NEURAL COMPUTATION: OPTIMAL PAIRWISE-MAX RELU REPRESENTATIONS
Michael Schroeder | Version 1.0 | 9 November 2026
DOI: 10.5281/zenodo.24051030
ORCID: https://orcid.org/0009-0004-3249-0195

THIS IS THE COMPLETE COMPACT DEPOSIT

main-paper.pdf                 Main paper, 26 pages.
technical-supplement.pdf       Integral technical supplement, 205 pages.
paper-source.zip              Editable TeX, direct exact checkers and inputs,
                              evidence routes and accepted verification records.
MAX5-MAX6-Lean-Compendium-2026-11-08.zip
                              Complete Lean sources and finite proof data.

No large archive accompanies this Zenodo deposit. The historical PM2/MAX5
research archives and the macOS offline runtime are retained locally. Their
names and hashes in EXCLUDED-ARCHIVES.json record provenance, not extra files
to obtain for the commands below. The complete Lean compendium is included;
no proof case, hypothesis or finite proof datum has been removed for size.

The six keywords are: ReLU neural networks; Neural network expressivity;
Maximum function; Sparse representations; Lower bounds; Lean formalization.

LICENSES

Papers, manuscript sources and publication documentation: CC BY 4.0.
Author's Lean proofs, code and supporting software data: Apache 2.0.
Existing third-party licenses and attributions are preserved. LICENSES.txt
specifies file scope; licenses/ contains both full license texts. The licenses
cover different components, rather than offering a choice for the same work.

VERIFY THIS ZIP (Python 3.10+; standard library only)

Extract the ZIP into a new directory. From that directory:

  python3 -I -B verify.py

To run all included direct checks and reconstruct every Lean source, choose
a NEW work directory outside this extracted package:

  python3 -I -B verify.py --reproduce --work ../max5-max6-check

The second command verifies paper-source.zip, checks references and proof
correspondence, checks both exact upper identities and the MAX5/MAX6 networks,
checks the outer geometry, regenerates byte-identical manuscript sources,
and reconstructs all 26,201 Lean files. It needs no historical archive or
network access. Results and stage logs are written to the new directory.

These commands do not execute Lean or repeat the archived global lower-bound
programs. File hashes establish identity, not mathematical proof. The accepted
Lean execution records and the present packaging checks are distinguished in
VERIFICATION.json and the preserved records within the compendium.

LEAN SOURCE COMPENDIUM

Unzip MAX5-MAX6-Lean-Compendium-2026-11-08.zip. From its extracted directory:

  python3 -I -B tools/compendium.py audit
  python3 -I -B tools/compendium.py expand ../MAX5-MAX6-expanded

Expansion restores 10,143 MAX5 and 16,057 MAX6 project modules plus the MAX6
scope-projection module. Keep max5 and max6 separate: their PM2 namespaces
contain different definitions. Each chapter container retains the original
module boundaries; it is not one merged compilable module.

Begin with sources/READING_GUIDE.txt and the model/final-theorem entry files.
Lean 4.32.2 and the exact external package revisions in DEPENDENCIES.json must
be installed separately to compile the sources. This is a complete project-
source deposit, not a bundled compiler or a zero-download Lean environment.
Historical receipts and the preserved compendium README may refer to the large
offline package; that package is not supplied by this deposit. The current
outer README and EXCLUDED-ARCHIVES.json define the deposited file set.

VERIFICATION AND MATHEMATICAL SCOPE

The accepted MAX6 checkpoint is 31 October 2026: a complete source rebuild of
all 16,057 project modules, independent kernel replay, and final scope and
completion audits. MAX5 has its separate accepted kernel verification. Records
of the 8 November isolated offline replay tests are retained. This release
applies licenses and clarifies availability; it does not rerun the Lean proof
or claim a new complete source rebuild.

The formalized optima W*(5,2)=6 and W*(6,2)=33 concern rational coefficients,
identities on every real input, the fixed pairwise-max dictionary, distinct
nonparallel hinge directions, and a free feature-linear term. The paper's
real-coefficient extension is a separate written rationalization argument.
These are not unrestricted neural-network neuron-optimality theorems.

The current paper-source evidence map is evidence/lean-correspondence-r21.json.
The adoption register covers 35 nodes and 122 historical sources. Follow the
completed dated records, not superseded status comments in preserved sources.
Archive-dependent commands in the historical documents require those excluded
archives; they are not the compact-deposit reproduction commands above.

CITATION

CITATION.bib supplies the paper citation. Both manuscripts carry Version 1.0,
the same DOI and the six keywords. DESCRIPTION.html and zenodo-metadata.json
contain the deposit description and license scopes.
