Skip to content

About

Certified global lower-bound improvement for the Erdos minimum-overlap problem: c_E > 0.38055925 via independent Arb and MPFI checks; the exact value remains open.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

21 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Erdos Minimum-Overlap Lower Bound

DOI

Project Overview

Field Value
Author Ruturaj R Raval
Affiliation Independent Researcher
ORCID 0000-0003-4930-8981
Field Additive combinatorics and rigorous harmonic analysis
Problem Determine the Erdos minimum-overlap constant c_E
Current result Certified global lower bound c_E > 0.38055925
Result type Project-original global lower-bound improvement
Release v0.3.1
Version DOI 10.5281/zenodo.22647743
Concept DOI 10.5281/zenodo.22260847
License Apache-2.0 for project-original material

This repository proves the certified lower bound

c_E > 0.38055925

for the Erdos minimum-overlap problem. The project-original contribution is an even replacement certificate for the two central mean bins. The other 170 mean bins use Liam Price's previously publicly released Arb-certified bounds. The exact value of c_E remains open.

Release v0.3.1 is an archival and documentation patch. It adds an explicitly named compiled paper PDF, a deterministic paper-source archive, and SHA256SUMS. The theorem, proof, certificates, data, and computations are unchanged from v0.3.0.

Problem And Background

Partition

{1, 2, ..., 2N} = A disjoint-union B

into two sets of size N. For an integer shift d, count the pairs

(a, b) in A x B with b - a = d.

Let R(A, B) be the largest such count over all shifts, and minimize it over all balanced partitions. The asymptotic constant is

c_E = liminf as N tends to infinity of
      (1/N) min over balanced partitions R(A, B).

The question is to determine c_E. In plain language, every balanced two-coloring of consecutive integers has a translation where many opposite-colored points coincide. The constant measures the smallest unavoidable peak overlap.

Paul Erdos posed the problem in 1955 and conjectured that the constant was 1/2. The exact value has resisted structural and computational methods for more than seventy years.

Starting Frontier And Longstanding Gap

Selected lower-bound developments are:

Year Result Status
1959 Moser proved the classical lower bound 0.35639395869... Published
2022 White raised the lower bound to 0.379005 Preprint
2026 Chung et al. released a certificate for c_E > 0.380552 Preprint and Apache-2.0 artifacts
2026-07 Price publicly released a repository certificate for c_E > 0.38055470 Public computational result
2026-09-03 Deng published a certified result c_E > 0.380557 Repository, preprint, and Zenodo archive
2026-09-04 This project certified c_E > 0.38055925 This repository

The dated 2026-09-04 frontier audit compared the project target with a strongest located stated claim of c_E > 0.380557 and a strongest located recorded certificate capability of 0.3805581795278582246. The new certified target exceeds that recorded capability by 0.0000010704721417754, or approximately 1.07047e-6.

The corresponding rounded comparison point is approximately 0.380558179527858....

Recent upper-bound constructions remain near 0.380856, so a numerical gap of about 0.00029675 remains. The exact constant and the remaining upper-bound gap are open.

Main Result

The manuscript in paper/main.tex proves:

Theorem. The Erdos minimum-overlap constant satisfies c_E > 0.38055925.

The proof separates the mean of the continuous overlap density into 172 bins.

Mean region Certificate used Verified denominator upper
Central bins 85 and 86, ` mu <= 1/320`
Other 170 bins Price's publicly released certificate 2.627538530873375791
Required threshold 1 / 0.38055925 2.627711716375308181...

Both denominator bounds are strictly below the threshold. The finite maximum over all 172 bins is therefore also strictly below the threshold, which gives the uniform strict lower bound.

Method And Proof Architecture

For the central mean range |mu| <= 1/320, the replacement certificate uses:

  • one exact second-moment inequality;
  • 107 pointwise Fourier inequalities C_xi <= sinc(xi)^2;
  • one Parseval energy inequality truncated at order 100;
  • nonnegative exact-decimal multipliers.

These rows define an even function

q(t) = 1
     + lambda_2 (B_2 - t^2)
     + sum_xi lambda_xi (sinc(xi)^2 - cos(xi t))
     + lambda_P (1/2 + sum_{n=1}^{100} cos(n pi t)).

The positive-part dual principle gives

||p||_infinity >= 1 / integral max(q(t), 0) dt.

The certificate file is certificates/center-038055925.tsv, with SHA-256:

b02a45a645337c74215a365e82f403990eeb9413e3f8771e719e5e5397da39e8

Frequencies and multipliers are interpreted as exact rational decimals. Each sinc(xi)^2 right-hand side is the exact transcendental value, enclosed independently by directed arithmetic in each verifier.

Price's previously released certificate supplies the other 170 mean bins. Those inputs are hash-pinned and replayed at the same target without redistributing unlicensed material. The finite center and noncentral certificates together establish the theorem.

Verification And Evidence

The center certificate is checked by two separately implemented directed-arithmetic verifiers:

  • Python with python-flint and Arb;
  • C with MPFI, MPFR, and GMP.

The Python verifier:

  • parses the canonical certificate without floating-point conversion;
  • computes the analytic bounds with python-flint and Arb;
  • encloses q, q', and q'' on adaptive cells using Taylor's theorem;
  • integrates certified-positive cells with an interval antiderivative;
  • charges unresolved terminal cells by a rigorous upper rectangle.

At 256-bit precision it reports:

D_upper = 2.627711172296609115765958739218722735...
margin  = 0.000000544078699065548777402393917083...

The independent C verifier has its own:

  • strict LF-ASCII parser;
  • exact decimal-to-rational conversion;
  • MPFI and MPFR interval calculations;
  • adaptive traversal and accounting;
  • endpoint extraction and strict target comparison.

