Fabian Arevalo

Computational research note · Design theory

Signed difference sets: closing two-thirds of an open database

Two classical nonexistence theorems and a validated exhaustive search close 45,370 of the 67,823 open cells in D. M. Gordon's 2023 signed difference set database, including ten new signed difference sets with no published source; an independent audit finds 147 of the database's own 280 example witnesses fail its own defining equation. A second session independently verifies an external researcher's closure of the last small open family (order 32 and 36) leg by leg, and adds a database-free proof for order 36.
August 9 & 12, 2026 · produced in Claude Code research sessions, merged and published the same day.
Code, data & certificates in the companion repository.

Start here: no maths required

Arrange twenty pegholes in a circle, like the hours on a strange clock. Drop a plus-marked peg, a minus-marked peg, or nothing into each hole. Now picture a second, identical ring of holes spinning underneath the first. At every hole, multiply the two marks lined up there (a peg times a peg, or zero if either hole is empty) and add up all twenty products. A signed difference set is a marking with an odd kind of regularity: spin the ring by any amount other than a full turn, and that sum comes out to exactly the same number every time.

Most markings do not have this property; the ones that do are rare, and get rarer to find as the circle grows or the rule is asked to hold in more exotic, non-circular versions (the finite abelian groups of the technical sections). Deciding whether a valid marking exists at all, for a given circle size, peg count and target sum, is a genuine search problem.

In 2023 the mathematician D. M. Gordon defined this question in full generality and built a reference database recording, for 70,543 different combinations of size, peg count and target sum, whether a marking is known to exist, known not to exist, or undecided. The large majority, 67,823 combinations, were simply Open: nobody, including Gordon's own exhaustive searches, had gotten around to most of the non-circular versions at all.

Hover or tap a segment.
Where the 67,823 open combinations went in one session. Two classical theorems, applied by computer to every cell (section 3), instantly rule out 45,328 of them. A validated exhaustive-search program decides 42 more that the theorems cannot reach (section 5). The remaining 22,453 are still open. (2,720 combinations were already decided, either way, before this session; that count does not change here.)

What this note adds

Gordon's exhaustive searches covered only the plain circular case (technically: cyclic groups), and not even all of those; the non-circular cases were open wholesale, and nobody appears to have applied classical number-theory shortcuts to the database at all.

  1. Two old shortcuts, applied to the whole database at once. Two classical number-theory arguments (one of them a century-old idea about character sums, the other the ordinary fact that certain differences must be perfect squares) turn out to transfer to this exact setting. Applied by computer to all 70,543 combinations in about a second, they instantly rule out 45,328 of the 67,823 open combinations, each with a one-line, hand-checkable reason. This is a proof, not a computation: the theorems themselves, and their application to each combination, are checkable by hand.
  2. A validated search program decided 58 more, by brute force. The program was checked against every combination whose answer is already known in the database (42 combinations, zero disagreements) and against a second, independently written program (exact agreement on 8 combinations). It found 10 new signed difference sets that appear in no published source, including one where the counts of plus- and minus-marks differ by only one, in a 5×5 non-circular clock, and 48 more combinations proved empty. Net effect: the open shelf drops from 67,823 to 22,453 in this one session.
  3. An audit of the database's own worked examples. A from-scratch checker was run against every one of the 280 example markings the database ships as evidence. 147 of them do not satisfy the database's own defining rule. The likely cause, traced on the smallest broken example: an export bug that swapped some plus-marks and minus-marks. 22 of the 147 broken examples could be repaired by undoing one or two such swaps, each repair independently re-checked.
  4. A pattern, not yet a theorem. Among the seven non-circular clocks of size 32, one particular peg-count-and-target combination has a valid marking in exactly the two clocks that happen to contain a smaller 4-by-4 clock inside them, and in none of the other five. Seven data points is not a proof; it is recorded as an open question (section 7).
  5. Session 2: a stranger's proof, checked leg by leg, plus one new fact. Three days later another researcher, Nicolas Masselot, published a closure of exactly the small clocks this note's search couldn't reach: the 20-peg, 4-sum family on the 32-hole clocks, and the 29-peg, 4-sum family on the 36-hole clocks. This session re-derived every one of his existence and nonexistence claims from scratch, with independently written code that shares nothing with his: all sixteen constructions check out, and all four "impossible" proofs reproduce exactly, including a 17-million-marking search that replaces his computer-checked proof certificate with a second, from-scratch one. One new fact came out of reading his method closely: the 36-hole impossibility can be proved without touching the database at all, from one small, complete calculation (section 8).

