WorksMathematics
Integral zero-sum incidence identities and Davenport constants in rank three
Abstract
We study the Davenport constant of finite abelian groups through signed identities among zero-sum subsequences. An integral incidence identity of degree for a sequence assigns integer weights to its nonempty zero-sum subsequences so that the weights add up to one, while those of the subsequences containing any given nonempty set of at most terms add up to zero. Over a finite abelian -group , the zero-sum subsequences of a sequence of any length have an indicator polynomial of degree at most , and this gives a transfer principle: if, for each , every sequence of terms over admits such an identity of degree , then for every finite abelian -group . For we construct integral identities at length over , and prove that these thresholds are sharp over fields of characteristic zero or characteristic not dividing . For a prime and , we prove that In particular and , which are the two smallest cases of the conjecture left open in the literature. The proofs of the base cases for are computer-assisted. All of the results in this paper have been formalised in Lean 4.
In particular and , which are the two smallest cases of the conjecture left open in the literature. The proofs of the base cases for are computer-assisted. All of the results in this paper have been formalised in Lean 4.
Introduction
Let be a finite abelian group, written additively. A sequence over is a finite list of elements of in which repetitions are allowed, and the Davenport constant is the least integer such that every sequence of terms over has a nonempty subsequence with sum zero. If
then copies of each basis element form a zero-sum free sequence, so
Olson proved that equality holds for -groups and for groups of rank at most two [16, 17]. Equality fails for infinitely many groups of rank four [11, 8]. In rank three, it is conjectured that
For background on the conjecture, see [7, 8, 19, 10]. In this paper, we develop a method for attacking particular cases of this conjecture. We write the group as with a -group, and read a sequence over it as a sequence over the quotient whose terms carry labels in the kernel . We first study identities among the zero-sum subsequences of sequences over a small quotient, and then use those identities for arbitrary labels in a -group kernel.
Our method allows us to prove the conjecture (1) for infinitely many mixed-prime groups of rank three with a quotient or , and the values it produces include the two smallest cases of (1) for that the classical theory leaves open, and .
The inductive method. The standard way to bound from above when is not a -group reduces the question to two smaller groups, a subgroup and the quotient . It goes back to the earliest work on the subject [17, 20], and is known in the zero-sum literature as the inductive method [19 Section 3], [3, 14]. A zero sum in a quotient need not be a zero sum in the original group. Project a sequence over to , and call a subsequence a block if its projection has sum zero. Each block has a sum in ; if we can find pairwise disjoint blocks, some nonempty collection of their sums adds to zero in , and the union of those blocks is a zero sum in [5 Proposition 2.6].
Short blocks make this argument efficient. Write for the least length such that every sequence over of that length contains a nonempty zero-sum subsequence of length at most . Extracting short blocks successively gives, in particular,
The same estimate holds for in the extension above.
The last block need not be short, and this improves the count. With , extract short blocks and then one arbitrary zero-sum block. This requires at most
terms. The first expression ensures that the last short block can be extracted; the second ensures that a final zero sum remains. The maximum is the second expression precisely when
For , the values and satisfy this inequality [17, 8]; this is the count that makes the inductive method exact in rank two. For and , however, the comparison is and , respectively [9, 15]. For the groups studied below, the second expression is the basis lower bound, so for the count overshoots it. Nor is the loss due to the insistence on short blocks: ten terms over need not contain two disjoint blocks of any size (Appendix § 8), although .
Once disjoint blocks have been chosen, this argument uses only their sums in . It forgets the other zero sums of the projection and the ways in which they overlap. Our approach keeps that incidence information and combines the zero sums with signed weights.
Incidence identities. We regard sequences as indexed, so equal group elements at different positions are considered different terms, since the kernel labels attached to those positions may differ. For over , write
for the family of its zero sets, and let be the incidence vector of .
Let be a non-negative integer, and let be a commutative ring with identity. An incidence identity of degree over , or -identity, is a family of elements of satisfying
The sum indexed by is the moment at . Thus the weights add to one, but the weights through any prescribed nonempty set of at most positions cancel. We say the identity is integral when .
For example, if and are disjoint nonempty zero sets, give weights , , to , , , respectively. The total weight is one, and the total through each individual position is zero. These are exactly the conditions for degree one.
The polynomial meaning is the reason for requiring all these cancellations. On Boolean inputs, the monomial records whether a set contains every position of . Consequently (3) is equivalent to
We prove this in Lemma 2.1. The empty subsequence receives no weight; nevertheless, these signed evaluations reproduce its value for every polynomial of the indicated degree.
For an abelian -group , the test for a selected set of kernel labels to have sum zero is such a polynomial over : its degree is at most , independently of the number of labels (Lemma 3.1). An identity of this degree for the sequence over the quotient can therefore be evaluated at the kernel’s zero-sum test, and this gives the transfer principle.
Theorem A (Transfer principle). Let be a finite abelian group, let be a prime, and let be integers such that every sequence of terms over has an -identity over . Then
If the identities are integral, the conclusion holds for every prime .