At 256-bit precision it reports:

D_upper = 2.627711172296609115765958701268201763538
margin  = 0.0000005440786990655487774403444380553601380208

The implementations agree beyond the precision needed for the theorem. Their shared boundary is limited to the certificate bytes, analytic formulas, Taylor enclosure argument, and elementary antiderivative. Repeated runs of one implementation are not treated as independent verification.

The complete machine-readable record is in:

Exact commands, evidence records, hashes, and the trust boundary are also documented under docs/. The theorem replay is CPU-only on a commodity workstation. No exploratory optimization run is needed to verify the result.

Reproduction

Prerequisites:

  • Python 3.12
  • uv
  • a C compiler
  • GMP, MPFR, and MPFI development libraries

Run the full repository checks:

make sync
make test
make lint
make typecheck
make verify
make audit
make paper-release
make verify-release-assets

Run the center checks directly:

uv run minoverlap verify-center \
  certificates/center-038055925.tsv

uv run minoverlap verify-center-mpfi \
  certificates/center-038055925.tsv

The C checker requires MPFI, MPFR, and GMP. The Python checker requires Arb through python-flint.

Resource profile: verification is CPU-only and fits an ordinary workstation. The directed-arithmetic checks need no GPU and no exploratory optimization run.

Claims

  • The frozen project center certificate rigorously covers both central mean bins.
  • Two independently implemented directed-arithmetic verifiers accept it.
  • Price's other 170 bins were replayed at the same target and all passed.
  • Together these finite certificates establish c_E > 0.38055925.
  • The center certificate and verification implementations are project-original work by Ruturaj R Raval.
  • The result is a global lower-bound improvement, not merely a bounded search or a local numerical observation.

Limitations And Nonclaims

  • The project does not determine the exact value of c_E.
  • The project does not prove that the retained multipliers are optimal.
  • Price's noncentral certificate is not claimed as project-original work.
  • Price's unlicensed source or certificate files are not redistributed.
  • Repeated runs of one implementation are not treated as independent verification.
  • The result depends on the stated finite analytic reduction and its explicit trust boundary.
  • Independent external mathematical review is not yet complete.

Significance And Use

Every lower-bound improvement strengthens a universal theorem over all balanced partitions of every sufficiently large interval. It rules out an additional range of hypothetical low-overlap constructions.

The project also demonstrates a reusable computer-assisted proof pattern:

  • separate optimization from verification;
  • freeze exact certificate semantics;
  • require independent arithmetic implementations;
  • bind evidence by cryptographic hashes;
  • preserve dependency and licensing boundaries;
  • state numerical and mathematical nonclaims explicitly.

The same architecture applies to extremal combinatorics, Fourier inequalities, rigorous optimization, and other finite-certificate proofs.

Remaining Work And Future Directions

The main mathematical problem remains open. High-value next directions are:

  1. Continue reduced-cost column generation and test whether the current Fourier relaxation has a structural ceiling.
  2. Add coupled-frequency or positive-semidefinite constraints.
  3. Narrow the remaining gap to the best upper construction.
  4. Seek independent external reproduction and peer review.
  5. Formalize the finite analytic reduction and certificate semantics in a formal proof system.

Any stronger theorem must exceed 0.38055925, cover all 172 mean bins, pass both independent directed-arithmetic verifiers and clean-checkout replay, refresh the prior-art audit, and preserve the explicit dependency boundary.

Repository Layout

Path Purpose
certificates/ Frozen project certificate
src/minoverlap/center_certificate.py Python-Arb center verifier
verification/center_mpfi.c Independent MPFI/C center verifier
paper/ Proof manuscript and publication sources
evidence/ Machine-readable verification and provenance records
docs/ Prior-art, proof, trust-boundary, and search analysis
upstream/station/ Licensed Station reproduction baseline
research/ Claim and release-gate records
CITATION.cff Citation metadata
RELEASE_NOTES.md Version history

Publication Citation And Archive

  • Public repository: ruturajr-raval/erdos-minimum-overlap-lower-bound.
  • Tagged release: v0.3.1.
  • Release commit: 965fcdec86a3232ed4895e3180c9c7b3d1a85b77.
  • Version archive: Zenodo DOI 10.5281/zenodo.22647743.
  • Concept archive: All versions are collected under Zenodo DOI 10.5281/zenodo.22260847.
  • Manuscript: paper/main.tex.
  • Paper assets: erdos-minimum-overlap-lower-bound-v0.3.1-paper.pdf, erdos-minimum-overlap-lower-bound-v0.3.1-paper-source.tar.gz, and SHA256SUMS.
  • Patch scope: The PDF and source bundle are newly explicit archival assets. The mathematical and computational content is unchanged.

Suggested citation:

Raval, Ruturaj R. (2026). The Erdos Minimum-Overlap Constant Exceeds 0.38055925 (v0.3.1). Zenodo. https://doi.org/10.5281/zenodo.22647743.

Machine-readable citation metadata is in CITATION.cff, and release history is in RELEASE_NOTES.md.

Authorship

Ruturaj R Raval, Independent Researcher.

ORCID: 0000-0003-4930-8981.

The replacement certificate for bins 85 and 86 and both center verifiers are project-original work by Ruturaj R Raval.

Licensing And Provenance

Project-original software, certificate data, and documentation are released under Apache-2.0. The retained Station baseline material preserves its Apache-2.0 upstream notices and provenance.

Price's noncentral source package is cited and hash-pinned but is not included because no license was declared at the audited commit. The project replays Price's certificate for the other 170 bins without claiming that material as original work and without redistributing the unlicensed source package.

References

About

Certified global lower-bound improvement for the Erdos minimum-overlap problem: c_E > 0.38055925 via independent Arb and MPFI checks; the exact value remains open.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages