Open computational mathematics. AI-audited, not peer-reviewed. All code and data open for independent verification.

by cahlen complete

Hardware

8x NVIDIA B200 (183 GB VRAM each) Intel Xeon 1.4 TB RAM
number-theorycontinued-fractionsprime-numberscomputational-mathematics b200 miller-rabincontinued-fractionsconvergent-recurrencemonte-carlo

Key Results

Problem
Verify and extend the Erdos-Mahler bound on greatest prime factors of CF convergents
Prior Work
Humphreys (2,013, NCUR/Boise State): Corollary 3.6 showed G(A_n) ≥ e^{n/(50 ln n)} for almost all irrationals
Status
First GPU runs complete. Bound verified for 10,000 random CFs. Doubly-prime convergent statistics collected.
Preliminary Findings
Erdos-Mahler bound holds 100% of the time across 10,000 random CFs to depth 38, Bound constant 50 is very conservative: worst-case ratio is 7.18x, average 117x, Average 4.9 prime numerators per CF, 4.95 prime denominators, Doubly-prime convergents: 0.95 per CF on average (about 1 in 40 terms), For e specifically: 3 doubly-prime convergents in first 42 terms (matches Humphreys 2,013)
Open Questions
Can the constant 50 be replaced by a smaller constant? Computational data suggests ~7 may suffice., Is the number of doubly-prime convergents always finite for algebraic irrationals?, What is the asymptotic density of doubly-prime convergents as depth grows?, Does e have exactly 3 doubly-prime convergents, or are there more beyond n=42?

Prime Convergents of Continued Fractions

Background

For an irrational number ζ=[a0;a1,a2,]\zeta = [a_0; a_1, a_2, \ldots], the nn-th convergent Cn=An/BnC_n = A_n / B_n has numerators and denominators that grow exponentially with nn. In 1939, Erdos and Mahler showed that for almost all ζ\zeta, the greatest prime factor G(Bn)G(B_n) of the denominator increases rapidly:

G(Bn)en/(50lnn)for all sufficiently large n.G(B_n) \geq e^{n / (50 \ln n)} \quad \text{for all sufficiently large } n.

Humphreys (2013, NCUR/Boise State) extended this to the numerator via Theorem 3.5 and Corollary 3.6, showing the corresponding result for G(An)G(A_n). The original experimental verification was limited to 500 convergents of nen \cdot e and nπn \cdot \pi for n500n \leq 500, computed in Maple.

What This Experiment Does

This GPU experiment massively extends the computational verification:

  1. Random CF sampling: Generate hundreds of thousands of random CF expansions (partial quotients drawn from the Gauss-Kuzmin distribution) and compute all convergents until overflow
  2. Primality testing: GPU-parallel deterministic Miller-Rabin (12 witnesses, valid for all 64-bit integers) on every numerator and denominator
  3. Greatest prime factor: Trial division by primes up to 997 plus Miller-Rabin on the cofactor
  4. Doubly-prime census: Count convergents where both AnA_n and BnB_n are prime — extending the observation from Humphreys (2013) Section 4

Each GPU thread handles one complete CF sequence independently. The B200’s 183 GB VRAM allows millions of simultaneous sequences.

Method

  • Convergent recurrence: An=anAn1+An2A_n = a_n A_{n-1} + A_{n-2}, Bn=anBn1+Bn2B_n = a_n B_{n-1} + B_{n-2}, computed in 128-bit arithmetic (overflow at depth ~38-42 for typical CFs with Gauss-Kuzmin partial quotients)
  • Primality: Deterministic Miller-Rabin with witnesses {2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37}, sufficient for all n<3.317×1023n < 3.317 \times 10^{23}
  • GPF computation: Trial division by first 168 primes (up to 997), then primality check on remainder
  • Modes: Random Gauss-Kuzmin CFs, CF of ee (known pattern), CF of π\pi (50 known terms + random)

Reproduce

nvcc -O3 -arch=sm_90 -o prime_convergents scripts/experiments/prime-convergents/prime_convergents.cu -lm
./prime_convergents 100000 500 0   # 100K random CFs, depth 500
./prime_convergents 1 500 1        # CF of e, depth 500
./prime_convergents 1 500 2        # CF of pi, depth 500

References

  1. Erdos, P. and Mahler, K. (1939). “Some Arithmetical Properties Of The Convergents Of A Continued Fraction.” London Math. Soc., pp. 12-18.
  2. Humphreys, C. (2013). “Prime Numbers and the Convergents of a Continued Fraction.” Proceedings of NCUR 2013, University of Wisconsin La Crosse.
  3. Gauss, C.F. (1812). Disquisitiones generales circa seriem infinitam.

Extends Humphreys (2013) with GPU computation. Human-AI collaboration (Cahlen Humphreys + Claude). All code and data open.

Recent Updates

findingUpdate certifications and finding metadata from review cycle
findingMark Zaremba density experiment as complete
updateAdd Convergent-7B model showcase to front page
updateTighten language: empirical observations are not laws or theorems
experimentUpdate Ramanujan Machine: v1 exhausted (7K false positives), v2 kernel built
findingUpdate README: 18 findings, 53 reviews, 7 models, 3 providers
infraMCP server: fetch manifest from GitHub instead of bundling a copy
infraUpdate MCP server manifest: 207/210 issues resolved
updateUpdate stats: 207/210 issues resolved (98.6%), up from 191 (91%)
reviewFix stale review counts in llms.txt, llms-full.txt, meta.json, certifications.json