To see the mechanism, let be the zero-sum indicator for the kernel labels and evaluate the quotient identity at . The result is
Some summand is nonzero, so some nonempty zero set in the quotient also has sum zero in the kernel. For a fixed degree, the weights depend only on the quotient sequence; the same identity works for every choice of kernel labels. Integer weights can be reduced modulo any prime, which makes a single certificate useful for many different kernels.
Over a field , the alternative to an -identity is a separator of degree at most $t: a polynomial of degree at most such that and for every . Lemma 2.1 shows that exactly one of these alternatives occurs. We write for the least such that no sequence of terms over has a separator of degree at most . In degree zero the only possible separator is the constant polynomial $1$, which works exactly when there is no nonempty zero sum. Thus .
The quotient hierarchy. For and with , Olson’s theorem and the Chinese remainder theorem give
This singles out the target length : a quotient identity at this length has exactly the degree and length needed to reach the basis lower bound after transfer.
Theorem B. Let and let be an integer. Every indexed sequence of terms over has an integral -identity. Consequently for every field , with equality whenever the characteristic of is zero or does not divide .
The case gives a small geometric model of the argument: a signed identity on the Fano plane with one point removed. For and , finite computations supply the base cases, and the three constructions of Lemma 2.2 propagate them to all degrees.
Combining Theorems A and B gives for every finite abelian -group . When has rank at most three and , the right-hand side is the basis lower bound, giving the following consequence.
Corollary C. Let , let be a prime with , and let $0 \le a \le b \le c$ be integers. Then
In particular for and for odd .
For example, . As and , the classical estimate (2) only gives the upper bound .
Trivial kernel factors are allowed, with . When , the rank-three groups displayed in (4) are themselves -groups, and their Davenport constants are already given by Olson’s theorem.
Earlier work and our contribution. Corollary C includes classical families: the doubled -groups and tripled -groups of van Emde Boas and Kruyswijk [20, 21], as well as special cases covered by later rank-three results [4, 3, 19, 14]. For earlier work covers, besides the prime powers, the exponents and [19, 13], but not or with .3
The group-algebra input goes back to Olson, and polynomial methods have a substantial history in zero-sum theory [16, 1]. Finite computational input to the inductive method also occurs in the work of Bhowmik, Halupczok and Schlage-Puchta [3]. Related signed zero-block and moment methods occur in Volkmann’s work on nonabelian groups of exponent three [22, 23, 24]. Here we establish complete incidence identities over the integers, throughout the degree hierarchy for three small quotients. The transfer principle explains how those identities control arbitrary labels in an abelian -group kernel, while integrality supplies a common certificate in every characteristic.
Scope. The bounded-degree indicator in Theorem A uses a -group kernel. Remark 3.3 explains why a degree bound independent of the number of labels fails for other kernels in characteristic . If has at least three distinct prime divisors, separating off one primary component of leaves a quotient with at least two prime divisors. An incidence hierarchy at length for such a quotient would already include the Davenport conjecture for that quotient in degree zero. Deleting short zero sets, as in Section 5, reduces a proposed hierarchy to finitely many base degrees for each fixed ; establishing those base cases is still a mathematical task.
Organisation. § 2 develops the identities and their three constructions. § 3 proves the transfer principle, and § 4 gives the lower bound. Section 5 establishes the quotient hierarchies and § 6 is dedicated to the proof of the announced Davenport values. § 7 records the computational dependencies and the formal verification, with the coverage arguments and certificate specifications in Appendix § 8.
AI disclosure. Large language models (OpenAI’s Codex and ChatGPT, Anthropic’s Claude) were used in an essential way for this work: in the exploration that led to the incidence identities, which play a key role in the main results of this paper, in coding the enumeration programs, in writing and revising the exposition and the proofs, and in a substantial part of the Lean formalisation. The authors made a significant effort to improve the exposition of the AI-produced mathematics and to give proper credit to prior work that uses similar ideas. Every argument was check independently by at least one author and the author take full responsibility for the paper.
Incidence identities and three constructions
For a sequence write ; the notation denotes occurrences of . A subsequence is specified by its set of positions.
Lemma 2.1. Let be a family of elements of indexed by .
is an -identity over if and only if
for every polynomial of degree at most .
An -identity over is an -identity over every commutative ring, by reduction of the coefficients.
If is a field, has an -identity over if and only if has no separator of degree at most over .
Proof. Both sides of (5) are linear in . On Boolean inputs, every polynomial agrees with a multilinear polynomial of no greater degree, so it is enough to test the monomials with , including . For these monomials the equality is , exactly the moment conditions.
Part (2) is clear, because we always have a ring morphism .
For (3), consider the vectors of monomial values, for , and the vector of the empty set. Observe that an identity is an expression of as a linear combination of the , and a separator is a linear functional vanishing on all and equal to on . Over a field, either belongs to the span of the vectors , or there is a linear functional vanishing on that span and taking the value at . The first alternative gives the identity; the second gives the separator.