Everything below is the technical version. Section 3 has the two theorems and their bulk application; section 5 has the ten new examples; section 6 has the audit; section 8 has session 2.

AbstractSummary of results

A signed difference set SDS(v,k,λ) in a finite abelian group G of order v is an element A of the group ring Z[G] with coefficients in {−1,0,+1}, exactly k of them nonzero, satisfying A·A(−1) = k·e + λ(G−e). D. M. Gordon introduced the notion (Des. Codes Cryptogr. 91 (2023) 2107–2115, arXiv:2212.10630) with a companion database of 70,543 parameter cells, of which 67,823 were marked Open in the snapshot fetched 2026-08-09 (sha256 39bab9fc…ca85). We prove two nonexistence criteria for SDS — verbatim transfers of the classical even-order square condition of symmetric-design theory and of Turyn's self-conjugacy argument — and apply them to every cell: 45,328 of the 67,823 open cells are closed (23,997 by the first, 21,331 by the second), each with a one-line checkable certificate. Independently, an exhaustive-search engine, validated by reproducing the entire decided database at v ≤ 24 (42 cells, zero discrepancies) and by exact witness-list agreement with a second implementation, decides 58 more open cells (every nonexistence consistent with the criteria, zero conflicts): 48 empty, and 10 containing signed difference sets that appear in no published source, including a λ = 1 set in Z5×Z5 and cells where, at fixed (v,k,λ), the group structure alone decides existence. In total the open shelf drops from 67,823 to 22,453. Finally, an audit with an independent checker found that 147 of the 280 witness sets stored in the database fail the defining equation (21 of 144 witness-bearing cells); ≤2-swap repair recovers independently re-verified witnesses for 12 of the 21 affected cells. Session 2 (2026-08-12) reviews an external closure of the family this left open: N. Masselot's census decides all remaining Open cells at v ≤ 36, and every leg of both directions (sixteen existence witnesses, four independent nonexistence searches) is re-derived here with review-owned code sharing no lineage with his — zero conflicts. One byproduct: the (36,29,4) nonexistence, which his statement imports from the database's cyclic exhaust, is shown here to follow instead from a single self-contained computation (the C18 quotient system is empty), removing that dependency.

Section 1Setting: Gordon's database, and the definitions

Let G be a finite abelian group of order v, written additively, and let Z[G] be its integral group ring with basis {xg : g ∈ G} and the involution A ↦ A(−1) induced by g ↦ −g. A signed difference set SDS(v,k,λ) in G is A = Σg A(g)·xg with A(g) ∈ {−1,0,+1}, #{g : A(g) ≠ 0} = k, and

A · A(−1)  =  k·e  +  λ·(G − e)
e = x0, and G also denotes Σg xg. Equivalently: for every d ≠ 0, the signed autocorrelation Σg A(g)A(g−d) equals λ.

Write P = {g : A(g) = +1}, M = {g : A(g) = −1}, σ = |P| − |M|. Difference sets are the case M = ∅; circulant weighing matrices are cyclic G with λ = 0. This is exactly the convention of Gordon's reference checker is_sds (his sds_code.py, shipped in the data directory), against which an independent checker built for this session agrees on all 133 valid stored witnesses.

