ERDŐS/DAILY

← back to the ledger

ERDőS #276 · PARTIAL

Erdős problem 276 — wave8q

Accessed and computed on 2026-07-28. The conclusion is PARTIAL, not a

solution: I obtain an exact finite obstruction for the Ismailescu--Son

construction and isolate the missing uniform lemma.

Claim labels used throughout:

0. Mandatory live-page gate

I fetched the rendered live problem page

and its discussion thread

through the Bright Data browser, not datacenter curl.

marker, not the current-worker stop condition.

formalised.

The mandatory stop conditions were therefore absent.

The live wording is reproduced below. To respect the short-quotation limit,

the verbatim excerpt ends at its last substantive noun; its completion is

given immediately afterward by the exact quantified formula.

> Is there an infinite Lucas sequence \(a_0,a_1,\ldots\) where

> \(a_{n+2}=a_{n+1}+a_n\) for \(n\geq 0\) such that all \(a_k\) are composite,

> and yet no integer has a common factor with every term

The completion means exactly

\[ \exists (a_k)_{k\geq0}\quad \bigl(a_{k+2}=a_{k+1}+a_k,\;a_k\text{ composite}\bigr) \quad\text{and}\quad \forall M>1\ \exists k\geq0:\gcd(M,a_k)=1. \]

This reading agrees with the linked

formal statement,

which uses ∀ n > 1, ∃ k, Nat.gcd n (a k) = 1.

Live known results and comments

\[ a_0=1786772701928802632268715130455793,\qquad a_1=1059683225053915111058165141686995. \]

Graham's 1964 primary paper

supplies the 18 covering triples and these starting values.

their sequence is provably all-composite, but they do not prove that it

lacks a finite covering set of primes.

covering-based pair

\((106276436867,35256392432)\), then accurately separates the two parts of

Ismailescu--Son: a finite cover for even indices and an algebraic

factorisation for odd indices. It explicitly says nonexistence of a finite

cover for the odd indices is unproved and asks whether Hough's ideas could

reduce it to finite computation.

problem is probably resolved and compares the missing step with a question

about \(5\cdot2^n+1\); it is not a proof claim.

be omitted from the question; the page notes that it was updated

accordingly.

1. Primary-source audit

1980 monograph, printed p. 27

asks whether an all-composite Lucas sequence can exist without an underlying

covering system.

example by a finite prime cover.

2004 paper

proves the smaller pair quoted in the live comment, again by a cover.

[A New Kind of Fibonacci-Like Sequence of Composite Numbers,

JIS 17 (2014), Art. 14.8.2](https://cs.uwaterloo.ca/journals/JIS/VOL17/Ismailescu/ism8.pdf),

is the directly relevant primary source. Its last paragraph reports 803

pairwise-coprime terms surviving a \(2\cdot10^6\) small-prime screen through

index 200000, but expressly says the finite-cover question remains

unproved. I treat “803” here as the paper's report, not as a recomputed

theorem.

JNT 128 (2008), 1916--1940,

explicitly asks whether the least prime factor of \(5\cdot2^n+1\) is

unbounded, confirming that the analogous uniformity step is genuinely

difficult.

Annals theorem

proves that the least modulus of a distinct-moduli covering system is at

most \(10^{16}\). The distinctness qualifier is essential below.

2002 paper, found during

the search, combines algebraic factorisation and a cover for a class of

recurrences, but its theorem assumes recurrence coefficient \(a>2\), so it

does not include the Fibonacci recurrence \(a=1\).

[c] Exact-title, author, citation-graph, and full-text searches found no

primary post-2014 paper resolving this specific no-cover question. OpenAlex

listed one direct citation, Lunev's

2017 Tribonacci generalisation,

which does not address the missing claim; a Semantic Scholar API query

returned HTTP 429. This is a reported search miss, not a claim that no

uncatalogued result exists.

2. Reconstructing the Ismailescu--Son candidate

Let

\[ \begin{split} q={}&129511502555081082458723990740612592095319437933512025195406541068\\ &394745828231264515958532145970461367703231950382110924410768870, \end{split} \]

and define

\[ x_0=1+q^2,\qquad x_1=q^2+2q,\qquad x_{n+2}=x_{n+1}+x_n. \]

A source-level ambiguity resolved

The journal TeX places -\\ between the two displayed digit blocks for \(q\).