An identity on a subsequence extends to the full sequence by assigning zero weight to the other zero sets. Conversely, a separator for the full sequence restricts to one for a subsequence by setting the omitted coordinates to zero. Thus existence of an identity, and absence of a separator, are inherited by longer sequences. The invariant is finite: terms contain disjoint zero sets, and Lemma 2.2(1) supplies an identity.
The identities that Section 5 obtains by computer are stored with integer weights. Reducing their coefficients modulo gives the identities needed for Theorem A, so one exact certificate check serves every characteristic. Solvability over each field separately would give neither the integral assertion of Theorem B nor such a common certificate.
We use three ways to build identities. Disjoint zero sets give inclusion–exclusion. Adjoining a disjoint zero set turns a finite difference into an identity of one higher degree. Taking complements inside a zero set cancels the highest-degree terms in a different way.
Lemma 2.2 (Three constructions). Let be an indexed sequence and . The following produce integral -identities for from integral identities of lower degree.
If has pairwise disjoint zero sets, then has an -identity.
Let . If and a subsequence of disjoint from has an -identity, then has an -identity.
Let . If and a proper subsequence has an -identity, then has an -identity.
Proof. For (1), let be disjoint nonempty zero sets. Give the union of every nonempty collection of blocks the weight , and give every other zero set weight zero. Distinct collections have distinct unions. If a nonempty set of at most positions is contained in such a union, it misses some block . The collections whose union contains are paired by adding or removing , with opposite weights. Thus every required nonconstant moment vanishes. The total weight is .
For (2), let be the identity of degree on a subsequence disjoint from . Define new weights by
and set the remaining weights to zero. These sets are nonempty zero sets and are pairwise distinct. Let be multilinear of degree at most . Restricting to the coordinates outside , the polynomial
has degree at most $t-1. Indeed, terms involving no coordinate of cancel, and each remaining term has spent at least one degree on a coordinate of . Applying the old identity gives
Rearranging yields .
For (3), let be the identity on , and put . Give weight to , weight to , and weight to , with zero weights elsewhere. The complements are nonempty zero sets. The assigned sets are distinct: each complement contains , whereas each lies in ; and is neither an nor the complement of a nonempty . Restrict a multilinear polynomial of degree at most to the coordinates of . Then
has degree at most $t-1: complementation multiplies each degree- term by , so the degree- parts cancel. Applying the identity on to and rearranging gives
All three constructions use only addition and subtraction of coefficients, so they preserve integrality.
The transfer principle
For a cyclic kernel , identified with the additive group of , the required indicator is familiar. Fermat’s theorem gives
At a Boolean vector , this is 1 precisely when , and it has degree at most $p-1, however many labels there are. Multiplying one such factor for each coordinate gives the same phenomenon for . The next lemma extends it to arbitrary abelian -groups, including cyclic factors of order .
Lemma 3.1. Let and . For arbitrary , the Boolean function
is represented by a multilinear polynomial over of degree at most .
Proof. Write for the basis element of indexed by , so that . Choose generators of the cyclic factors, with of order , and put . Then
and the augmentation ideal satisfies , since a monomial of total degree in the has some exponent at least . Now we consider the multilinear polynomial
where extracts the identity coefficient in the group basis. Multiplication records the selected sum: setting omits , and setting includes it. For a Boolean vector the product collapses to , so . The coefficient of in is . Before taking , this is a product of elements of the augmentation ideal, and therefore lies in . Once , that product, and hence the corresponding scalar coefficient of , is zero. This is why the degree is bounded by rather than by the number of labels. Thus the degree of our polynomial is at most . □
Olson’s theorem. The same nilpotence gives . If terms were zero-sum free, the product would have identity coefficient , contributed only by the empty selection. But this product lies in . Hence . Conversely, copies of each form a zero-sum free sequence of length , giving equality. For this basis sequence, only the empty subset has sum zero. Its unique multilinear indicator is , of degree . Thus the indicator degree bound is best possible in general.
Proof of Theorem A. Put and take terms in . The quotient sequence has an -identity over . Let be the indicator polynomial for the kernel labels supplied by Lemma 3.1. Since ,
Some summand is nonzero. For its nonempty set , by definition, while gives . Thus the original sequence has a nonempty zero sum, and . If the identities are integral, reducing their coefficients modulo makes the same argument available for every prime. □
Remark 3.2 (The elementary block bound). A sequence of length contains disjoint short zero sets, obtained by repeated extraction. Their unions, with inclusion–exclusion weights, give an -identity by Lemma 2.2(1). For a -group kernel, Theorem A therefore recovers (2), and in the same way every bound obtained from disjoint blocks, whatever their sizes. Incidence identities allow further cancellations among zero sets that do not arise from a disjoint block decomposition. The six-term Fano example below already requires such overlap.
Remark 3.3 (The kernel hypothesis). The degree bound in Lemma 3.1 cannot hold uniformly in the number of labels when is not a -group. Choose an element of prime order and take all labels equal to . The coefficient of in their zero-sum indicator is
where the middle equality is Frobenius and because . A function on has a unique multilinear representative over , so the indicator has degree exactly . Since is arbitrary, there is no bound depending only on .
For example, when has order two and is odd, the indicator is
the test that the number of selected labels is even. Thus the bounded-degree indicator argument cannot be applied unchanged to such kernels.
Sharpness
Proposition 4.1. Let and , and let be a field of characteristic zero or of characteristic not dividing . Then
Proof. We choose a sequence whose zero sets are especially simple:
of length . A zero set contains no first- or second-axis term, because there are fewer than of each. It therefore consists of third-axis positions, and its size is for some .
We look for a separator among the linear combinations of the elementary symmetric polynomials in those third-axis coordinates. On a set of size , their values are . It suffices to show that the integer matrix
is invertible over .
To compute its determinant, let count the -element subsets of a fixed union of disjoint -element blocks that meet every block. Put and for . A -element subset of a union of blocks first chooses the blocks it meets, giving
Thus , where is lower triangular with diagonal entries one, and is upper triangular. Moreover, : a -element subset meeting blocks chooses one element from each. Consequently
This identity was obtained over the integers without division. The determinant has nonzero image in under the stated characteristic hypothesis. We can therefore choose a linear combination of taking the values at subset sizes . It is a separator for of degree at most . Hence .
Remark 4.2. The restriction on the characteristic cannot be dropped, and the smallest case shows the threshold moving: , one less than the value 6 that Theorem B gives over every field of characteristic different from two. Indeed, for five terms over the linear map , , has a null space of dimension at least two, and its nonzero elements are the incidence vectors of the zero sets. If and are two independent vectors of this null space, an affine separator with linear would have to satisfy , which is impossible; so no five-term sequence has an affine separator over . On the other hand has the affine separator , its only zero set being . Thus . The integral six-term identity supplied by Theorem B remains valid in characteristic two; the least length is simply smaller there.
The quotients , and
We now prove that every sequence of terms over has an integral -identity. Call a sequence short-zero-free if it has no nonempty zero set of size at most the exponent. For , if a sequence of length has a zero set with , deleting it leaves at least
terms. An identity of degree on the remainder therefore gives one of degree by Lemma 2.2(2). We call this step the short-zero recurrence. Once , it is automatic. The work is concentrated in the degrees before that point.
Identities of degree one and two are tested on affine and on quadratic polynomials, and we call them affine and quadratic identities. The following table summarises where the finite work lies; the computations behind the finite certificates are described in Appendix § 8, and sharpness is Proposition 4.1.
The case
The Fano configuration shows the mechanism in miniature. Degree zero is . For degree one, consider six terms. If a term is zero or repeated, there is a zero set of at most two positions; deleting it leaves at least four terms and hence a second, disjoint zero set. Lemma 2.2(1) applies.
Otherwise the six terms are distinct nonzero vectors of : the Fano plane with one point removed. Its lines are the triples , each of which has sum zero. Give weight 1 to the four lines not through , and weight to the complements, in the seven-point plane, of the three lines through . The latter are zero quadruples and do not contain . Each remaining point lies on two of the selected lines and in two of the selected quadruples. Its total weight is therefore , whereas the total weight of the seven sets is .
This identity uses overlap in an essential way. Every zero set among six distinct nonzero vectors has at least three terms. Two disjoint zero sets would therefore exhaust the six points, but their total is the omitted nonzero point , since all seven Fano points sum to zero.
Finally, any eight terms contain zero or a repetition, so . The short-zero recurrence gives all degrees at length . These are integral identities, valid after reduction in every characteristic; Proposition 4.1 gives equality in characteristic zero and in odd characteristic.
The case
We need the ten-term incidence identity and two elementary facts about supports.
Lemma 5.1 (Ternary base facts). The following assertions hold.
Every ten-term sequence over has an integral -identity.
The support of a short-zero-free sequence over has at most eight points. Every such support with seven points has nonzero sum, and every such support with eight points has sum zero.
We start by proving the second assertion.
Proof of Lemma 5.1(2). Let be the support of a short-zero-free sequence and put . Then is a cap, that is, it contains no three distinct collinear points (this is the connection with affine caps studied in [6]). Indeed, over , three distinct points are collinear exactly when their sum is zero. A line through would give an opposite pair in , and any other line would give a zero triple in .
An affine plane contains at most four cap points. Otherwise choose five. In each of the four directions, the three parallel lines would contain of these points, accounting for two pairs in that direction. This accounts for at most eight pairs, rather than .
Put and let be the number of affine planes meeting in exactly points, for . There are parallel classes of planes, each with three members, hence 39 planes. A point belongs to 13 planes, a pair to , and a noncollinear triple to one. Counting these incidences gives
For , these equations give , a contradiction. A larger cap would contain a ten-point cap, so and .
Suppose , so . The same equations give , hence some affine plane is disjoint from . Write it as , where is a nonzero linear functional. Since , one has . The other two parallel planes contain four points each, and therefore
Thus .