Lemma 1 · character identities

Let A be an SDS(v,k,λ) in G and χ a character of G. Then |χ(A)|2 = k − λ for χ nontrivial, and σ2 = k + λ(v−1) for the trivial character. In particular s := √(k+λ(v−1)) ∈ Z≥0 with s ≡ k (mod 2), |P| = (k+s)/2 and |M| = (k−s)/2 after the global sign flip A ↦ −A (which preserves the defining equation) normalizing σ ≥ 0.

Proof. Extend χ linearly to Z[G]. Since χ(A(−1)) = Σ A(g)χ(−g) = conj(χ(A)) and χ(G) = 0 for nontrivial χ (column orthogonality), applying χ to the defining equation gives |χ(A)|2 = k − λ; the trivial character gives σ2 = k + λ(v−1). The parity claim: σ ≡ |P| + |M| = k (mod 2). ∎

Section 2Two nonexistence criteria

Both criteria are classical machinery from symmetric-design and difference-set theory; what is demonstrably new is their systematic application to this database (see the novelty remark below).

Theorem 1 · even order, the classical even-v square condition

If v is even and k − λ is not a perfect square, then no SDS(v,k,λ) exists in any abelian group of order v.

Proof. An abelian group of even order has a subgroup of index 2, hence a character χ of order 2, whose values are ±1. Then χ(A) ∈ Z and Lemma 1 gives k − λ = χ(A)2. ∎

Theorem 2 · self-conjugacy, Turyn's argument

Let m > 2 divide exp(G), and let p be a prime with p ∤ m such that pj ≡ −1 (mod m) for some j ≥ 1. If vp(k−λ) is odd, then no SDS(v,k,λ) exists in G.

Proof. Take χ of order exactly m (it exists because m | exp(G)) and α = χ(A) ∈ Z[ζm], so α·conj(α) = n := k−λ by Lemma 1. The Galois group of Q(ζm)/Q is (Z/m)×, and the decomposition group of any prime ideal ℙ above p is ⟨p mod m⟩, which by hypothesis contains −1, i.e. complex conjugation; hence ℙ̄ = ℙ. Since p ∤ m, p is unramified, so vℙ(n) = vp(n) is odd, while vℙ(α·conj(α)) = vℙ(α) + vℙ̄(α) = 2·vℙ(α) is even — a contradiction. ∎

The hypothesis p ∤ m is essential: for p | m, ramification defeats the argument, and Gauss sums realize |α|2 = p in Z[ζp].

On novelty. Theorem 1 is the signed analogue of the standard “v even ⇒ k−λ square” condition for symmetric designs and abelian difference sets, and Theorem 2 is Turyn's self-conjugacy test, standard in difference-set and weighing-matrix nonexistence. Their transfer to signed difference sets is immediate from Lemma 1 (the {−1,0,+1} coefficients play no role beyond χ(A) lying in Z[ζm]). Resolved 2026-08-12 (session 2, with web access): Gordon's paper and He–Chen–Ge's were read in full. Neither states either transfer, so the possible-rediscovery hedge on the two criteria is lifted; they remain classical in substance, but their statement for signed difference sets appears to be new to this line. Lemma 1's trivial-character identity is Gordon's own Lemma 1.1; its nontrivial-character identity is He–Chen–Ge's Lemma 2.2 — both are now checked against the papers directly, not inferred from abstracts. What is demonstrably new regardless is the systematic application to the database: cells as small as SDS(18,15,2,[3,6]) (k−λ = 13, v even) were Open in the 2026-08-09 snapshot, so the database had not absorbed even Theorem 1. Lemma 1's trivial-character part reproduces, as the special case λ = 0, the classical fact that the weight of a circulant weighing matrix is a perfect square.

