Erdős problem #172 — wave 5l
Date: 2026-07-26 (UTC)
Claim labels
- (a) elementary-rigorous: proved from first principles here.
- (b) rigorous-modulo-named-theorem/source: a statement checked in the cited primary source.
- (c) plausible/structural-unverified: a route, diagnosis, or literature-search inference, not used as a theorem.
- (d) computational-only: a finite result backed by the standalone checker; it is not promoted to an infinite theorem.
Step 0: mandatory live-page audit
I fetched the fully rendered live page through the Bright Data browser on 2026-07-26. It displayed OPEN, 0 comments on this problem, 0 claimed proofs for this problem, Interested in collaborating: None, and Currently working on this problem: None. It says it was last edited 06 April 2026. Therefore neither mandatory stop condition applied. (d)
Verbatim current statement
> Is it true that in any finite colouring of ℕ there exist arbitrarily large finite 𝐴 such that all sums and products of distinct elements in 𝐴 are the same colour?
Everything currently listed as a known result
The page says:
- Hindman first asked the question. (d)
- Hindman [Hi80] proved that the assertion is false, with seven colours, if \(A\) is required to be infinite. Erdős [Er77c] asked about infinite \(A\) with only two colours, referring to problem #1198. (b)
- Moreira [Mo17] proved that every finite colouring of \(\mathbb N\) has \(x,y\) for which \(\{x,x+y,xy\}\) is monochromatic. (b)
- Alweiss [Al23] proved that every finite colouring of \(\mathbb Q\setminus\{0\}\) has arbitrarily large finite \(A\) whose sums and products of distinct elements are all one colour. Bowen and Sabok [BoSa22] had earlier proved the first nontrivial rational case \(|A|=2\). (b)
The page additionally shows Formalised statement? Yes, but gives no comments, proof claims, worker, or collaboration marker. (d)
1. Primary-source literature audit
I searched arXiv and the web by the exact statement, the phrases “pairwise sums and pairwise products” and “Hindman conjecture,” the cited authors, and forward-looking 2024–2026 variants. I then downloaded and inspected the primary TeX/PDF rather than relying on search snippets. The optional --sources mode of the companion checker re-downloads the primary TeX and asserts the theorem phrases and identifiers below. (d)
1. Hindman's Partitions and sums and products—Two counterexamples, JCTA 29 (1980), 113–120, DOI 10.1016/0097-3165(80)90052-790052-7), has exactly the seven-colour infinitary counterexample described on the live page. (b)
2. Moreira's arXiv:1605.01469, Corollary 1.5, gives monochromatic \(\{x,xy,x+y\}\) in every finite colouring of \(\mathbb N\). (b)
3. Bowen–Sabok arXiv:2210.12290 proves monochromatic \(\{x,y,xy,x+y\}\) for arbitrary finite colourings of \(\mathbb Q\). (b)
4. Alweiss arXiv:2307.08901 proves the stronger rational statement in which all nonempty subset sums and subset products, including the singleton terms, are monochromatic. Its introduction still states the natural-number version as a conjecture. (b)
5. Bowen arXiv:2205.12921, published in Advances in Mathematics 462 (2025), proves for two colours configurations consisting of the \(x_i\), prefix products, and one total sum. This does not control every one of the \(\binom{k}{2}\) pairwise sums and products. (b)/(a)
6. Alweiss–Bowen–Sabok arXiv:2512.09598 proves that every two-colouring has monochromatic \(\{x,y,xy,x+iy: i\leq k\}\). This again is not the page's all-pairs configuration. (b)/(a)
I found no primary source claiming the full natural-number statement or the finite computations below. This is an honest search result, not a proof of absence or a priority claim. (c)
