}_p^2:u_1u_2\neq 0\text{ and } |A_u(f)|\leq 3\}.\] Thus \(B(f)\) records mixed directions whose corrected derivative image is too small for the box contribution.
Theorem 16. Let \(p\) be an odd prime. Choose \(f:\mathbb{F}_p^2\to\mathbb{F}_p^2\) uniformly at random, with the values \(f(x)\) independent and uniformly distributed. Then \[\Pr_f\bigl(|T_f-T_f|<(2p-1)^2\bigr) \leq p^8\left(\frac{p^2 3^p}{p^{2p}}\right)^p.\] In particular, \[\Pr_f\bigl(|T_f-T_f|\geq (2p-1)^2\bigr)\longrightarrow 1 \qquad (p\to\infty).\] The same conclusion holds for uniformly random normalized liftings \(f(0)=0\).
Proof. Before normalization, uniform functions give the uniform model on transversals for \(G_p/H_p\); subtracting \(f(0)\) only translates \(T_f\) by an element of \(pG_p\).
Fix \(f\). By Lemma 1, \[|T_f-T_f|=\sum_{u\in\mathbb{F}_p^2}|A_u(f)|.\] The zero direction contributes \(1\), and Lemma 2 gives at least \(2\) from every nonzero direction. There are \(2(p-1)\) nonzero axis directions and \((p-1)^2\) mixed directions; the mixed directions outside \(B(f)\) contribute at least \(4\), while those in \(B(f)\) still contribute at least \(2\). Hence \[\begin{align} |T_f-T_f| &\geq 1+2\cdot 2(p-1)+2|B(f)|+4\bigl((p-1)^2-|B(f)|\bigr) \\ &= (2p-1)^2-2|B(f)|. \end{align}\] Thus \(B(f)=\varnothing\) is sufficient for the box lower bound, and failure forces \(B(f)\neq\varnothing\).
We bound this last event. Fix a mixed direction \(u\) and a set \(S\subseteq\mathbb{F}_p^2\) with \(|S|\leq 3\). The translation \(x\mapsto x+u\) partitions \(\mathbb{F}_p^2\) into \(p\) disjoint \(p\)-cycles. On one cycle, write \(x_i=x_0+iu\) and \(y_i=f(x_i)\). The condition \(d_u f(x_i)\in S\) is \[y_{i+1}-y_i\in S-c_u(x_i).\] After choosing \(y_0\), there are at most \(3\) choices for each edge increment; ignoring the closing condition only overcounts. Thus one cycle has probability at most \[\frac{p^2 3^p}{p^{2p}}.\] The \(p\) cycles use disjoint values of \(f\), hence are independent. The probability that this fixed \(S\) contains all values of \(A_u(f)\) is therefore at most \(\left(\frac{p^2 3^p}{p^{2p}}\right)^p\). Hence there are at most \(\sum_{j=0}^3 \binom{p^2}{j}\leq p^6\) subsets of \(\mathbb{F}_p^2\) of size at most \(3\). It follows that, for the fixed mixed \(u\), \[\Pr_f\bigl(|A_u(f)|\leq 3\bigr) \leq p^6\left(\frac{p^2 3^p}{p^{2p}}\right)^p.\] Finally, there are \((p-1)^2<p^2\) mixed directions. A union bound gives \[\Pr_f(B(f)\neq\varnothing) \leq p^8\left(\frac{p^2 3^p}{p^{2p}}\right)^p.\] Since failure of the box lower bound implies \(B(f)\neq\varnothing\), this proves the estimate. Its logarithm is \(-2p^2\log p+p^2\log 3+O(p\log p)\), which tends to \(-\infty\). ◻
The random theorem shows that counterexamples, if they exist, are rare among all functions. We next rule out another natural source: liftings defined by one fixed polynomial rule as the prime varies.
Let \(g=(g_1,g_2)\in\mathbb{Z}[X,Y]^2\) be fixed. For a prime \(p\), let \(g_p:\mathbb{F}_p^2\to\mathbb{F}_p^2\) be its reduction modulo \(p\), and set \[T_{g,p}=\{[z]+p[g_p(z)]:z\in\mathbb{F}_p^2\} \subseteq (\mathbb{Z}/p^2\mathbb{Z})^2.\] The word fixed is essential: for a single prime \(p\), every function \(\mathbb{F}_p^2\to\mathbb{F}_p^2\) has polynomial representatives of degree less than \(p\) in each variable, for instance by finite-field interpolation. The theorem excludes only one integer-polynomial rule independent of \(p\). This fixed-polynomial setting is close to value-set questions for polynomial maps over finite fields [21] and to finite-field polynomial/rational image-expansion results [22].
Theorem 17. For every fixed \(g=(g_1,g_2)\in\mathbb{Z}[X,Y]^2\), there is an integer \(B_g\) such that, for every prime \(p>B_g\), \[|T_{g,p}-T_{g,p}|\geq (2p-1)^2.\]
The proof uses a nonzero-Jacobian image bound, large-rectangle estimates, and a degenerate-Jacobian classification.
Lemma 4. Fix \(E\geq 1\). Let \[F=(P,Q):\mathbb{A}^2_{\mathbb{F}_p}\to\mathbb{A}^2_{\mathbb{F}_p}\] be a polynomial map with \(\deg P,\deg Q\leq E\) and \(\det JF\not\equiv 0\). If \(R\subseteq\mathbb{F}_p^2\), then \[|F(R)|\geq \frac{|R|-C_Ep}{E^2},\] where one may take \(C_E=\max\{1,2(E-1)\}\).
Proof. Let \(\Delta=\det JF\). Its zero set has at most \(C_Ep\) points by the elementary polynomial zero bound over finite fields. For \(a\in\mathbb{F}_p^2\), the fibre \(F^{-1}(a)\) is cut out by two plane curves of degrees at most \(E\). Discard any common fibre components contained in \(\{\Delta=0\}\); those points are already excluded below. Off the critical curve, the remaining fibre curves have no common component: otherwise their gradients would be dependent at a smooth point of that component, forcing \(\Delta\) to vanish generically there. Bezout’s theorem for plane curves [23] then bounds every noncritical fibre by \(E^2\) geometric, hence rational, points. Thus the points of \(R\) outside \(\{\Delta=0\}\) map with fibres of size at most \(E^2\), and \[|F(R)|\geq \frac{|R|-C_Ep}{E^2}.\] ◻
We also use three elementary estimates on large rectangles. Let \(I,J\subseteq\mathbb{F}_p\) be intervals of consecutive standard residues, each of size at least \(p/2\), and put \(R=I\times J\).
A nonzero linear form \(\ell\) has \(|\ell(R)|\geq p/2\). This is immediate if one coefficient vanishes, and otherwise follows from Cauchy–Davenport [9].
If \(P\in\mathbb{F}_p[S]\) is nonconstant of degree at most \(d\), then \(|P(S_0)|\geq |S_0|/d\) for every \(S_0\subseteq\mathbb{F}_p\).
If \(Q\in\mathbb{F}_p[X,Y]\) is nonconstant of degree at most \(d\), then every nonempty fibre has at most \(dp\) points by the elementary zero bound, so \(|Q(R)|\geq |R|/(dp)\geq p/(4d)\).
The second input handles the degenerate Jacobian case, using singular subspaces of \(2\times 2\) matrices.
Lemma 5. Let \(k\) be an infinite field and let \(L\subseteq M_2(k)\) be a linear subspace such that every element of \(L\) is singular. Then either all matrices in \(L\) have a common nonzero kernel vector, or all their images lie in a common line in \(k^2\).
Proof. If \(L=0\), both alternatives hold. Otherwise choose a nonzero rank-one matrix in \(L\) and change source and target bases so that it is \(E=\begin{pmatrix}1&0\\0&0\end{pmatrix}\). For \(B=\begin{pmatrix}a&b\\ c&d\end{pmatrix}\in L\), the matrix \(E+tB\) is singular for every \(t\in k\). Since \(k\) is infinite, \[\det(E+tB)=td+t^2(ad-bc)\] vanishes as a polynomial in \(t\). Hence \(d=0\). Since \(B\) is singular, also \(bc=0\). Linearity forces either \(b=0\) for all \(B\in L\), or \(c=0\) for all \(B\in L\); otherwise a linear combination would have both off-diagonal entries nonzero and would be nonsingular. These are exactly the common-kernel and common-image cases. ◻
This gives the needed classification for polynomial maps with singular Jacobian differences.
Lemma 6. Let \(k\) be a field of characteristic zero, and let \(g:k^2\to k^2\) be a polynomial map satisfying \[\det(Jg(w)-Jg(z))=0 \qquad \text{for all } z,w\in k^2.\] Then there are \(A\in M_2(k)\) and \(c\in k^2\) such that one of the following holds: \[g(z)=Az+G(\ell(z))+c\] for a nonzero linear form \(\ell:k^2\to k\) and a one-variable polynomial map \(G:k\to k^2\), or \[g(z)=Az+w\varphi(z)+c\] for some nonzero \(w\in k^2\) and some polynomial \(\varphi\in k[X,Y]\).
Proof. Fix \(z_0\) and let \(L\) be the linear span of \[\{Jg(z)-Jg(z_0):z\in k^2\}.\] The determinant quadratic form vanishes on each generator and on each difference of two generators. Its polar form therefore vanishes on pairs of generators, so the determinant vanishes on all of \(L\). Lemma 5 gives a common kernel vector or a common image line.
Write \(Jg(z)=A+N(z)\) with \(N(z)\in L\). If \(L\) has a common kernel vector \(r\neq 0\), the directional derivative of \(g-Az\) along \(r\) is zero. After a linear change of coordinates, \(g-Az\) is independent of one coordinate and has the form \(G(\ell(z))+c\). If \(L\) has common image line \(kw\), all derivatives of \(g-Az\) lie in \(kw\). A linear form annihilating \(w\) is constant on \(g-Az\), so \(g-Az-c\) takes values in \(kw\) and \(g(z)=Az+w\varphi(z)+c\). ◻
Proof of Theorem 17. We first remove affine terms, which do not change the support size. Adding a constant translates \(T_{g,p}\) by an element of \(pG_p\). Adding a linear term \(Az\) applies the automorphism \[U_A(t)=t+pA\overline{t}, \qquad t\in(\mathbb{Z}/p^2\mathbb{Z})^2,\] where \(\overline{t}\) denotes reduction modulo \(p\); its inverse is \(U_{-A}\). Thus both operations preserve \(|T_{g,p}-T_{g,p}|\), and affine-linear liftings have the coordinate-box support \((2p-1)^2\). Let \(D=\deg g\). If \(D\leq 1\), we are done.
Assume first that \(D\geq 2\) and \[\Delta_g(z,v)=\det\bigl(Jg(z+v)-Jg(z)\bigr)\] is not the zero polynomial over \(\mathbb{Q}\). Expand \(\Delta_g\) in the \(z\)-variables and choose a nonzero coefficient polynomial \(H(v)\). After excluding finitely many primes depending only on \(g\), the reduction of \(H\) is nonzero; hence, for all sufficiently large \(p\), all but \(O_g(p)\) directions \(v\in\mathbb{F}_p^2\) have \(H(v)\neq 0\). For these directions, the finite difference \[F_v(z)=g_p(z+v)-g_p(z)\] has degree at most \(D-1\) and nonzero Jacobian determinant.
For each good \(v\), choose a carry cell \(R_v\): a rectangular cell of standard residue representatives on which \(c_v(z)\) is constant, with both side lengths at least \(p/2\), so \(|R_v|\geq p^2/4\). On \(R_v\), the map \(z\mapsto d_v g_p(z)\) is a constant translate of \(F_v(z)\). By Lemma 4, with \(E=D-1\), \[|A_v(g_p)|\geq \frac{p^2/4-C_Ep}{E^2}.\] There are \(p^2-O_g(p)\) good directions, each contributing \(\gg_g p^2\), so Lemma 1 gives \[|T_{g,p}-T_{g,p}|=\sum_v |A_v(g_p)|\gg_g p^4.\] For all sufficiently large primes this exceeds \((2p-1)^2\), proving the generic case.
It remains to treat the case \(\Delta_g\equiv 0\). Then \[\det(Jg(w)-Jg(z))=0\] for all \(z,w\) over \(\mathbb{Q}\), so Lemma 6 applies. Excluding the finitely many primes where the rational classification data have denominator problems, where the nonzero linear form or target vector vanishes after reduction, where the relevant leading forms vanish, or where the characteristic is too small for the degree bounds, we remove the affine part as above. One of two forms remains.
First suppose \(g(z)=G(\ell(z))\), with \(\ell\neq 0\). If \(m=\deg G\leq 1\), the map is affine-linear, already handled. Otherwise, for all sufficiently large primes and every \(v\) with \(\ell(v)\neq 0\), the difference \(G(S+\ell(v))-G(S)\) has a nonconstant coordinate of degree \(m-1\). On a large carry cell, the linear-form and univariate image estimates above give \[|A_v(g_p)|\geq \frac{p}{2(m-1)}.\] The \(p^2-p\) such directions contribute \(\gg_g p^3\), which exceeds \((2p-1)^2\) for all sufficiently large \(p\).
Second suppose \(g(z)=w\varphi(z)\), with \(w\neq 0\). If \(m=\deg\varphi\leq 1\), the map is affine-linear, already handled. Otherwise, for all sufficiently large primes the highest homogeneous part \(\varphi_m\) remains nonconstant. The directions \(v\) for which \(D_v\varphi_m\) vanishes identically form a proper linear subspace, so all but at most \(p\) directions give a nonconstant difference \(\varphi(z+v)-\varphi(z)\) of degree at most \(m-1\). On a large carry cell, \(d_v g_p\) is a constant translate of \(w(\varphi(z+v)-\varphi(z))\), so multiplication by \(w\neq 0\) preserves cardinality and the bivariate image estimate gives \[|A_v(g_p)|\geq \frac{p}{4(m-1)}.\] Again the total contribution is \(\gg_g p^3\), which exceeds \((2p-1)^2\) for all sufficiently large \(p\).