Finally suppose , so . No affine plane is empty, because the other two parallel planes could hold at most eight cap points. Thus , and the incidence equations give . For every nonzero linear functional , the three parallel slices have sizes or a permutation of . In either case their sizes are congruent modulo three, so . This holds for every , hence . □
Proof of Lemma 5.1(1), computer-assisted. If a ten-term sequence has a zero set of at most three positions, its deletion leaves at least seven terms and hence a disjoint zero set, since . Lemma 2.2(1) then gives the required affine identity.
It remains to consider the short-zero-free sequences. Their multiplicities are at most two, and their supports satisfy the cap conditions used above. The complete enumeration and its symmetry reductions are given in Appendix § 8. Each case is supplied either with two disjoint zero sets or with an explicit integral affine identity. The certificate check verifies the zero sums, total weight one, and zero total weight through each individual position. The transports between normalised representatives and the indexed sequences they represent are checked together with the certificates.
Call a set allowed if it contains neither zero nor a pair of opposite points, and no three distinct points of sum to zero. Write for the sequence containing two copies of each point of .
Lemma 5.2. Every short-zero-free sequence of thirteen terms over has a zero set of size at least eleven.
Proof. The plan is to write as a sum of at most two terms of : deleting them leaves a zero set of size at least eleven. The representation will come from the maximality of an allowed support, and the only point to watch is whether it uses a term more often than contains it.
We first observe that if is an inclusion-maximal allowed support, then every is a sum of at most two terms of . For take no terms, for one term, and if take two copies of . Otherwise is a nonzero point outside whose opposite is also outside , so adjoining to can only violate the third condition: there is a zero triple with distinct , which represents .
Now let be short-zero-free of length thirteen. Its multiplicities are at most two, so its support has seven or eight points, and there are three cases.
Case 1: the support is an eight-set . Then for three distinct points of multiplicity one. By Lemma 5.1, , so , and is maximal by the support bound. A representation of by at most two terms of is unavailable in , that is, it uses a term more often than contains it, only if it uses one of twice, say ; this gives , impossible in an allowed support.
Case 2: the support is a seven-set which is maximal. Then for one point of multiplicity one, and the only unavailable representation is , which forces , contradicting Lemma 5.1.
Case 3: the support is a seven-set which is not maximal. Again , and extends to an allowed eight-set , which is maximal, with and hence . A representation of by at most two terms of is unavailable in only if it uses , which does not contain, or twice, so it is one of
and these force , , and respectively. Each contradicts the support conditions.
Proof of Theorem B for . The target lengths are . Degree zero is ; degree one is Lemma 5.1(1).
For degree two, consider thirteen terms. A zero set of at most three positions leaves at least ten terms, so Lemma 2.2(2) applies to the affine identity on the remainder. Otherwise Lemma 5.2 gives a zero set of at least eleven positions. It properly contains a ten-term subsequence, and Lemma 2.2(3) raises that subsequence’s affine identity to degree two.
For degree three, a short zero set among sixteen terms again reduces to the preceding degree. If there is none, the support bound and the multiplicity bound force eight support points, each appearing twice. Their sum is zero by Lemma 5.1(2), so the whole sequence is a zero set. A thirteen-term subsequence properly inside it has a degree-two identity, which Lemma 2.2(3) raises to degree three.
The support and multiplicity bounds also show that every seventeen terms contain a short zero set: (in fact , see [9]). For , the target length is at least nineteen, so the short-zero recurrence proves all further degrees. Proposition 4.1 gives sharpness under its characteristic hypothesis.
The case
Lemma 5.3 (Quaternary finite input). Every fourteen-term sequence over has an integral -identity, and every eighteen-term sequence over has an integral -identity.
The proof is computer-assisted; its reductions are as follows. Since , a zero set of at most four positions among fourteen terms leaves a disjoint zero set. The two sets give an affine identity by Lemma 2.2(1). Each remaining fourteen-term case receives either a disjoint-pair witness or an explicit integral affine identity.
For eighteen terms, a short zero set reduces to the fourteen-term claim by Lemma 2.2(2). Three disjoint zero sets also suffice, by part (1). A zero set of at least fifteen positions gives a third route: it properly contains fourteen positions, whose affine identity lifts by part (3). The remaining cases are supplied with explicit integral quadratic identities.
Appendix § 8 gives the complete finite domains, the coverage arguments for two normalisations, and the counts of the positive witnesses. The checks verify the group sums and every required incidence moment over the integers. The enumeration includes multiplicities and accounts for sequences in proper subgroups; all transports preserve indexed occurrences.
The following example shows what the finite search can find beyond these length-based reductions. Take
in , with positions numbered in this order.
Inspecting the first, third and second coordinates shows that a nonempty zero set of at most four terms would require four equal terms, which are unavailable. Also . A zero set of at least fifteen terms would leave at most three terms with this sum. Their first coordinates force them to be 010 or 001, requiring three copies of 010 and one of 001, a contradiction. Three disjoint zero sets would have a zero union of at least fifteen terms, so they are impossible too.