Section 3Bulk closure of the database

A script applies the admissibility test T0 (s ∈ Z, s ≡ k mod 2, s ≤ k, k ≤ v, all following from Lemma 1) and Theorems 1 (T1) and 2 (T2) to each of the 70,543 cells. Results on the 2026-08-09 snapshot (see the figure above for where these numbers land in the database as a whole):

ShelfCount
Open cells closed by T123,997
Open cells closed by T2 (and not T1)21,331
Open cells closed, total45,328 of 67,823
Yes/All cells violating any criterion0 (146 cells checked)
Exhaust decisions conflicting with criteria0
Gordon's exhaust-No cells the criteria would also close984 of 2,574

Each closure is recorded as (cell, criterion, m, p, vp(k−λ)) — a certificate checkable by hand or by rerunning the closure script (1.1 s for the 45,328 open-cell closures; the 984-cell retro-coverage figure in the last row is reproducible from the same test applied to the database's No cells, independently confirmed for this page, though the committed script currently only persists the Open-cell pass). The zero-violation rows are the soundness controls: the criteria never contradict a cell known to contain an SDS, and never contradict an exhaust decision.

Section 4The exhaustive-search engine

A C engine performs a depth-first search over assignments A: G → {−1,0,+1} in a fixed element order, with |P| and |M| forced by Lemma 1, exact incremental correlation sums, and an interval prune (each unresolved pair can move a correlation by at most 1). One reduction is applied in production mode:

Lemma 2 · translation reduction

For p ∈ P, the translate A′(g) = A(g+p) is an SDS with the same parameters and 0 ∈ P′. Hence restricting the search to A(0) = +1 preserves existence and nonexistence. (|P| ≥ 1 holds in every admissible cell since s ≤ k.) ∎

Validation, all completed before any new claim was accepted:

Section 5New decisions of open cells

58 previously-open cells were decided (10 EXIST, 48 NONEXIST; 42 beyond the criteria's reach, 16 also covered by them), at wall times from under a second to 31 minutes per cell on 4 cores. Every EXIST row ships an explicit witness verified by the independent checker; every NONEXIST row is an exhaust with a node count, plus, where the criteria also apply, an independent proof.

New signed difference sets. Representative witnesses, elements given as coordinates in the invariant-factor groups:

Group structure decides existence. Two exhibits from the session's table, both independently re-verified against the committed data for this page:

Hover or tap a cell.
Existence at fixed parameters, across every group of a given order. Filled cells contain a signed difference set; open cells were exhausted empty; dash-marked cells are Gordon's own prior "No", not decided this session. First row: (32,28,12) across all seven abelian groups of order 32 — existence lands in exactly the two groups containing a Z4×Z4 subgroup (seven data points; an observation, not a theorem). Second row (added in session 2, Masselot's census, verified here — section 8): (32,20,4) across the same seven groups — existence lands in exactly the six noncyclic ones, a clean cyclic/noncyclic split rather than a subgroup condition. The two rows together say the same order can be governed by two different mechanisms at two different parameters. Bottom grid: every parameter triple at order 27, across its three groups — the cyclic group Z27 carries nothing anywhere on the shelf, while the two non-cyclic groups carry signed difference sets at three of the ten triples.

At order 27 the cyclic group is an obstruction, not an aid, across the whole shelf; SDS(36,11,2) exists exactly when the 3-Sylow subgroup is non-cyclic (Z6×Z6 and Z3×Z12 both have 3-Sylow Z3×Z3; Z2×Z18 and Z36 both have cyclic 3-Sylow Z9) — the same direction as the order-27 pattern. A 94-second node-limit probe of the surviving (32,20,4) family (all seven groups genuinely open at the time: verified against the snapshot for this page, though the probe's own throughput numbers are not backed by a committed data row) put a complete exhaust at dozens of core-hours per group, out of this session's range. That family did not stay open: section 8 covers its closure three days later, verified here leg by leg.

Section 6The witness audit

An independent checker was run over every witness set stored in the database: 133 of 280 verify; 147 fail, concentrated in 21 of the 144 witness-bearing cells (all cyclic; the other 123 cells, including every Paley and He–Chen–Ge cell, verify perfectly, which pins the convention). No symmetry of the defining equation (translation, group automorphism, inversion, global sign flip — all of which preserve the multiset of off-peak correlations) can repair a set whose correlation profile is non-constant, so the 147 are invalid as stored, not differently encoded.

