Computational research note · van Doorn's consecutive-powerful-triples conjecture, 2026
Break a number apart into its prime factors. Call it powerful if none of those primes shows up just once: 8 = 2×2×2 is powerful, but 12 = 2×2×3 is not, because the 3 only appears once.
Now look for three powerful numbers in a row, evenly spaced, like beads on a ruler: 1728, 1764 and 1800 differ by 36 each, and all three are powerful. In May 2026, a paper by van Doorn conjectured that infinitely many such triples exist — which would settle an old question of Erdős in the negative — and listed the complete evidence for it: 18 such triples below ten trillion (1014).
This note reruns that search a hundred thousand times further, out to 1019, and finds 346 triples instead of 18. But most of that growth is an illusion. Stretch a triple by the right whole number and it stays powerful, so almost every new hit is a rescaled copy of something already on the list. Strip the copies away and only 16 genuinely different triples remain below 1019. The figure below is the whole story in one picture.
Running an exact census five orders of magnitude past the published table:
Everything below is the technical version. Section 2 has the two proofs; section 3 has the full census.
A powerful number is a positive integer n such that p² | n for every prime p | n. Van Doorn (arXiv:2605.06697, May 2026, secondary) conjectured that infinitely many three-term arithmetic progressions consist of three consecutive terms of the powerful-number sequence — which would answer a question of Erdős in the negative — and reported the complete list of such triples below 1014: there are 18, the smallest being (1728, 1764, 1800).
We extend the census five orders of magnitude, to 1019: there are exactly 346 such triples, and they collapse, under componentwise rational scaling, to 16 primitive triples. Each scaling class has a canonical gcd-1 root, an integer arithmetic progression that need not itself be powerful, and we prove the exact criterion (Lemma 1′) for which multiples of a root are componentwise powerful. Beyond the published horizon the census shows a primitive triple none of whose elements is a perfect square (the other fifteen contain exactly one), three primitives with odd common difference, no chain saturated for long (the dominant chain, 272 of the 346 triples, first loses a multiple at m = 288), and no four-term progression of consecutive powerful numbers below 1019. Every triple is verified by an independent implementation; the sweep reproduces the published powerful-number counts at 1010–1012 and van Doorn's 18 at 1014.
A positive integer n is powerful if minp|n vp(n) ≥ 2 over primes p dividing it; equivalently n = a²b³ with b squarefree, and this representation is unique. Write the powerful numbers in increasing order as q₁ = 1 < q₂ = 4 < …
A (consecutive) triple is three neighbours (qi, qi+1, qi+2) with qi+1−qi = qi+2−qi+1 = d: three powerful numbers, evenly spaced, with no powerful number squeezed between them. Two triples are equivalent if one is a componentwise rational multiple of the other. Within the census below X, the smallest member of a class is its primitive; the class itself is its chain.
The root of a class is T₀ = T / gcd(T) for any member T (independent of which member is chosen). T₀ is an integer arithmetic progression with coprime-as-a-tuple entries, in general not itself powerful; every member of the class is m·T₀ for a positive integer m (a rational c = p/q with q > 1 cannot map a gcd-1 integer tuple to an integer tuple).
The conjecture (van Doorn, arXiv:2605.06697, secondary): there are infinitely many consecutive triples.
Everything in this section is exact and elementary; both proofs are short enough to check by hand.
Let T₀ = (w₁, w₂, w₃) be any triple of positive integers and m ≥ 1. Then m·T₀ is componentwise powerful if and only if
(1) every prime p with vp(wi) = 1 for some i divides m, and
(2) for every prime p | m, vp(m) ≥ 2 − mini vp(wi).
Proof. m·wi is powerful iff for every prime p | m·wi, vp(m) + vp(wi) ≥ 2. If p ∤ m, the condition reads vp(wi) ∈ {0} ∪ [2,∞) for each i; a prime with some vp(wi) = 1 therefore forces p | m, giving (1). If p | m, then p | m·wi for all i, and the condition over all i is exactly vp(m) ≥ 2 − mini vp(wi), giving (2). Conversely, (1) and (2) cover every prime dividing any m·wi. ∎
A perfect-square k is admissible for every already-powerful triple (the special case wi themselves powerful, condition (1) is vacuous and (2) always holds for even exponents). Scaling by an admissible multiplier always preserves the arithmetic-progression property, but not automatically consecutiveness: a stray powerful number can land inside the scaled window. Which admissible multiples actually stay consecutive is exactly what the census in section 3 decides.
If a consecutive triple (x, y, z) with common difference d contains two perfect squares, then d > √x.
Proof. Say s² < t² are two of the elements, so t ≥ s+1 and t²−s² ∈ {d, 2d}. Since s² ≥ x, t²−s² ≥ (s+1)²−s² = 2s+1 > 2√x, hence d > √x (indeed d > 2√x unless the two squares are the outer pair x, z). ∎
Corollary. A consecutive triple with d ≤ √x contains at most one perfect square. Fourteen of the sixteen primitives below 1019 satisfy d < √x (exceptions: P7, d/√x ≈ 1.82, and P14, ≈ 1.34), so for those fourteen "at most one square" is a theorem about that specific triple, not merely an observation.