There is nevertheless a useful five-term zero set
Deleting it leaves only thirteen terms, one fewer than the length at which an affine identity is guaranteed for every sequence. For this particular remainder, such an identity can be written down. The following five sets are zero sets disjoint from :
Position by position, they satisfy the integer identity
The coefficients add to one and cancel through every position, so they give an integral -identity on the remainder. Lemma 2.2(2) now produces an -identity on : keep the five weights, give their negatives to the unions of the five sets with , and give weight 1 to .
Thus the eleven-set quadratic certificate is a lower-degree identity lifted by one of the elementary constructions. The useful extra information is the affine identity on this particular thirteen-term remainder; its existence does not follow from the universal fourteen-term threshold. The certificate itself, in the form in which it is stored and checked, is reproduced in the supplement.
Proof of Theorem B for . The target lengths are . Degree zero is , and degrees one and two are Lemma 5.3. The bound follows from the classical result of Harborth [15], see [9 Theorem 2.2(2)]. For , the target length is at least 22; deleting a short zero set leaves at least terms. Lemma 2.2(2) therefore proves every further degree. Proposition 4.1 gives sharpness under its characteristic hypothesis.
Davenport constants
Proof of Corollary C. Theorems A and B give
for every finite abelian -group . Take , allowing trivial factors. Olson’s theorem gives
Since , the Chinese remainder theorem identifies with . Its invariant factors are in divisibility order, so substitution yields
The basis lower bound gives the reverse inequality.
The general bound (6) holds for kernels of every rank and for every prime. Coprimality and the rank bound are used in the comparison with above.
For example, take and . Here , so the quotient identity is used in degree 28, at length . Transfer gives
No enumeration at length 91 is needed: the constructions have already extended the finite base cases to degree 28. The basis sequence of length shows why the resulting value is sharp.
Verification, reproducibility and dependencies
The proofs of Lemma 5.1(1) and Lemma 5.3 are computer assisted. The support statement in Lemma 5.1(2) follows from affine-plane counting, and the proof of Lemma 5.2 uses it. The later incidence degrees are deductions from the finite base identities by the three constructions. Table 1 records these dependencies. Olson’s small -group values for the Davenport constant , and also supply degree zero and the short-zero reductions in the relevant rows.
| Quotient | Target length | Before the automatic recurrence | Automatic recurrence |
| : ; : the Fano configuration | , using | ||
| : ; : finite affine certificates; : the support facts and the constructions | , using | ||
| : ; : finite certificates | , using |
Table 1.
For each computed case, the verification checks a positive witness: disjoint zero sets, a suitable long zero set, or an integer incidence identity. An incorrect negative classification, such as failing to detect a disjoint pair, cannot by itself certify a case; a valid positive witness must still pass the check. Appendix § 8 specifies the full finite domains and justifies the reductions used to enumerate them. The certificate verification uses exact arithmetic and does not rely on the optimisation software used to discover the weights. All certificates and the instructions for reproducing their results are collected in a separate Supplement available on our Github repository [18].
The definitions, the lemmas, Proposition 4.1, Theorems A and B and Corollary C have also been formalised in Lean 4 with Mathlib [18], with proofs that follow the ones given here. In developing the formalisation, we made use of Scott Armstrong’s LeanAutomatizationSkills collection [2]. For Lemma 5.1(1) and Lemma 5.3, the formalisation performs its own enumerations; their completeness is proved in Lean, together with the soundness of the certificate checks (see the supplement).
| Statement | Nature | Essential inputs |
| Lemma 2.1 (duality) | ordinary | Boolean multilinearisation; linear algebra over a field |
| Lemma 2.2 (three constructions) | ordinary | Lemma 2.1(1); inclusion–exclusion; two degree-cancellation arguments |
| Lemma 3.1 (indicator) | ordinary | truncated group algebra; nilpotence of the augmentation ideal |
| Theorem A | ordinary | evaluation identity; Lemma 3.1; Olson’s value of |
| Proposition 4.1 (sharpness) | ordinary | integer triangular factorisation of the binomial-value matrix |
| Theorem B, | ordinary | Fano identity; ; Lemma 2.2; Proposition 4.1 |
| Lemma 5.1(1) | computer-assisted | complete ten-term coverage (Appendix § 8); positive affine or disjoint-pair certificates and their transports |
| Lemma 5.1(2) | ordinary | cap–plane incidence counts; linear functionals |
| Lemma 5.2 | ordinary | maximal allowed supports; Lemma 5.1(2); multiplicity checks |
| Theorem B, | deduction using finite input | Lemmas 5.1, 5.2, 2.2; ; ; Proposition 4.1 |
| Lemma 5.3 | computer-assisted | two quaternary enumerations (Table 3); exact certificates; the proper-subgroup bound used by the support-basis enumeration |
| Theorem B, | deduction using finite input | Lemmas 5.3, 2.2; ; ; Proposition 4.1 |
| Corollary C | deduction using finite input | Theorems A and B; Olson’s theorem; Chinese remainder theorem; basis lower bound |
Table 1 (PDF p. 14). Dependency map. “Ordinary” refers to the proof and its stated inputs; “computer-assisted” means an exhaustive enumeration with positive witnesses checked by an independent program; “deduction” means an ordinary argument some of whose inputs are computer-assisted statements.