Hover or tap a segment.
The 280 example markings the database ships as evidence. 147 fail the database's own defining equation. Forensics on the smallest affected cell, SDS(20,11,2) in Z20, found the cause: a complete re-enumeration (5,426,081 nodes, 1.3 s) shows the cell genuinely has exactly 40 valid markings, and the four stored ones are true markings with a plus-mark and a minus-mark exchanged — an export-stage defect, not a wrong answer. Undoing such swaps (≤2 per set) repairs 22 of the 147, independently re-verified.

The repair mechanism generalizes: searching the ≤2-swap neighborhood of every invalid stored set repairs 22 of the 147, recovering at least one valid witness for 12 of the 21 affected cells (every stored set of SDS(20,11,2,[20]) and SDS(35,21,10,[35]), all four of SDS(247,127,63,[247]), both of SDS(499,250,123,[499]), and one set each of eight more cells up to v = 499). The Yes/All statuses of those 12 cells are thereby confirmed constructively, independent of Gordon's original exhausts. The remaining 9 cells — SDS(51,19,3), SDS(78,53,28), SDS(104,29,4), SDS(111,66,17), SDS(181,136,11), SDS(182,38,2), SDS(182,101,20), SDS(200,151,102), SDS(277,208,17) — hold only sets more than two swaps from any valid signed difference set; their statuses rest on Gordon's original exhaust claims alone. A draft report for Gordon covering all three findings above is prepared in the companion repository, to be sent from a machine with outbound access.

Section 7Open questions

  1. Do sporadic signed difference sets live in the smallest surviving cells? The (32,20,4) family, across all seven abelian groups of order 32, is the natural next compute target. A short timing probe put a complete exhaust at roughly 48 core-hours per group by extrapolation from a 94-second node-limited run; that specific throughput figure is reported by the research session's own notes but has no surviving row in the committed run log, so it is stated here as reported, not independently re-certified — the qualitative fact that all seven cells remain genuinely open is independently confirmed against the database snapshot for this page. Settled by Masselot (v1.0, 2026-08-12), verified here in section 8: exists iff the group is noncyclic. The method lesson stands regardless of who closed it: a quotient ladder (C8→C16→C32) turns the flat 48-core-hour estimate above into 6.7 measured seconds, because most of the correlation constraint already bites at the smallest quotient. The successor question is where else that ladder pays off — the smallest cells still Open past order 36.
  2. Gordon's equivalence convention for "All" cells (how his set lists are normalized) is not pinned here; complete enumerations in this session report labeled counts and translation classes instead. Pinning it matters before contributing complete enumerations upstream.
  3. Do stronger classical tests (Mann-type, multiplier theorems, field-descent) transfer as cleanly and close more of the 22,453 surviving open cells? The 2-adic behavior at ramified primes, the Gauss-sum escape hatch in Theorem 2, is where signed sets genuinely differ from ordinary difference sets.
  4. Is there a uniform construction behind the non-cyclic (27,10,1) / (27,14,5) / (27,17,8) family, beyond the case-by-case exhaust witnesses reported here?
  5. Does the Z4×Z4 pattern at (32,28,12) generalize — is there a transfer or projection argument through a Z4×Z4 quotient or subgroup that explains both the two existences and the five nonexistences?

Section 8 · Session 2, 2026-08-12Reviewing an external closure, and one new fact

Three days after the session above, Nicolas Masselot released Certified census of small signed difference sets v1.0, deciding all 68 Open cells at v ≤ 36 in the same frozen database snapshot: the 58 this note already decided, with zero conflicts, plus the 10 it had left open — signed (32,20,4) difference sets exist in an abelian group of order 32 iff the group is noncyclic, and no abelian group of order 36 admits a signed (36,29,4) difference set. At his request, this repository reviewed the note. What follows is a review, not a rediscovery: every number below was re-derived by code written from the definitions alone, sharing no lineage with his pipeline or with Gordon's. The constructions are his; this session verified them.