2. Exact finite reduction
For a finite set \(A\subset\mathbb N\), put
\[ \Phi(A)=\{a+b,ab:a,b\in A,\ aFor \(N,k\geq2\), let \(\mathcal H_{N,k}\) be the hypergraph with vertex set \([N]\) and one hyperedge \(\Phi(A)\) for every \(k\)-set \(A\) satisfying \(\Phi(A)\subseteq[N]\). Such an \(A\) automatically lies in \([N]\): for each \(a\in A\), pairing it with another element produces either a sum or a product at least \(a\). (a)Define
\[ R(r,k)=\min\{N:\mathcal H_{N,k}\text{ has no }r\text{-colouring without a monochromatic hyperedge}\}, \]with \(R(r,k)=\infty\) if there is no such \(N\). Then the live problem is equivalent to
\[ R(r,k)<\infty\qquad\text{for every }r,k\geq2. \tag{2.1} \](a)
Proof: if all these thresholds are finite, restrict an arbitrary \(r\)-colouring of \(\mathbb N\) to \([R(r,k)]\) for each \(k\). Conversely, if avoiding colourings of every \(\mathcal H_{N,k}\) exist for some fixed \(r,k\), arrange them in the finitely branching tree whose parent map is restriction from \([N+1]\) to \([N]\). Every level is nonempty, so König's lemma gives an infinite \(r\)-colouring with no \(k\)-element witness. Finally, witnesses of unbounded size are equivalent to witnesses of every exact size because every subset of a witness is a witness. (a)
This isolates the remaining task exactly: prove a finite obstruction for every \((r,k)\), or construct one infinite avoiding branch. (a)
3. Verified finite results
With the convention above, the computation proves
\[ \boxed{R(2,2)=8,\qquad R(3,2)=100,\qquad R(2,3)>5000.} \tag{3.1} \]The first equality is elementary; the other equality and the lower bound are computational-only. (a)/(d)
3.1 The pair case is ordinary graph colouring
When \(k=2\), every hyperedge is the pair \(\{x+y,xy\}\). Let \(G_N\) be the ordinary graph on \([N]\) with those pairs as edges for distinct \(x,y\) and with both values at most \(N\). Thus \(R(r,2)\) is the first \(N\) with \(\chi(G_N)>r\). (a)
For \(N=7\), \(G_7\) is the path
\[ 2-3-4-5-6-7 \]plus isolated vertex \(1\), so it is two-colourable. At \(N=8\), vertices \(6,7,8\) form a triangle: its edges arise respectively from
\[ (x,y)=(1,6),\quad(1,7),\quad(2,4). \]Hence \(R(2,2)=8\). (a)
For three colours, the checker constructs \(G_{99}\) with 230 edges and validates this explicit colouring of vertices \(2,\ldots,99\):
10101201021202120101021012010121021202120101212102120212010120121202021201010121012101010101021202
It constructs \(G_{100}\) with 234 edges and exhaustively proves that no three-colouring exists. The new vertex has neighbourhood
\[ N_{G_{100}}(100)=\{25,29,52,99\}, \]with witnesses \((5,20),(4,25),(2,50),(1,99)\). The independent standard-library DSATUR search uses saturation degree, ordinary degree, and a fixed vertex tie-break; it explores unused colours only up to permutation. It finds the \(G_{99}\) colouring in 152 recursive calls and exhausts \(G_{100}\) in 148 calls. Thus \(R(3,2)=100\). This upper certificate is exact but computer-assisted, not a hand proof. (d)
3.2 An explicit two-colouring through 5000 avoiding every triple
For \(a \[ \Phi(\{a,b,c\})\subseteq[N]\quad\Longleftrightarrow\quad bc\leq N. \tag{3.2} \]
The number of triples to check is therefore
\[ \sum_{b=2}^{\lfloor\sqrt N\rfloor} (b-1)\bigl(\lfloor N/b\rfloor-b\bigr). \tag{3.3} \]At \(N=5000\), this is exactly \(210{,}497\). (a)
The file erdos172_wave5l_reverify.py embeds an explicit 5000-bit colouring, compressed losslessly as a base85/zlib constant. Its uncompressed SHA-256 is
a8599f44248a5ff52c954f4ee2e7b0687dd73eafed89545ecb40081d170b682b
It has 3,980 zeros and 1,020 ones. Direct enumeration confirms that every one of the 210,497 footprints contains both colours; 209,919 footprints have six distinct values and 578 have five. Consequently \(R(2,3)>5000\). (d)
