Computational research note · Arithmetic Ramsey theory
Take the whole numbers 1, 2, 3, …, m and paint each one with one of three colors. The rule to avoid, in its simplest form: no color has three numbers x, y, z of that color with x + y = z (x and y are allowed to be equal). For small m this is easy to dodge. Past one exact m it becomes impossible, however cleverly you color — that threshold is a Schur number.
This note studies a family of harder versions of the same game: each of the three colors gets its own equation length. Color i must avoid a monochromatic solution where ti numbers of its color — repeats allowed — have the first ti − 1 of them adding up to the last. (The everyday Schur number is every ti = 3.) The entire published table of these thresholds, every combination of lengths anyone had pinned down exactly, was 26 numbers, and it had not grown since 2016.
A modern SAT solver — the kind of automated reasoning tool that, in 2017, finally settled the notorious "Schur number five" after decades of not knowing — certifies eleven more of these thresholds here, each with a proof a separate program can check from scratch on both sides of the boundary. Four of them answer open cases of a specific 2016 conjecture about what formula these numbers should follow, and all four confirm it. Seven map a smaller family with no conjectured formula at all — and inside that unmapped territory, a pattern showed up clean enough to correctly call two more numbers before they were computed.
The 26-number table below is the complete published record for this family (Ahmed–Schaal, 2016); the eleven dark cells are new as of this session.
Everything below is the technical version.
For an equation L(t): x1 + x2 + … + xt−1 = xt (repetitions among the xi permitted), the generalized Schur number S(k; t0,…,tk−1) is the least m such that every k-coloring of {1,…,m} admits, for some color i, a monochromatic solution of L(ti) in color i. Ahmed–Schaal (Experimental Mathematics 25(2), 2016) computed the 26 values that constituted the entire published frontier for k = 3 and conjectured S(3;s,t,u) = stu − tu − u − 1 for 4 ≤ s ≤ t ≤ u (their Conjecture 2.1); for the s = 3 sub-family they conjectured only a strict lower bound (Conjecture 2.2), proved by Song–Mao in April 2026 (secondary) — no exact formula for S(3;3,t,u) is known or conjectured prior to this note. No new exact value in this family had been published since 2016; the diagonal S(3;t,t,t) = t3−t2−t−1 became a theorem in 2019 (Boza–Marín–Revuelta–Sanz (secondary)).
We compute eleven new exact values using a SAT-based boundary search with an independently checked certificate on both sides: a DRUP unsatisfiability proof at S, verified by a from-the-definition proof checker sharing no code with the encoder, and an explicit coloring at S−1, verified by an independent bitset algorithm. Four values — S(3;4,4,8) = 87, S(3;4,4,9) = 98, S(3;4,5,8) = 111, S(3;4,6,7) = 118 — are previously open instances of Conjecture 2.1, all four confirming it. Seven — S(3;3,3,8) = 59, S(3;3,3,9) = 68, S(3;3,3,10) = 77, S(3;3,3,11) = 86, S(3;3,4,8) = 67, S(3;3,4,9) = 78, S(3;3,5,8) = 91 — extend the s = 3 sub-family, where no formula was previously conjectured. We determine the complete extremal structure of the (3,3,u) family at every computed size u ∈ {4,5,6,8,9,10}: exactly one mirror-symmetric skeleton with u−2 free ternary slots (2·3u−2 valid colorings), except at u = 7, where the count (846) matches no such skeleton. From this structure we state Conjecture A: S(3;3,3,u) = 9u−13 for all u ≥ 4 except u = 7. Formulated from data at u ≤ 9, it predicted the exact value, extremal count, slot layout and class maximum at u = 10, and the exact value at u = 11, both subsequently certified. Eleven published boundary values, twelve published climb values and all ten published enumeration counts of Ahmed–Schaal are reproduced exactly as controls, including one extremal coloring matching the paper's printed string character for character. No case of Conjecture 2.1 is refuted; Conjecture A is new and open beyond u = 11.
Fix k ≥ 2 and lengths t0,…,tk−1 ≥ 2. A k-coloring of {1,…,m} is a function c: {1,…,m} → {0,…,k−1}. It is valid for (t0,…,tk−1) if for every color i there is no monochromatic solution of L(ti): x1+…+xti−1 = xti with all xj in color i (the xj need not be distinct). The generalized Schur number S(k;t0,…,tk−1) is the least m with no valid coloring of {1,…,m}; it exists for every choice of lengths by a compactness argument, and {1,…,S−1} always has at least one valid coloring. The classical Schur number is the case ti = 3 for every i. This note works entirely with k = 3, and follows the paper's convention of writing S(3;s,t,u) with s ≤ t ≤ u.