Claim in Masselot's noteIndependently re-derived hereAgree
16 existence witnesses (6 of them the order-32 constructions)16/16 pass this repository's independent checker (§1); the six order-32 file hashes match his table byte for byteyes
C8 quotient: 9,528 candidates after sum+norm, 56 solutions, 5 affine orbitssame counts; orbit sizes [8,8,8,16,16] reproduced independentlyyes
C16 quotient: 12 survivors over orbit representatives12 — and, re-run with no orbit reduction anywhere: 144 survivors over all 56 parentsyes
C32 (cyclic group): refines to 0 solutions2,985,984 ternary refinements tried, 0 surviveyes
C2×C18 quotient: 144 candidates, 0 solutions144 on the canonical pair; 0 over a complete enumeration of every marginal pair, no normalization usedyes
C3×C12 quotient: 420 candidates, 0 solutions420 on the canonical pair; 0 over every marginal pairyes
C6×C6: SAT/DRAT-certified 0 solutions16,964,640 marginal-consistent ternary vectors, complete integer search, 0 survive (43.7 s) — a solver-free second proof, not an audit of the DRAT chainyes
Census: 68 Open cells at v ≤ 36; 58 replicate this repository, 10 novelindependent 2026-08-09 fetch of the same commit has the same 68 cells; verdict-by-verdict diff: 58 agreements, 0 disagreements; the 10 remaining are exactly this repository's 10 undecided onesyes

Full table and methodology: masselot-review/REVIEW.md in the companion repository. Total independent verification cost: under one minute of laptop CPU.

This closes open question 1 above. Six of the seven order-32 groups need no search at all — Masselot's constructions exist by direct witness, checked here against the same independent checker used throughout this note. Only the seventh, cyclic Z32, needs a nonexistence proof, and that is where a quotient ladder does the work a flat search could not.

Hover or tap a stage.
The quotient ladder that proves SDS(32,20,4) impossible in the cyclic group of order 32. Each stage refines the survivors of the last into a bigger quotient group; nearly everything is killed at the cheapest stage, C8, because the correlation identity already binds hard there. Measured on this session's independent code, sharing no lineage with Masselot's: 6.7 s total, against roughly 48 core-hours for the flat depth-first search this note's first session had priced the identical Z32 proof at (§5).
Lemma 3 · quotient projection

Let K ≤ G with |K| = m and Q = G/K. Projecting an SDS(v,k,λ) A ∈ Z[G] to cell sums over the fibers of G → Q gives B ∈ Z[Q], coefficients in [−m,m], satisfying B·B(−1) = (k−λ)·e + λm·Q in Z[Q]; equivalently Σq bq2 = k + (m−1)λ, and every nonzero shift of the Q-autocorrelation equals mλ.

Proof. The projection Z[G] → Z[Q] is a ring homomorphism sending e ↦ e and G ↦ m·Q (G here also denoting the group sum, as in §1); apply it to the defining equation (1). ∎

The moment identities are Gordon's own Lemma 5.2; the full convolution form used here is the engine of Masselot's note.

Observation (C18 emptiness). For (v,k,λ) = (36,29,4) and m = 2, the C18 system of Lemma 3 — coefficients in [−2,2], Σb = 13, Σb2 = 33, every nonzero shift equal to 8 — has no solutions (complete enumeration, 0.6 s; re-verified directly against the committed run output for this page). Hence no SDS(36,29,4) exists in any abelian group of order 36 with a C18 quotient — that is, C36 and C2×C18 at once. CERTIFIED. Masselot's own version of the order-36 result imports the cyclic case from Gordon's database exhaust; this removes that dependency, so "no abelian group of order 36 admits a signed (36,29,4) difference set" now rests entirely on computations independent of the database — worth doing given §6, where 147 of 280 of the database's own stored witnesses turned out to be corrupted on export.

What this review does and does not cover