Fabian Arevalo

Computational research note · Arithmetic Ramsey theory

Reciprocal Rado numbers: an independent certificate for nine exact values

Nine values already reported, without a certificate, by Myers–Parrish or Gaiser–Ramezanpour (July 2026), reproduced by an independent machine-checked pipeline, plus the extremal coloring structure, a mechanism for part of their Conjecture 1.3, and a new bound at four colors.
August 8, 2026 · produced in a Claude Code research session, fact-checked against the primary sources during publication.
Code, data & certificates in the companion repository.

Start here: no maths required

Pick two whole numbers and add their reciprocals: 1/x + 1/y. Physicists know this shape well: it is the lens equation, the rule for where a lens focuses light. Now paint every whole number from 1 up to some limit either red or blue. Can you always avoid a same-colored triple x, y, z with 1/x + 1/y = 1/z?

Below a certain limit, yes: you can always dodge it with a clever paint job. Above it, every paint job fails. Call that breaking point f(2). The same question makes sense with more terms on the left of the equation, and with more than two colors; each choice gets its own breaking point. Several of these breaking points were already known, computed by two different research groups, most recently one month before this note. Neither computation carried a proof beyond "the program said so."

This note recomputes nine of those breaking points from scratch, with a second, independent kind of evidence attached to each: not just a number, but a machine-checked proof that the number is exactly right. The figure below is that certified table.

Hover or tap a bar.
The threshold f2(k), independently certified, against the reference value 3k2 (Gaiser–Ramezanpour's theorem and conjecture, not a fit). All seven values were already reported, without a certificate, by Myers–Parrish (k = 2, 3, 4) or Gaiser–Ramezanpour (k = 5, 6, 7, 8); this figure is this repository's independent, certified reproduction. Filled bars are the odd prime powers (k = 3, 5, 7): each clears the reference line. Outlined bars are k = 2, 4, 6, 8; at 4, 6 and 8 the threshold lands exactly on 3k2, as their Conjecture 1.3 predicts for every k ≥ 4 that is not an odd prime power. What happens past k = 8 is in section 8: the gap does not keep shrinking.

What this note adds

Gaiser (2023) introduced fr(k) and proved f2(k) = O(k3). Myers and Parrish computed f2(2), f2(3), f2(4) and f3(2). Gaiser and Ramezanpour (arXiv:2607.04373, July 2026) proved f2(k) = 3k2 whenever k = 3·2m, proved f2(k) ≥ 3k2+1 whenever k is an odd prime power, extended the computed table to f2(k) for k up to 25 and f3(k) for k up to 5, noted from their own table that the bound is never quite attained at an odd prime power, and conjectured (their Conjecture 1.3) that f2(k) = 3k2 at every k ≥ 4 that is not an odd prime power. None of that work carries a certificate: no proof of unsatisfiability, no independently re-checked witness coloring; the paper says so of itself.

  1. Nine independently certified values. A from-scratch pipeline, sharing no code with either prior computation, reproduces f2(2)=60, f2(3)=40, f2(4)=48, f2(5)=80, f2(6)=108, f2(7)=150, f2(8)=192, f3(2)=3276 and f3(3)=585 exactly, each with a proof that no valid coloring exists one step past the value, and a witness coloring at the value itself, checked by two independently written programs. Two differently built pipelines landing on the same nine integers is evidence neither paper offered on its own.
  2. The pattern their own table already showed, now with a proof attached. Gaiser–Ramezanpour's own concluding remarks note that their bound is never quite attained at an odd prime power. This note supplies the missing certificate for three of those cases (k = 3, 5, 7), and section 8 reports what their own table, not independently certified here, shows happens next: the gap does not shrink to zero.
  3. A candidate mechanism for half of their Conjecture 1.3. Gaiser–Ramezanpour conjecture f2(k) = 3k2 for every k ≥ 4 that is not an odd prime power, purely from their table, with no proposed reason. This note gives a structural argument for the even-k half of that conjecture (an extra family of forced same-colorings that exists only when k is even), certified at k = 4, 6 and 8, having recorded the k = 8 value as a prediction before computing it.
  4. The coloring structure itself, which neither prior source examines: at k = 8 the optimal coloring is a plain two-block pattern with zero exceptions; at k = 7 it needs six sparse corrections. The extremal three-coloring at f3(2) keeps every number a different color from its own double, all 1637 times it is possible to check.
  5. A new bound at four colors, f4(2) > 60000, independently verified; no prior published value was found.
  6. An error, caught by the method built to catch it. A first estimate of f2(7) was wrong, briefly recorded as "at least 168," because a background computation was read before it finished. A second, independently sound proof lane flatly contradicted it the same session, and the record was corrected before anything left the sandbox. The true value is 150; the near-miss is on the record in section 3.

Everything below is the technical version. Sections 3 and 4 spell out exactly what was already known, from where, and what is newly certified.

AbstractSummary of results

The reciprocal Rado number fr(k) is the least n such that every r-coloring of {1,…,n} contains a monochromatic solution of 1/x1 + … + 1/xk = 1/xk+1 (repeats allowed). Myers and Parrish computed f2(2), f2(3), f2(4) and f3(2). Gaiser and Ramezanpour (arXiv:2607.04373, July 2026) proved f2(3·2m) = 3k2 for m ≥ 1 and f2(pm) ≥ 3k2+1 for odd prime powers, extended the computed table to f2(k) for k ≤ 25 and f3(k) for k ≤ 5, observed from that table that their bound is never exactly attained at an odd prime power, and conjectured (their Conjecture 1.3) that f2(k) = 3k2 at every k ≥ 4 that is not an odd prime power. Neither computation carries a certificate. We build an independent pipeline, sharing no code with either source, and reproduce all nine values that overlap our computed range exactly, each with a proof of unsatisfiability at f, checked by an independent forward proof-checker, and a witness coloring at f−1 verified by two independently written checkers. We give a structural mechanism, absent from the source paper, for the even-k half of Conjecture 1.3: an extra family of forced same-colorings, parametrized the same way as the k = 2 case itself, that exists only when k is even; we certify it at k = 4, 6 and 8, having recorded the k = 8 prediction before computing it. We determine the extremal coloring structure at every certified size: a plain two-block skeleton with zero deviations at k = 8, six sparse corrections at k = 7, and, at f3(2), a three-coloring in which every integer differs in color from its own double at all 1637 applicable pairs. We give a new bound at four colors, f4(2) > 60000 (verified witness), and note, from Gaiser–Ramezanpour's own uncertified table, that their reported f2(9) = 252 breaks the shrinking-excess pattern visible at k = 3, 5, 7.

AI-assisted (Claude); every computational claim of this repository ships a certificate reproducible from the companion repository. The research session that ran the underlying computation worked from an egress-blocked sandbox and could not reach arxiv.org; its own notes accordingly marked the source paper's contents as secondary. This page was built on a machine with ordinary internet access, and the claims about arXiv:2306.04029 and arXiv:2607.04373 below were checked directly against the papers during publication, not carried over unverified.

Section 1Setting, and what was already known

fr(k): the least n such that every function χ : {1,…,n} → {0,…,r−1} admits x1, …, xk, y ∈ [1,n], monochromatic under χ, with ∑i 1/xi = 1/y. The xi need not be distinct. Positivity forces y < xi for all i, so a solution's support set {y, x1, …, xk} has between 2 and k+1 elements, and the integer 1 can appear in a solution only as y (then ∑ 1/xi = 1 is an Egyptian representation of 1 by k unit fractions). Monotonicity: a valid coloring of [1,n] restricts to [1,n′] for n′ < n, so fr(k) is exactly the first n with no valid coloring; every boundary claim below certifies the pair (SAT at f−1, UNSAT at f).

What was already known, and by whom. Checked directly against the primary sources during this page's publication: arXiv is reachable from the machine that built this page, though it was not reachable from the sandbox that ran the underlying computation, which is why that session's own notes hedge this material as secondary. Gaiser introduced fr(k) and proved f2(k) = O(k3), down from Brown–Rödl's O(k6), in On Rado numbers for equations with unit fractions (arXiv:2306.04029; Discrete Math. 347 (2024) 114156). Myers and Parrish computed f2(k) for k ∈ {2,3,4} and f3(2); this note has not located their paper directly and cites it exactly as it appears in Gaiser–Ramezanpour's own bibliography (their reference [12]), (secondary). Gaiser and Ramezanpour, A sharp lower bound for some reciprocal Rado numbers (arXiv:2607.04373, July 2026), prove f2(3·2m) = 3k2 for m ≥ 1 and f2(pm) ≥ 3k2+1 for k a power of an odd prime p (their Theorem 3.1), prove fr(2) ≥ 4r/2 and fr(k) ≥ (2r−1)kr for k ≥ 3 (their Theorem 1.1), and report their own computed table of f2(k) for k ≤ 25 and f3(k) for k ≤ 5, using the Cadical and Glucose solvers inside the PySAT toolkit, with, by the paper's own acknowledgement, help from ChatGPT 5.5 and Codex. The paper documents no certificate or independent proof-check on any of these computed values.