Literal subtraction gives a negative 80-digit integer, contradicting the

paper's statement that \(q\) has 129 digits.

in the paper's Table 3 using CRT. The least nonnegative solution is exactly

the concatenated 129-digit integer above; it is smaller than the product of

the 30 prime moduli. The subtraction reading fails these checks.

\[ 4(q^2+1)-(2q-1)(2q+1)=5. \]

Since \(q\equiv0\pmod5\), it cannot divide \(x_0\), so the gcd is 1.

Why every term is composite

Write \(F_0=0,F_1=1\) and \(L_0=2,L_1=1\).

\[ x_{2k+1}=(F_k+qF_{k+1})(L_k+qL_{k+1}). \]

Both factors exceed 1. This follows by substituting

\(x_n=x_0F_{n-1}+x_1F_n\) and the standard doubling identities. The checker

verifies the identity with full integers for every odd index through 20000.

satisfy: \(p_i\) is prime, \(p_i\mid F_{m_i}\), their classes

\(r_i\bmod m_i\) cover every even residue modulo

\(\operatorname{lcm}(m_i)=5040\), and the two initial congruences force

\(p_i\mid x_n\) when \(n\equiv r_i\pmod {m_i}\). The checker verifies every

row from scratch, including deterministic 64-bit primality tests and all

2520 even residues.

\(764940961\mid x_0\). Thus there is no omitted \(n=0\) case.

Consequently this is a rigorously certified all-composite sequence. What is

not certified is the absence of a second finite prime set covering its odd

terms.

3. Exact reduction of the open part

Three equivalent missing statements

Let \(b_k=x_{2k+1}\), the odd-indexed subsequence. The even terms already have

the fixed 30-prime cover above.

[a] The following are equivalent:

1. no finite set of primes covers all \(b_k\);

2. the least prime factor of \(b_k\) is unbounded as \(k\) varies;

3. the \(b_k\) contain an infinite pairwise-coprime subsequence.

Proof:

\(\leq B\) would cover every \(b_k\). Conversely, a finite cover bounds the

least prime factor by its largest member.

pairwise-coprime choices, the union of their prime divisors is finite, so

some further term avoids it. The converse is immediate: one prime can hit

at most one member of a pairwise-coprime family.

Because the even terms have a finite cover, each of these is also equivalent

to the original sequence having no covering integer.

The induced congruence system

[a] Suppose a prime \(\ell\mid x_r\). Consecutive terms are coprime, and

the addition formula gives

\[ x_{r+t}=F_{t-1}x_r+F_t x_{r+1}. \]

Hence

\[ \ell\mid x_{r+t}\quad\Longleftrightarrow\quad \ell\mid F_t. \]

If \(z(\ell)\) is the rank of apparition of \(\ell\) in the Fibonacci

sequence, the zero indices of \(x_n\bmod\ell\) are therefore one residue class

modulo \(z(\ell)\). On the odd subsequence the zero indices are one class

modulo

\[ \mu_\ell=\frac{z(\ell)}{\gcd(z(\ell),2)}. \]

Thus a hypothetical covering integer produces an ordinary finite covering of

the \(k\)-line by these special residue classes. This is exact, but the moduli

need not be distinct.

\(z(31)=30\) and \(31\mid b_k\) for \(k\equiv4\pmod {15}\), whereas

\(z(61)=15\) and \(61\mid b_k\) for \(k\equiv8\pmod {15}\).

Both induced moduli are 15 but their residues differ. The checker verifies

these four assertions directly.

Therefore Hough's \(10^{16}\) least-modulus theorem cannot simply be invoked:

its hypothesis is distinct moduli. Even if it did apply, factoring Fibonacci

numbers up to the corresponding ranks would not be a practical “small”

check—\(F_{2\cdot10^{16}}\) has about \(4.18\cdot10^{15}\) decimal digits.

4. New exact finite certificate

Define the forbidden prime set

\[ \mathcal P=\{p:p\leq10^7\}\ \cup\ \{p_1,\ldots,p_{30}\}, \]

where the second set is Ismailescu--Son Table 2.