The generic and degenerate cases give an integer \(B_g\) such that the claimed inequality holds for every prime \(p>B_g\). ◻
Remark 18. These results do not settle Conjecture 14: they exclude random counterexamples with high probability and fixed algebraic rules for large \(p\), while the conjecture allows arbitrary \(p\)-dependent liftings.
The cyclic and one-nonsplit-coordinate cases are settled by the fibre lower bound, singleton reduction, and Kneser estimate. The remaining obstruction is the same-prime nonsplit rank-two case: the odd-prime box statement is still conjectural, and the fixed-prime certificates and asymptotic evidence have only the limited scope proved above. The fixed-polynomial evidence also resonates with the broader finite-field polynomial-method tradition, exemplified by Dvir’s finite-field Kakeya breakthrough [24]; however, our result is not a Kakeya theorem and uses only the polynomial image constraints developed above.
The central open problem is Conjecture 14. In the notation of Lemma 1, it asks whether \(\sum_u |A_u(f)|\geq (2p-1)^2\) for every \(f\), with equality for the coordinate box. Theorem 15 leaves gap \((p-1)(p-2)\), so any counterexample would have to realize a structured loss among low corrected-derivative images.
Sharper bounds for low corrected-derivative images may require incidence technology beyond the cycle-curl identities used here, such as Rudnev’s [25] point-plane incidence theorem and the Stevens–de Zeeuw point-line incidence bound over arbitrary fields [26].
Problem 19. Let \(p\) be odd and let \(f:\mathbb{F}_p^2\to\mathbb{F}_p^2\). Classify, or sharply bound, the oriented nonzero directions \(u\) for which \(|A_u(f)|\leq 3\). In particular, decide whether any loss from these low images is necessarily compensated by larger corrected-derivative images in the remaining directions.
The coordinate box suggests the following higher-rank analogue.
Conjecture 20. For every odd prime \(p\) and every \(r\geq 1\), \[\delta\bigl((\mathbb{Z}/p^2\mathbb{Z})^r, p(\mathbb{Z}/p^2\mathbb{Z})^r\bigr)=(2p-1)^r.\]
For higher-rank quotients, one natural next layer could be iterated sumset-growth technology in the Plünnecke–Ruzsa tradition [27], [28], together with arbitrary-abelian-group small-doubling structure [11].
The cases \(r=1\) and \(r=2\) are respectively the cyclic residual theorem and Conjecture 14. For \(r\geq 3\), the same graph-and-fibre identity holds, and the rank-two argument suggests looking for a frame of low directions and product-type lower bounds for mixed directions.
The primitive quotient reduction separates singleton directions from the primitive quotient part, so the remaining classification can be phrased in primitive terms.
Problem 21. Classify the pairs \(H\leq G\) for which every primitive transversal \(T\) satisfies \(|D(T)|\geq 2|G/H|-1\), and identify the primitive quotient geometries that force a strict improvement.
Cyclic residual quotients attain the bound \(2|G/H|-1\); residual quotients containing a \(C_2\times C_2\) plane do not. The product results control one nonsplit cyclic coordinate with arbitrary split factors, leaving the next case.
Problem 22. Let \(p\) be prime, \(G=\prod_{i=1}^r C_{p^{a_i}}\), and \(H=\prod_{i=1}^r p^{b_i}C_{p^{a_i}}\), with \(1\leq b_i<a_i\) for at least two indices. Determine \(\delta(G,H)\), or give sharp lower bounds in terms of the nonsplit \(p\)-primary coordinates.
The square-plane case is the smallest instance of Problem 22. Resolving it would clarify the first residual nonsplit geometry left open by the transversal difference number.
The computations mentioned in Subsection 6.4 were carried using Python and are released in the form of an online companion on GitHub at the link below
The authors used OpenAI ChatGPT Deep Research, accessed through ChatGPT to help locate and organize potentially relevant bibliographic material. The authors also used OpenAI Codex with GPT-5.5, accessed through the Codex agentic programming environment, to assist with Python experiments, realization of the GitHub repository and copy-editing of the manuscript. These tools were used as research aids, not as authors. In line with the transparency and human-responsibility principles of the Leiden Declaration on Artificial Intelligence and Mathematics [29], the human authors checked the sources, code, numerical results and take full responsibility for the results presented in the article.