Section 2Methodology and controls

Solutions are enumerated by exact integer arithmetic (64-bit with 128-bit intermediate products, overflow-checked) via a bounded Egyptian-fraction depth-first search; for k = 2 an independent divisor-form enumerator (x = z+d1, y = z+d2, d1d2 = z2, via a smallest-prime-factor sieve) cross-checks it. The CNF encoding needs no at-most-one clauses: for each integer, an at-least-one-color clause; for each solution support set and each color, a clause forbidding all of that support taking that color. Any satisfying assignment projects to a valid coloring.

Bracketing runs on CaDiCaL 3.0.0 without proofs; the claimed value rests only on the boundary pair. Glucose 4.2.1, built standalone so its pure-RUP proof output streams to disk rather than filling RAM, proves UNSAT at f, checked by an independent forward proof-checker (tools/satcert/rup_check); the SAT witness at f−1 is re-verified by two independently written checkers sharing no code with the encoder (Python and C). At k ≥ 7, where full enumeration is infeasible, values come from a counterexample-guided (CEGAR) loop whose UNSAT verdicts are sound by construction, since every clause is the support of an exactly verified solution, so a subset of true constraints being UNSAT implies the full encoding is, and whose SAT verdicts are accepted only when the complete independent checker passes the model.

Controls, run before any new claim. The enumerator reproduces OEIS A002966 (1, 3, 14, 147 Egyptian representations of 1 for k = 2..5, confirmed directly against the OEIS record during this page's publication); the C depth-first search agrees with an independent Python implementation and with brute-force Fraction arithmetic on a grid of small (k,n), and with the k = 2 divisor enumerator at n = 60, 500, 3276; and the two Gaiser–Ramezanpour theorem anchors reproduce exactly: f2(6) = 108 = 3·62 (their m = 1 case) and SAT witnesses exist at n = 3k2 for k = 3, 5, 7, precisely where their odd-prime-power bound requires them. Every proof of unsatisfiability is re-checked by rup_check, itself validated against injected non-proofs and truncated proofs; a deliberately corrupted witness is rejected by both independent checkers, which agree on the exact violating tuple 1/55 + 1/70 + 3/77 = 1/14.

Section 3The two-color ladder

kf2(k)3k2Δ = f2(k) − 3k2already reported by
26012+48Myers–Parrish
34027+13Myers–Parrish
448480Myers–Parrish
58075+5Gaiser–Ramezanpour
61081080Gaiser–Ramezanpour's own theorem (k = 3·2m)
7150147+3Gaiser–Ramezanpour
81921920Gaiser–Ramezanpour

All nine values in this table, together with f3(2) = 3276 and f3(3) = 585 in section 5, were already reported, without a certificate, by one of these two sources. This repository's contribution to the raw numbers is independent reproduction with proof, not discovery. What is new: the certificate itself, a proof of unsatisfiability at f checked by an independent forward proof-checker, plus a witness at f−1 verified by two independently written programs sharing no code with the encoder or with each other; and, in section 6, the structure of the extremal colorings, which neither source examines.

Gaiser–Ramezanpour's own concluding remarks already note the pattern their table shows: their computed values exceed their own lower bound at every odd prime power in range, and they pose improving that bound as their own open problem. This note supplies the missing certificate for the three cases inside its computed range, k = 3, 5, 7 (Δ = 13, 5, 3), and section 8 reports, from their own table, what happens at k = 9 and beyond.

An error, caught

The first estimate of f2(7) was wrong: "≥ 168" was briefly written into a draft, from a background enumeration that had not finished. The completion check was cat logfile && echo DONE, which succeeds on a logfile that exists but is still empty; the probe then read the stripe files mid-write and reported a satisfying coloring at n = 167 against an incomplete constraint set.

