Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.
Updated
Erdős 699: computations and reproducibility
Start with the mathematical proofs at seven and nine. Their combined checkers join the numerical infinite tails to the finite exclusions:
With the pinned dependencies
installed in the repository's .venv/, run these commands from the
repository root:
.venv/bin/python wiki/research/erdos_699/computations/erdos_699_index_7_complete.py
.venv/bin/python wiki/research/erdos_699/computations/erdos_699_index_9_complete.pyThese commands check the retained proof records without replacing them.
They verify numerical source bounds, all retained initial interval
inequalities, exact polynomial coverage, the joint interval bindings, and
the overlap with the finite range. They reuse the interpolation ranks.
By default each then regenerates every joint witness sequence, 9,811,438
witnesses at seven and 11,599,987 at nine, and compares the count and
sequence digest of every direction with the retained joint record. The
replay builds the directed prime-log cache through under the
ignored repository directory tmp/erdos699/ on first use, about 2.3 GB
in under a minute, and the regeneration takes roughly an hour of processor
time per index, under ten minutes of wall time with the default eight
worker processes (--workers) on the machine used here. --quick skips
the regeneration, checks only the source and endpoint bindings of the
retained joint record, and says so in its output; --compact-cache
relocates the cache.
Longer replays
The seven and
nine joint-source drivers accept
--check for their retained source and endpoint checks, and can be run
on their own as the combined checkers run them. They load the compact
prime-log cache that a combined checker builds on first use; build it
separately with erdos_699_compact_prime_prefix.py --limit 10000000000 --base-limit 100000000 --cache-dir tmp/erdos699/compact-prime-prefix-10000000000,
and pass another location with --compact-cache. Use a separate output
file under the ignored repository tmp/ when reconstructing sequences:
mkdir -p tmp/erdos699
.venv/bin/python wiki/research/erdos_699/computations/erdos_699_index_7_height_cover.py --replay --reuse wiki/research/erdos_699/computations/erdos_699_index_7_height_cover.json --output tmp/erdos699/erdos699-seven-replay.json
.venv/bin/python wiki/research/erdos_699/computations/erdos_699_index_9_height_cover.py --output tmp/erdos699/erdos699-nine-replay.jsonAt seven, --replay compares each reconstructed witness count and sequence
digest with the --reuse record. At nine, compare those fields in the
output with the retained record; the driver has no --replay option.
Execution times and other diagnostic fields need not match.
The initial-source certificates retain every witness. Their separate
drivers can additionally reconstruct all pointwise coefficient-gcd bounds:
seven does this by default,
while nine uses
--replay-gcd. Both accept --cache. The proof notes distinguish these
replays from the checks performed by the combined drivers.
How the files are kept
The scripts, source tables, proof records, and older experiments remain in one directory because they import one another and locate their inputs as siblings. The relocation preserved every computation file byte for byte. A retained record names its inputs by basename and carries no digest of them; a checker validates an input by loading it and checking its domain and coverage, never by a hash. The digests a record does carry are the expected digests of recomputed objects listed under "Computed-result digests" in the verification guide. Follow the links from the relevant mathematical note to find its supporting computation; this page is not an inventory of every intermediate file.
Some older docstrings and three historical JSON records retain the former
research/finite_experiments/ paths. These are preserved provenance;
the applicable readers resolve the data by sibling basename. For current
commands, use wiki/research/erdos_699/computations/ as above.
The pinned requirements cover the principal proof checkers; individual
older experiments may state additional dependencies in their docstrings.