Sourcing, and a rare exception. As in every session on this repository, the sandbox's egress proxy blocked arxiv.org, oeis.org, erdosproblems.com and every other scholarly host with HTTP 403; only web search and (unusually) raw.githubusercontent.com were reachable. That second exception mattered: Tanbir Ahmed hosts the author preprint of Ahmed–Schaal's paper on his own GitHub Pages site, and it was downloaded and read in full this session — the deciding publication boundary for this entire note is therefore a primary source, not a search snippet. Table 1 of the paper was transcribed in full to data/published_values.csv (URL and sha256 of the PDF in data/environment.txt), and its ten enumeration theorems and conventions were read directly rather than inferred. Two supporting citations remain secondary: Song–Mao's April 2026 proof of Conjecture 2.2 (abstract recovered verbatim via an arXiv-RSS mirror) and Boza–Marín–Revuelta–Sanz's 2019 diagonal theorem (a search digest). Minor venues — INTEGERS, theses, MathSciNet — were unreachable; an isolated prior value could in principle predate one reported new here, though a full sweep of the arXiv math.CO RSS feed for "Schur numbers" from 2024 through this session turned up no other paper computing exact values in this family.
Deciding S(k;t0,…,tk−1) ≤ m reduces to Boolean satisfiability: one variable xn,i per (integer n ≤ m, color i), an exactly-one clause per n, and for each color i and each solution (a1,…,ati) of L(ti) with all aj ≤ m, a clause forbidding all ti variables xaj,i from holding simultaneously. The number of such clauses grows with the number of ways to write a sum with ti−1 (not necessarily distinct) parts, so support sets are deduplicated before encoding. S = m exactly when the instance at m is unsatisfiable and the instance at m−1 is satisfiable.
Every boundary ships a certificate on both sides, checked by a program sharing no code with the encoder:
Unsatisfiable side (at S). Glucose 4.2 emits a DRUP proof; tools/satcert/rup_check, a from-the-definition reverse-unit-propagation checker built and validated in an earlier session, verifies it independently of the solver.
Satisfiable side (at S−1). Cadical finds a model; verify_witness.py, an iterated-sumset bitset algorithm sharing no code with the CNF encoder, confirms directly from the definition that no color contains a monochromatic solution of its equation.
CNF files are not stored: the encoder is deterministic, so every instance regenerates byte-identically from schur3.py s,t,u N, and every run's CNF sha256 is recorded in data/results_a1.csv and data/cnf_hashes.txt. Extremal-coloring counts (section 5) use a separate exactly-one-per-class enumeration mode, cross-validated against brute-force product enumeration on fifteen small (t,n) cells with zero mismatches.
Every new value below was computed only after the identical pipeline reproduced the existing literature exactly:
| Check | Result |
|---|---|
| Convention-independent classics | S(2;3,3) = 5, S(3;3,3,3) = 14 — both reproduced before the publication boundary arrived |
| 11 published boundary values | reproduced with full DRUP + witness certificates |
| 12 published S(3;t,u) climb values | reproduced by independent Cadical climb from Song–Mao's proved lower bound |
| 10 published enumeration counts (Ahmed–Schaal Theorems 2.1–2.10) | matched exactly: 18, 54, 162, 846, 8, 1, 112, 96, 3584, 9488 |
| The unique (3,4,5) extremal coloring | matches the paper's printed string character for character |
| Negative control | a deliberately corrupted witness is rejected by verify_witness.py |
Every anchor of Song–Mao's Conjecture 2.2 lower bound (2tu−u−1 for s = 3) held as satisfiable, as their theorem requires — a further, independent consistency check running underneath every s = 3 value in section 4.
Re-run independently while building this page — not merely re-read from the session log — against the committed scripts and data:
| Triple (s,t,u) | S(3;s,t,u) | Family | Note |
|---|---|---|---|
| 3, 3, 8 | 59 | s = 3 (no formula) | re-verified: DRUP at 59, witness at 58 |
| 3, 3, 9 | 68 | s = 3 (no formula) | |
| 3, 3, 10 | 77 | s = 3 (no formula) | re-verified independently — predicted by Conjecture A (section 6) |
| 3, 3, 11 | 86 | s = 3 (no formula) | re-verified independently — predicted by Conjecture A (section 6) |
| 3, 4, 8 | 67 | s = 3 (no formula) | |
| 3, 4, 9 | 78 | s = 3 (no formula) | |
| 3, 5, 8 | 91 | s = 3 (no formula) | |
| 4, 4, 8 | 87 | s ≥ 4 (Conjecture 2.1) | re-verified independently — confirms 4·4·8−4·8−8−1 = 87 |
| 4, 4, 9 | 98 | s ≥ 4 (Conjecture 2.1) | confirms the formula |
| 4, 5, 8 | 111 | s ≥ 4 (Conjecture 2.1) | confirms the formula |
| 4, 6, 7 | 118 | s ≥ 4 (Conjecture 2.1) | confirms the formula |
CERTIFIED. Full per-run details, times and CNF hashes: data/results_a1.csv. Three re-verifications performed independently while building this page (S(3;3,3,10)=77, S(3;3,3,11)=86, S(3;4,4,8)=87) matched the session's own certificates exactly, on both the DRUP and witness sides.