The mistake surfaced because a second, independently sound proof lane (the CEGAR loop, whose UNSAT verdicts are sound by construction) reported UNSAT at n = 151, flatly contradicting a claim of SAT at 167: a subset of true constraints can never be UNSAT where the full constraint set is SAT. Chasing the contradiction found two junk tuples from the mid-flight read. The claim was struck the same session; the certified value is 150.

Section 4Conjecture 1.3, and the even-k mechanism

Conjecture 1.3 (Gaiser–Ramezanpour, stated from their computational table; not proved by them or here)

If k ≥ 4 is not an odd prime power, then f2(k) = 3k2.

This is strictly more general than what is certified here: it also covers odd composite k with more than one distinct prime factor, such as k = 15 and k = 21, both confirmed in Gaiser–Ramezanpour's own uncertified table (Δ = 0 at both). This repository certifies only the even-k instances k = 4, 6, 8, and offers a structural mechanism for the even case specifically, which the source paper does not attempt.

Mechanism sketch, not a proof. At n = 3k2 the diagonal solutions (k copies of ky, target y) give edges {y, ky} for y ≤ 3k, whose chains {j, jk, jk2} (j ≤ 3) force χ(jk2) = χ(j) in any valid 2-coloring. For even k there are additionally "half-diagonal" solutions: k/2 copies each of a and b with 1/a + 1/b = 2/(ky), parametrized the same way as the k = 2 equation itself (d1d2 = (ky/2)2). These knit the diagonal chains together into a web rigid enough to force the coloring at n = 3k2 to fail. For odd k this family does not exist (k/2 is not an integer), the constraint web at 3k2 is strictly poorer, and the computed colorings survive to 3k2 + Δ, exactly the sparse-correction structure in section 6.

This mechanism is specific to even k. It does not explain why the odd, composite, non-prime-power cases k = 15 and k = 21 also land exactly on 3k2 in Gaiser–Ramezanpour's table: a different argument must be doing the work there, and finding it is an open thread (section 8), effectively Gaiser–Ramezanpour's own Problem 5.3 (find a quadratic upper bound for f2(k), k ≥ 4) stated as a mechanism question rather than a bound.

Section 5Three and four colors

rkfr(k)already reported byGR lower boundratio
323276Myers–Parrish43/2 = 32102×
33585Gaiser–Ramezanpour(23−1)·33 = 1893.1×
42> 60000 (verified witness)not found in either source44/2 = 128> 468×

Gaiser–Ramezanpour also report f3(4) = 960 and f3(5) = 1650 in the same table (their Table 3, uncertified; not attempted by this repository's pipeline). f3(2) = 3276 = 22·32·7·13 and f3(3) = 585 = 32·5·13. The extremal 3-coloring of [1, 3275] satisfies χ(z) ≠ χ(2z) for all 1637 applicable pairs (figure below) while being otherwise highly fragmented: 1804 maximal runs, class sizes 1849 / 1032 / 394. No r = 4 table appears in either prior source; a 4-color instance at n = 150000 (600000 variables, 1,724,780 clauses) was still undecided by CaDiCaL at this repository's session close, so f4(2) > 60000 (verified witness) stands as the honest bound rather than an exact value.

Section 6Structure of the extremal colorings

Neither prior source examines the shape of an extremal coloring, only its length. The k = 4 witness at n = 47 and the k = 8 witness at n = 191 are exactly the plain two-block construction A = {1,2} ∪ [3k, 3k2−1], B = [3, 3k−1], with not a single deviation at k = 8. B is solution-free because k terms drawn from [3, 3k−1] sum to more than 1/y for any y ≥ 3 in range; A is solution-free because a diagonal solution needs ky ≤ n. At odd k the optimum exceeds 3k2 through sparse corrections to this skeleton, whose exact-hit avoidance is delicate: no uniform parametrization was found. The picture: at even k the plain skeleton is optimal; at odd k it is beatable, by a shrinking-then-not-shrinking margin (section 8).

