back to table · edit · history · where entries came from · files · download
5342 bytes, as of the version from 2026-11-17 12:35 (current). Recorded here, not run.
# T733: rational widths and rational heat times
This extension adds 23 arguments to the original 46. It preserves every
original parameter key, decimal string, definition and formula. It is a
computation and validation record, not an independent editorial acceptance.
The current selection and its noncanonical cutoffs are in
`argument-selection.md` and `selection-plan.json`.
## Computation and precision
`generate.py` evaluates all reduced rational widths `a/b`, `1<=a<=b<=12`,
using the unconditional prime/archimedean explicit formula. The added slice
is `alpha=(a/b)*pi` for coprime `1<=a,b<=6`; here `t=alpha/(4*pi)=a/(4*b)`
is evaluated as an exact rational. Symbolic keys and their LaTeX labels
identify exact inputs, not rounded decimal approximations to pi.
The second slice uses FLINT's rigorously isolated, consecutively indexed
zeta zeros. The first 138 are below 300 and the next is above 300. The
isolation/counting algorithm certifies completeness and the critical-line
location of this finite prefix; this is not a global RH assumption.
Write `t=alpha/(4*pi)>0`. For every remaining zero in the critical strip,
`|exp(t*(rho-1/2)^2)| <= exp(t/4-t*Im(rho)^2)`. Corollary 1.2 of
[Hasanalizade–Shen–Wong](https://arxiv.org/abs/2107.06506) gives
`N(u) <= u^2/(2*pi)+0.3611*u+9.3675 <= u^2` for `u>=100`, hence also
for `u>=300`. Integration by parts, discarding the negative boundary term,
bounds the two-sided tail above `T` by
`4*t*exp(t/4)*integral_T^infinity u^3*exp(-t*u^2)du`, equal to
`2*exp(t/4-t*T^2)*(T^2+1/t)`. Multiplicities and possible off-line zeros
are included. `generate.py` propagates the finite-zero enclosures and this
tail through Arb/FLINT ball arithmetic.
All 69 values were computed at 512 bits. Exact rational endpoint comparisons
verify that each resulting ball lies inside the full interval represented by
its stored decimal, namely the written midpoint plus or minus one last-place
unit. Every decimal has 100 significant digits. The worst production
relative-radius bound retains more than 136.41 decimal digits, at `alpha=1`.
Reparsing the printed diagnostic balls gives the more conservative bound
135.71; these diagnostic strings are not substituted for the production balls.
The 46 old decimal strings are unchanged. `components.json` contains the
components and method per argument; `extension-results.json` records checks.
## Independent checks and their limits
- `check_values.py` recomputes all 69 intervals using certified zeros at
512 bits. This is independent mathematics and data for the 46 prime-formula
values, but only reproduction/implementation checking for the 23 new
zero-generated values. Every check passed. The two certified routes share
the Arb/FLINT arithmetic library.
- `check_pi_values.py` imports neither the production generator nor Arb.
It recomputes the first 30 zeta zeros with mpmath at both 160 and 200 decimal
digits and uses the 29 below height 100. All 23 new values lie within the
recorded intervals, and the two precisions agree relatively to better than
`1e-145`. The theoretical omitted-tail bound is less than `1e-150` times
the finite sum throughout this slice. The mpmath roots and rounding errors
are not certified: this is an independent numerical agreement check, not
another proof of the error bounds. Results: `pi-independent-results.json`.
- The prime/archimedean route also independently cross-checks `pi/6`, `pi/3`
and `pi/2`, with respective prime cutoffs 100000, 150333 and 2293473.
These enclosures lie inside the same stored intervals. The pi uncertainty
is propagated by ball arithmetic. Larger widths use zeros because this
deliberately elementary prime-tail bound makes the prime route expensive.
- All 23 new values were recomputed at 640 bits, as were the original
rational checkpoints `1/12`, `1/2` and `1`. All enclosures agree.
`check_routes.py` reproduces these controls and the three prime-formula
cross-checks, and checks exact rational endpoint containment for all 69
stored decimals; its output is `routes-results.json`.
- Known-answer Gaussian and mock-spectrum controls passed. Missing
multiplicities, incorrect heat-time normalization, omitted prime powers,
NaN and vacuous-enclosure controls were rejected. No output is promoted
to an exact value on the basis of agreement.
The executable checkers, numerical reports and `verify(sample=None)` serve
different purposes: mathematical cross-checking, evidence recording and
reproduction. None is an independent human review. All added values and
their associated metadata remain subject to the website's review process.
## Historical experiment scope
The retained `remainder-results.json` is the original bounded experiment
on the 46 rational widths: 22 denominator-at-most-eight training widths
and 24 denominator-nine-to-twelve withheld widths. Its two uniform-sign
guesses both fail; it makes no all-width or RH claim. `summary_check.py`
still reproduces that experiment and deliberately skips the new zero-only
rows, which do not contain its prime and archimedean components. This report
must not be read as an experiment on all 69 arguments. The original research,
runtime and skill reports describe the original build unless explicitly
updated; the earlier complete source/evidence set remains in revision history.