The model was discovered from the exact not-all-equal CNF. If \(z_v\) is the colour bit of \(v\), each footprint \(E\) contributes
\[ \left(\bigvee_{v\in E}z_v\right) \land \left(\bigvee_{v\in E}\neg z_v\right). \tag{3.4} \]The discovery instance had 5,000 variables and 420,994 clauses. The recorded model was found in 44.6 seconds on this VM. The essential generator was:
clauses = []
for b in range(2, N + 1):
for c in range(b + 1, N // b + 1):
for a in range(1, b):
edge = sorted({
a + b, a + c, b + c,
a * b, a * c, b * c,
})
clauses.append(edge)
clauses.append([-v for v in edge])
SAT is used only for discovery. The run used PySAT 1.9.dev7 with its cadical195 backend (CaDiCaL 1.9.5). The final verifier contains no SAT library and checks the decoded colouring against all triples from scratch. (d)
4. Independent reproduction
Run:
$ python runs/erdos172_wave5l_reverify.py --sources
Triple certificate passed: N=5000, triples=210497, footprint sizes={5: 578, 6: 209919}, 0/1 counts=3980/1020.
Pair thresholds passed: R(2,2)=8 (G7/G8 edges 5/7; UNSAT calls 6); R(3,2)=100 (G99/G100 edges 230/234; SAT/UNSAT calls 152/148).
ALL MATHEMATICAL CHECKS PASSED
Primary-source TeX/metadata checks passed.
ALL SOURCE CHECKS PASSED
The mathematical run takes under one second here. Source mode re-downloads six arXiv source archives and Hindman's Crossref record; it does not attempt to bypass Cloudflare or silently substitute a cached Erdős Problems page. (d)
5. Exact wall and next computation
The rational theorem does not transfer by simply clearing denominators. If \(q_i\in\mathbb Q\) and \(Dq_i\in\mathbb Z\), then for \(b_i=Dq_i\),
\[ b_i+b_j=D(q_i+q_j),\qquad b_ib_j=D^2q_iq_j. \]Thus sums and products land in two different dilation scales. Alweiss's rational colouring theorem supplies no compatibility between an arbitrary natural-number colouring at the \(D\) and \(D^2\) scales. A sufficient missing ingredient would be a denominator-synchronisation lemma that produces the rational configuration and a common \(D\) while forcing the original colouring to agree on all these linearly and quadratically scaled values. No such lemma was found. (a)/(c)
The two-variable Moreira/Bowen machinery likewise controls selected sums and products, not the simultaneous complete graph of \(\binom{k}{2}\) pairs. The exact missing combinatorial statement is already the first unsettled finite target \(R(2,3)<\infty\). Our certificate only pushes its lower bound past 5000; it gives no infinite colouring and no evidence that the threshold is infinite. (a)/(c)/(d)
For scale, \(N=7500\) has 393,628 hyperedges and 787,256 clauses; a CaDiCaL decision run was stopped unresolved after about 75 seconds. At \(N=100{,}000\), (3.3) gives 20,424,017 hyperedges, 40,848,034 clauses, and up to about 245 million literal occurrences. That is roughly 0.91 GiB for raw 32-bit literals alone and several times more with solver/Python overhead. A modest next portfolio of 16 cores for two hours would cost 32 core-hours, about US$1.60–$3.20 at US$0.05–$0.10 per vCPU-hour, but has no guaranteed outcome; an UNSAT certificate could be much more expensive. I did not run that heavier search here. (d)/(c)
The finite pair thresholds do not address unbounded \(|A|\), and the \(N=5000\) colouring is not a counterexample on \(\mathbb N\). The live problem remains open. (a)/(d)
PARTIAL: Reduced the problem exactly to finite weak hypergraph colouring, proved R(2,2)=8, computationally certified R(3,2)=100, and supplied a solver-free-verified 2-colouring of [5000] showing R(2,3)>5000; no uniform or infinite conclusion.