| Quantity | Length 14 | Length 18 |
| Generated by projection and lifts | 308,051 | 18,224 |
| Generated with a fixed support basis | 24,445,377 | 1,267,272 |
| Distinct linear orbits | 106,655 | 4,941 |
| Disjoint-zero witnesses | 284,994 | 13,549 |
| Long-zero witnesses (after disjoint cases) | — | 4,171 |
| Direct integer certificates | 23,057 | 504 |
| Orbits receiving direct certificates | 6,963 | 64 |
Table 3. Quaternary finite coverage. Generated cases may repeat orbits; orbit counts refer to the full linear group.
Computational coverage and certificates
This appendix collects the parts of the proofs of Lemma 5.1(1) and Lemma 5.3 that justify the finite computations: the exact domains that were enumerated and the arguments showing that these domains cover every required sequence. Nothing is inferred from samples: each enumeration establishes a universal statement about all sequences of the stated length. The positive witnesses themselves, their statistics, the stored form of the eighteen-term example of Section 5 and the instructions for an independent replay are collected in the supplement.
Ternary coverage
A short-zero-free sequence over has all multiplicities at most two, and its support contains no zero, no pair of opposite points and no three distinct points with sum zero; equivalently, is an affine cap in , a connection developed in [6]. A support with at least five points spans , since a plane contains only four pairs of opposite nonzero points. Mapping three independent support points to the standard basis leaves ten pairs of opposite points, each contributing none or exactly one of its two points, and the complete enumeration of these choices produces the support counts of Table 2; no support of size nine or more occurs. All 296 seven-point supports have nonzero sum and all 18 eight-point supports have sum zero. A separate enumeration by direct basis completion confirms these two facts. These are the three facts of Lemma 5.1(2); they are proved in Section 5 by counting affine planes, and the enumeration is a redundant check of them.
| Support size | Supports | Cases | Disjoint pair | Other identities |
| 5 | 108 | 108 | 81 | 27 |
| 6 | 296 | 4,440 | 2,769 | 1,671 |
| 7 | 296 | 10,360 | 7,052 | 3,308 |
| 8 | 18 | 504 | 504 | 0 |
| Total | 15,412 | 10,406 | 5,006 |
Table 2. The ternary ten-term computation, after normalising a support basis. Support counts are not orbit counts.
For Lemma 5.1(1), a zero set of size at most three leaves seven terms and hence a disjoint zero set, since . It remains to treat the short-zero-free multiplicity vectors of total ten; their exact counts are given in Table 2. Most of them contain two disjoint zero sets. The remaining 5 006 cases form 43 orbits under the linear group, and each orbit representative receives an explicit integral identity with 5, 7 or 9 zero sets and coefficients . An independent checker regenerates the complete case space and verifies every zero set and every moment of every identity, and a second one checks all transport matrices and index permutations between representatives and cases. No affine translations are used anywhere.
Quaternary coverage by reduction modulo two
Represent as and reduce modulo two. In a short-zero-free sequence, the fibre over contains at most three terms, which must be distinct and independent in . A nonzero fibre contains at most four distinct elements, since its eight elements form four pairs of opposite elements, and at most one of these elements can repeat, because two distinct repeated elements in the same fibre satisfy . Each multiplicity is at most three, so a fibre carries at most six terms.
We enumerate every projected multiplicity vector of the prescribed total with and otherwise, up to the 168 elements of . Every such linear map lifts to an automorphism of , because a binary lift has odd determinant. Write a lifted element as with , where is the binary representative. The automorphisms that reduce to the identity modulo two, and their action on high parts, are
so choosing one term in each of the first independent occupied nonzero fibres, a suitable makes each chosen high part zero. Since zero is the least high part, requiring the first high part in each of these fibres to vanish, after sorting high parts within fibres, retains a representative of every orbit. This enumerator does not assume that the projection spans, so it covers sequences contained in proper subgroups as well.
Backtracking then generates all lifts in nondecreasing order within each fibre. It maintains the sets of sums of one, two and three previously placed terms with distinct indices. Appending is legal exactly when and , and the update is
with the old sets on the right. This is an exact test for short-zero-freeness that keeps track of repetitions.
Independent support-basis enumeration
A second enumeration normalises a support basis instead. An exhaustive computation in the subgroup shows that short-zero-free sequences there have length at most thirteen. Every proper subgroup of lies in an index-two subgroup, all index-two subgroups are linearly equivalent, and therefore every short-zero-free sequence of length at least fourteen spans modulo two; three of its support vectors then have odd determinant and form a basis of . We generate all short-zero-free multisets containing the standard basis and canonicalise each by the least sorted image under the maps that send an ordered support basis to the standard basis; this is a complete invariant for linear equivalence. The two enumerations produce identical canonical orbit lists.