Hover or tap a cell.
The extremal coloring, position by position, against the plain skeleton A = {1,2} ∪ [3k, 3k2−1], B = [3, 3k−1]. Filled cells are class B, outlined cells class A; a tick marks any position that deviates from the skeleton, computed live against each certified witness, not hardcoded. At k = 8 there is not one deviation. At k = 7 there are six: 17 and 19 sit in A where the skeleton would put them in B, and 119, 133, 136 and 147 sit in B where the skeleton would put them in A. Those six are the sparse corrections that let the odd-k optimum beat 3k2.
z up to220
Each z plotted at its class (lane 0, 1 or 2 of the certified 3-coloring of [1, 3275]), with a line to 2z when 2z is also in view. No line ever connects two points in the same lane: z and 2z never share a color, checked live against the full witness for every z in view. This structural picture is not in either prior source.

Section 7Reproducibility

Environment: 4-core cloud sandbox, 15 GB RAM, Python 3.11.15, gcc 13.3.0, python-sat 1.9.dev8 (bracketing via Cadical153), standalone Glucose 4.2.1 (built from the python-sat sdist's own bundled source so its certified mode streams proofs to disk instead of filling RAM) and CaDiCaL 3.0.0 for the f4 bracketing. No randomness anywhere: enumerators and encoders are deterministic, solvers run default single-thread configurations.

Every UNSAT boundary ships a certs/*.drup file checked by tools/satcert/rup_check; sweep CNFs regenerate byte-identically from the deterministic encoder (sha256 recorded in data/results.csv); CEGAR CNFs are retained on disk with per-clause solution provenance in certs/*.sols, so an auditor can re-verify every constraint in exact arithmetic without trusting the generator. Every SAT boundary ships a witness verified by two independently written checkers (Python and C).

Exception recorded honestly: the k = 8 witness at n = 191 passed the C checker twice (in-loop and at certification) but the slower Python checker was still running at session close. The C checker's validation battery is the warrant in the interim; a future session should close the loop by re-running verify_witness.py certs/f2_8_n191.witness.

Runtimes: every f2 value used at most 30 seconds of solver time; f3(2) = 3276 took about 4 minutes including its 10 MB proof; the f2(7) certificate re-derives in about 40 seconds and reproduced byte-identically (equal CNF sha256) on a second execution.

filerole
enum.c / enum2.cfull solution enumerator (exact rational depth-first search) and an independent k = 2 divisor-based enumerator
sweep.pyfull-enumeration driver: enumerate, bracket with CaDiCaL, certify the boundary pair
cegar.pylarge-k driver: verified-clause subsets grown by counterexamples; --certify-at re-derives a known boundary
verify_witness.py / check_class.ctwo independent witness checkers, Python and C
make_table.pyaggregates the append-only run log into the authoritative table, aborting on any conflict

Full reproduction commands are in the conjecture README; code, certificates and data are in the companion repository.

Section 8Open questions

What Gaiser–Ramezanpour's own table already shows past k = 8. Their Table 2 (uncertified, not independently checked by this repository) reports:

k91113151719212325
f2(k), their table2523725136758801092132316081887
Δ = f2(k) − 3k2996013902112

k = 15 and k = 21 are the odd, composite, non-prime-power cases (Δ = 0, consistent with Conjecture 1.3); every other k above is an odd prime power, and Δ is nowhere close to shrinking to zero: it rises to 9 at k = 9 and 11, and reaches 21 at k = 23. The session's own recorded prediction for k = 9, made before its certified run finished, was Δ ∈ {2, 3}; the true value (per Gaiser–Ramezanpour's table, not certified here) is 9. Certifying k = 9 through 25 with this repository's own pipeline, to find out whether their reported values are exactly right, is the most direct next step.

  1. Is there a closed form, or even a conjectural shape, for Δ(k) at odd prime powers? The data above rules out simple monotone shrinkage; Gaiser–Ramezanpour pose improving their own bound as their Problem 5.2 without a proposed answer.
  2. Gaiser–Ramezanpour's Problem 5.3 asks for a quadratic upper bound for f2(k), k ≥ 4. The mechanism in section 4 supplies one argument for even k; the odd composite non-prime-power cases (15, 21, and whatever comes next) need a different argument, not yet found here.
  3. Growth of fr(2) with r: 2, 60, 3276, more than 6×104, runs far past the 4r/2 bound. Is a closed form, or even the right growth rate, accessible? The dyadic pair constraint {z, 2z} alone forces alternation on dyadic chains; the interaction with 3-support solutions is where the truth likely lives.
  4. No OEIS sequence for this family was found by search, (secondary, absence of evidence). A submission is a candidate once this repository's own values are certified against the fuller range Gaiser–Ramezanpour's table already covers.