[d] Exact prefix result. Among all \(0\leq n\leq20000\), exactly the

following 59 terms avoid every prime in \(\mathcal P\):

123, 719, 1143, 1383, 1799, 1815, 1827, 1887, 2823, 3843,
4095, 4167, 4583, 4823, 5123, 5735, 5819, 6027, 6927, 8363,
8955, 9239, 9323, 9443, 9515, 9539, 10199, 10379, 10415, 10499,
10647, 10763, 11067, 11715, 11795, 11855, 11895, 12143, 12147,
12719, 13379, 13619, 13715, 14247, 14975, 15839, 15855, 15959,
16919, 16979, 17495, 18375, 18659, 18675, 18839, 19079, 19343,
19595, 19887

All 59 are odd-indexed and algebraically composite. Their 1711 pairwise gcds

are all exactly 1.

[a+d] Consequence. If an integer \(M>1\) has

\(\gcd(M,x_n)>1\) for every \(n\), choose for each of these 59 terms a prime

dividing both it and \(M\). Pairwise coprimality makes those 59 primes

distinct. The forbidden-prime check makes every one larger than \(10^7\) and

outside the 30-prime table. Therefore

\[ \omega(M)\geq59,\qquad M>(10^7)^{59}=10^{413}. \]

This does not prove that \(M\) cannot exist.

The journal paper's 803-term computation has a larger cardinality at cutoff

\(2\cdot10^6\). The present certificate is an independent, exactly

reproducible higher-cutoff result, not a claim to dominate the paper in every

parameter.

[d] Cross-check of two source claims. PARI/GP isprime proves that the

four factors associated with \(k=913,943\) have digit lengths

319, 320, 326, 326 and are prime. Hence \(x_{1827}\) and \(x_{1887}\) are

indeed semiprimes as reported by Ismailescu--Son.

5. Reproduction and independence checks

Run:

python runs/erdos276_wave8q_reverify.py

The standalone checker:

1. reconstructs \(q\) by CRT;

2. proves all 30 table moduli prime in the 64-bit range and verifies the

covering certificate;

3. enumerates all 664579 primes through \(10^7\);

4. scans every prime/term modular recurrence pair;

5. regenerates the full integer terms and checks all odd factorizations;

6. independently multiplies all small primes in a balanced product tree and

checks each survivor by a full-integer gcd against that primorial and all

30 table primes;

7. checks all 1711 pairwise full-integer gcds;

8. checks the four large primes with PARI/GP; and

9. hashes the indexed full terms as

19168f99acc0b4c3b6b655f6786218a557ccd2baceb179d8645c835b79bebf80.

The final clean run passed in 104.44 seconds. Its last substantive output was:

Independent primorial-gcd check: PASS
Full-integer pairwise gcds: PASS (1711 pairs)
PARI proven-prime checks at x_1827 and x_1887: PASS
ALL CHECKS PASSED in 104.44 seconds

6. Precise wall and cost

\((F_k+qF_{k+1})(L_k+qL_{k+1})\) is unbounded. Equivalently, its prime

divisibility classes do not contain a finite covering system; equivalently,

these products have an infinite pairwise-coprime subsequence.

distinct moduli, while the induced rank moduli here can repeat.

particular multiplicities

\(\#\{\ell:z(\ell)/\gcd(z(\ell),2)=m\}\), or directly force an uncovered

residue in this rank-residue family. No such applicable theorem was found.

\(20000\pi(10^7)\approx1.33\cdot10^{10}\) vector residue updates. At its

measured rate, a segmented scan through \(10^9\) and 200000 indices would

require roughly \(1.02\cdot10^{13}\) updates, about 22 core-hours. At a

realistic \$0.05--\$0.10 per core-hour this is roughly \$1--\$2 of raw CPU;

as a current anchor, Google's listed 56-vCPU c2d-highcpu-56 on-demand rate

is about \$0.0375 per vCPU-hour

(official pricing).

This also needs a few GB of working-memory traffic and engineering for segmentation.

It would still be only another finite lower bound, not a proof.

The obstruction is therefore mathematical uniformity, not the next finite

sieve range.

PARTIAL: the Ismailescu--Son candidate is re-certified, and any covering integer must contain at least 59 distinct primes above 10^7 (and outside its 30 covering primes), but ruling out every finite cover remains equivalent to the unproved unbounded-least-prime-factor lemma.

This is the AI working report, labelled by outcome — not an independently verified claim unless marked PROVED. ← ledger