Both lemmas are proved statements, labelled PROVED. Their consequences for specific triples — which multipliers actually occur, which primitives actually contain a square — are exact finite computations, labelled CERTIFIED in section 3. Neither computation is described here as a proof of anything beyond the range it covers.
The census reaches 18 triples at 1014 (matching van Doorn's table exactly, the calibration gate for this whole run), 152 at 1018, and 346 at 1019. Full decade-by-decade counts, for both raw triples and genuinely-new primitives, are the figure in the opening section above. Full list: data/census_1e19.txt; chain decomposition: data/analysis_1e19.txt.
Ordered by size. The root columns show T₀ (first element only) and its common difference d₀ = d / gcd(T); the multiplier column is the smallest admissible m taking that root to the primitive. A dagger marks the three primitives with odd common difference (Result C6).
| # | x (first element) | d | root, first elt. | d₀ | multiplier | squares |
|---|
(1728, 1764, 1800) = 36·(48, 49, 50) is P1: the smallest primitive, and the one van Doorn's own smallest triple belongs to.
Fifteen of the sixteen primitives contain exactly one perfect square. P15, at 414,781,000,521,010,200 ≈ 4.15×1017, is the unique primitive below 1019 with no square element: the first known consecutive triple of this type. ("At most one" is forced by Lemma 2 for the fourteen primitives with d < √x; "at least one" is simply false in general, as P15 shows.)
No four consecutive powerful numbers below 1019 form an arithmetic progression.
In root coordinates: P1's class is generated by (48, 49, 50), the smallest consecutive triple being 36·(48, 49, 50). A root with d₀ = 1 is a scaled arithmetic progression of consecutive integers; by the Erdős–Mollin–Walsh conjecture (open, secondary) the root itself should never be powerful, and indeed no root in the census is (checked for all 16, directly from the definition).
No chain stays consecutive for every admissible multiplier. The dominant chain (272 of the 346 triples, primitive P8, whose gap is 0.5% of the local mean powerful-number gap) keeps every admissible multiple below m = 288 consecutive, then loses m = 288 (root-relative; 72× the primitive) to an intruder. The figure below shows exactly this chain. The smallest failure anywhere in the census is P1's own: 2·(1728, 1764, 1800) = (3456, 3528, 3600) is an arithmetic progression of powerful numbers, but the powerful number 3481 = 59² lands inside it, so this particular doubling is not a consecutive triple.
One chain is saturated across the whole range checked: P6, root (28440875, 28440882, 28440889) with d₀ = 7 — all 3 admissible multipliers m ≤ 200,000 occur as consecutive triples, with no gap yet found. P15's chain contains members at ratios 5/2 and 45/2 to its own primitive: consecutive triples in one class need not be integer multiples of the smallest triple in it, only of the root (P15's observed multipliers of the root are 1800, 4500, 5400, 9000, 16200, 40500). Every observed multiplier satisfies Lemma 1′ (mechanically re-checked against every one of the 346 triples; zero exceptions).
Three primitives (P3, P10, P12) have odd d. The earliest, P3 at 1.49×1011, is within van Doorn's own published range; whether his paper remarks on the parity of d is not something this session could check (his PDF is unreachable, section 6).
Enumeration: every powerful number is uniquely a²b³ with b squarefree, so all values below X are generated by sieving squarefree b ≤ X1/3 and looping a ≤ √(X/b³), then sorted and scanned for adjacent equal gaps; a segmented variant carries a 3-element window across segment boundaries for large X. All arithmetic is 64-bit unsigned integer with 128-bit overflow guards. No floating point appears in any critical path (the "tightness" ratios quoted in section 3 are display cosmetics computed after the fact, not part of the search; 1019 < 264).
Evidence for completeness of the sweep:
Cost (single thread, 4-core / 15 GB box): census to 1018, 3m29s; to 1019, 7m52s; monolithic run to 1017, 77s / 5.6 GB. Verification of all 346 triples: 1m51s. Chain analysis: 1.2s. No randomness anywhere in the pipeline.
Raw triple counts are dominated by the multiplier supply of one tight chain (P8 alone contributes 272 of 346), so mostly they measure Lemma 1′ arithmetic, not new events. The informative curve is the primitive count: 6 below 1014, then 8, 10, 11, 15, 16 across 1015…1019. New primitives keep arriving across the entire observed range, consistent with van Doorn's conjecture and five decades more evidence than its published basis, but this is evidence, not proof: no growth model is fitted here, and sixteen data points would not usefully constrain one.
No primary source was read for this note. arxiv.org, oeis.org and every academic mirror attempted returned a proxy-level refusal (HTTP 403 on CONNECT) from the machine that ran the census; a later attempt to reach oeis.org directly from the machine that built this page returned the same 403. Everything attributed to the literature below rests on search-engine summaries and must be treated accordingly.
The basis for every novelty claim on this page — that the census beyond 1014 is new, that the 16-primitive structure is not already in van Doorn's paper under a different name — rests on search summaries only. If a primary source shows otherwise, the extension is a rediscovery and should be relabelled; that would not affect its correctness, only its novelty, and is flagged as the single most important thing to check next.