What did not finish. The (4,4,u) ladder beyond u = 9 is open for memory, not time: the DRUP proof for (4,4,10) outgrows the sandbox's 15 GB RAM before the solver can finish, and three attempts were killed by the OS. The (4,4,·) equations are anomalously hard for their size — two short forbidden sums and one long one give a SAT solver very little to propagate on. A disk-streaming proof pipeline (a standalone solver binary plus drat-trim, rather than holding the whole proof in a Python process's memory) is the identified fix, untried this session.
Ahmed–Schaal's ten enumeration theorems count, for several small triples, every valid coloring at the critical size S−1 — not just exhibit one. Extending that to every computed (3,3,u) triple reveals a single, almost-uniform structure:
For u ∈ {4,5,6,8,9,10}, every valid coloring of {1,…,9u−14} for (3,3,u) is an instance of one mirror-symmetric pattern with exactly u−2 free ternary choices, at positions 2u+1+5j for j = 0,…,u−3. This gives exactly 2·3u−2 valid colorings, and the largest same-color class within any of them has size exactly 5(u−2).
Re-verified directly from the committed certificates while building this page: the enumeration file for u = 10 has 13,122 colorings (2·38 = 13,122, exactly), and for u = 4 has 18 (2·32). Both match the formula on the nose.
At u = 7, S(3;3,3,7) = 49 — one below the 9u−13 = 50 that every neighboring u satisfies. The 846 valid colorings of {1,…,S−1} (re-verified: the certificate file has exactly 847 lines, one a header) form no single skeleton: 846 = 2·32·47, not a power of 3 times 2, and no mirror-symmetric pattern with free ternary slots accounts for the count. The class-maximum law still holds (25 = 5·5); the deficit traces to the equation pair itself, not to the free-slot structure.
Nothing here is claimed as a proof of the skeleton for all u: it is a certified structural fact at every size actually computed, u ∈ {4,5,6,8,9,10}, with one genuine exception at u = 7 whose cause is identified but not yet reduced to a short argument.
The skeleton's slot count immediately suggests a closed form for the value itself. Fit to the values known partway through the session (u ≤ 9: 23, 32, 41, 59, 68, skipping the anomalous 49 at u = 7):
Before u = 10 or u = 11 had been computed, Conjecture A made four further predictions at u = 10 (the exact value 77, the extremal count 13,122, the free-slot positions 21+5j, and the class maximum 40) and one at u = 11 (the value 86). All five landed exactly. This is stated plainly because it is easy to overstate: a formula predicting its own next few inputs correctly is suggestive, not a proof, and five confirmations is a small number against "for all u." The figure in the plain-language section above plots every computed value against the line.
CERTIFIED at u ≤ 11 (every value on this page independently re-verified, section 4). The conjecture itself — the claim for all u ≥ 4, u ≠ 7 — is open. No construction proving the lower-bound direction for general u was found this session: the five skeletons share their slot arithmetic and their first several cells, but the prefix block changes shape with u, and no single parametrization covering u ∈ {4,5,6,8,9,10} survived the session (see WRITEUP.md). Whether u = 7 is the only exception, or the first of a sparse family of exceptions, is unknown.
Settled here. Eleven new exact values (CERTIFIED, DRUP + independent witness on every boundary); the complete (3,3,u) extremal structure at every computed size, with the u = 7 exception identified and localized (CERTIFIED); Conjecture A, stated and certified at u ≤ 11 (a conjecture, not a theorem — open for u ≥ 12 and unproven in general).
Not settled. Conjecture A for general u. The (4,4,u) ladder beyond u = 9 (a memory limit, not a mathematical obstruction). Whether the s = 3 family has any closed form beyond the (3,3,u) sub-case. Whether an isolated prior value sits in an unreachable minor venue (section 1).
Sharpest open threads, in order:
drat-trim, proof written to disk rather than held in memory) to reopen the (4,4,u) ladder past u = 9, and to reach further (t,u) at s = 3.