References
- [1]N. Alon, Combinatorial Nullstellensatz, Combin. Probab. Comput. 8 (1999), 7–29. doi:10.1017/S0963548398003411.DOI
- [2]S. Armstrong, LeanAutoformalizationSkills: Portable Lean 4 autoformalization skills for Codex and Claude Code, GitHub repository (2026). github.com/scott narmstrong/LeanAutoformalizationSkills.
- [3]G. Bhowmik, I. Halupczok, and J.-C. Schlage-Puchta, Inductive methods and zero-sum free sequences, Integers 9 (2009), A40, 515–536. doi:10.1515/INTEG.2009.042; arXiv:0711.1262.DOI
- [4]G. Bhowmik and J.-C. Schlage-Puchta, Davenport’s constant for groups of the form ℤ₃ ⊕ ℤ₃ ⊕ ℤ₃d, in: Additive Combinatorics, CRM Proc. Lecture Notes 43, Amer. Math. Soc., Providence, RI, 2007, pp. 307–326. arXiv:math/0610416.
- [5]C. Delorme, O. Ordaz, and D. Quiroz, Some remarks on Davenport constant, Discrete Math. 237 (2001), 119–128.DOI
- [6]Y. Edel, C. Elsholtz, A. Geroldinger, S. Kubertin, and L. Rackham, Zero-sum problems in finite abelian groups and affine caps, Q. J. Math. 58 (2007), 159–186. doi:10.1093/qmath/ham003.DOI
- [7]W. Gao, On Davenport’s constant of finite abelian groups with rank three, Discrete Math. 222 (2000), 111–124. doi:10.1016/S0012-365X(00)00010-8.DOI
- [8]W. Gao and A. Geroldinger, Zero-sum problems in finite abelian groups: a survey, Expo. Math. 24 (2006), 337–369. doi:10.1016/j.exmath.2006.07.002.DOI
- [9]W. D. Gao, Q. H. Hou, W. A. Schmid, and R. Thangadurai, On short zero-sum subsequences. II, Integers 7 (2007), Paper A21. doi:10.5281/zenodo.8281159.DOI
- [10]A. Geroldinger, D. J. Grynkiewicz, and Q. Zhong, Combinatorial Factorization Theory. Factorization Theory of Krull Monoids via Additive Combinatorics, Math. Surveys Monogr. 296, Amer. Math. Soc., Providence, RI, 2026. doi:10.1090/surv/296.
- [11]A. Geroldinger and R. Schneider, On Davenport’s constant, J. Combin. Theory Ser. A 61 (1992), 147–152. doi:10.1016/0097-3165(92)90061-X.DOI
- [12]A. Geroldinger and W. Yang, On a classical zero-sum invariant, preprint (2026). arXiv:2608.19090v1.arxiv.org/abs/2608.19090
- [13]B. Girard, An asymptotically tight bound for the Davenport constant, J. Éc. polytech. Math. 5 (2018), 605–611. doi:10.5802/jep.79.DOI
- [14]B. Girard and W. A. Schmid, Direct zero-sum problems for certain groups of rank three, J. Number Theory 197 (2019), 297–316. doi:10.1016/j.jnt.2018.08.016.DOI
- [15]H. Harborth, Ein Extremalproblem für Gitterpunkte, J. Reine Angew. Math. 262/263 (1973), 356–360. doi:10.1515/crll.1973.262-263.356.DOI
- [16]J. E. Olson, A combinatorial problem on finite Abelian groups. I, J. Number Theory 1 (1969), 8–10. doi:10.1016/0022-314X(69)90021-3.DOI
- [17]J. E. Olson, A combinatorial problem on finite Abelian groups. II, J. Number Theory 1 (1969), 195–199. doi:10.1016/0022-314X(69)90037-7.DOI
- [18]C. G. Pop, G. C. Rad, and G. C. Turcaș, Davenport_Project: a Lean 4 formalization of “Integral zero-sum incidence identities and Davenport constants in rank three”, 2026. https://github.com/CiprianRad/Davenport_Project.
- [19]W. A. Schmid, The inverse problem associated to the Davenport constant for C₂⊕C₂⊕C₂ₙ, and applications to the arithmetical characterization of class groups, Electron. J. Combin. 18(1) (2011), Paper P33. doi:10.37236/520.DOI
- [20]P. van Emde Boas, A combinatorial problem on finite abelian groups. II, Report ZW 7/69, Mathematisch Centrum, Amsterdam, 1969. CWI institutional repository.
- [21]P. van Emde Boas and D. Kruyswijk, A combinatorial problem on finite abelian groups. III, Report ZW 8/69, Mathematisch Centrum, Amsterdam, 1969. CWI institutional repository.
- [22]A. Volkmann, The small Davenport constant of H₂₇×C₃ʳ, preprint (2026). arXiv:2609.19810.
- [23]A. Volkmann, A twelve-term exclusion for the small Davenport constant of E₂×C₃ʳ, preprint (2026). arXiv:2609.22972.
- [24]A. Volkmann, The small Davenport constant of E₂ × C₃ʳ for 0 ≤ r ≤ 3, preprint (2026). arXiv:2609.37262.