The Radix Riemanniana
The explicit formula connects the zeros of an L-function with the primes. If one looks only through a window [−L, L], a finite piece of prime data is enough to say something exact about the zeros. The Radix Riemanniana is the root of that window: the reproducing kernel of the Weil form, built from the Gamma factor, the conductor and the primes alone, without knowing a single zero.
For an elliptic curve E over ℚ and a window [−L, L], with G the Gram matrix of the Weil form of E and Φ(z) the Fourier basis on the window,
What it solves
How large is the rank of an elliptic curve? This question lies behind the conjecture of Birch and Swinnerton-Dyer. Assuming the Generalized Riemann Hypothesis, the explicit formula bounds the analytic rank from the traces ap. Until now this bound was computed with a fixed test function. The Radix Riemanniana yields the best possible one in the window, separately for every curve:
The Radix in three dimensions
How it is checked
Every bound is certified with interval arithmetic (Arb), so rounding cannot produce a false result. As an independent cross-check, the same quantity is computed a second way, from the actual zeros of L(E, s). Both agree, and the gap halves with each doubling of the zeros included. Of more than fifteen thousand certified bounds, not one falls below the known rank.
Results
| Sample | Curves | Primes saved | Worse than fixed |
|---|---|---|---|
| Cremona tables, N < 4.1·10⁵ | 1642 | 8.3–15.2 % | 0 |
| Stein–Watkins, N up to 7.5·10⁹ | 2500 | 17.0–22.7 % | 0 |
| Single curves, N up to 5·10¹⁵ | 9 | 0–43 % | 0 |
Primes needed to determine the analytic rank, compared with the fixed Fejér-type kernel of Bober and of Klagsbrun–Sherman–Weigandt. The range depends on the stopping criterion. The gain grows with the conductor. Curves of rank 8 and 9 are among the test cases.
Record curve ICARM #4239. On 16 October 2026 the Radix gave the first upper bound for curve #4239 of the ICARM Elliptic Curve Rank Leaderboard, found on 15 October by Elkies, Klagsbrun and Li with 32 independent points. With all 2 529 822 599 primes below 6.02·1010, the certified bound is [33.99820448 ± 1.83·10−9] < 34. The root number is +1, so the analytic rank is at most 32 under GRH, and the rank is exactly 32 under GRH and BSD. Certificate: record_4239.md. Independently, on 19 October 2026, rankbound (R. Jain) obtained rank ≤ 32 by a different method, a bound on the 2-Selmer group.
What it does not claim
The bounds assume GRH and concern the analytic rank. Weil Gram matrices as such are known (Yoshida; Connes–Consani–Moscovici; Suzuki; Shi). What is new is the per-curve optimisation from the actual ap, its certification, and the systematic comparison. For the record curve #4239 the bound holds under GRH; that the rank is exactly 32 also needs BSD.
Sealed
The archived release is fixed twice: by a DOI at Zenodo, and by a timestamp in the Bitcoin blockchain through OpenTimestamps. Anyone can check that the files below existed unchanged on this date.
- DOI
- 10.5281/zenodo.23299461
- Archive radix-rank-0.3.0.zip · SHA-256
- 7b5ae83e5718485143c066fd5274b2b966812c2a7a7043e8b3f6abe89611b91b
- Paper radix_rank_v0.3.0.pdf · SHA-256
- 8a0b8a65658549944a0236a46628431ce35e28682d5f38306d2654a50ce48064
- Release commit
- 519e70cfb39adf1e8bfaa36483a59d3cfb0cd250
- Bitcoin timestamp
- submitted 11 October 2026 to four OpenTimestamps calendars; block confirmation pending
Optimality theorem
A proof published on 25 October 2026 (preprint) shows: assuming GRH, the Radix bounds converge to the best bound that any test function supported in [−2L, 2L] with non-negative Fourier transform can give, and the limit is attained by the limiting radix kernel, with an exact error identity. No choice of test function in this class can do better.
Note: Optimal explicit-formula bounds for analytic ranks: the limiting radix kernel, doi:10.5281/zenodo.23649057.
Paper & Code
Paper: radix_rank.pdf · Code, data and all certificates: github.com/sectio-aurea-q/radix-rank, archived as doi:10.5281/zenodo.23299461. One command re-derives and re-certifies seven reference curves:
git clone https://github.com/sectio-aurea-q/radix-rank cd radix-rank && ./reproduce.sh quick