Michael Schroeder · Preprint · Version 1.0 · 9 November 2026 · ORCID 0009-0004-3249-0195

Neural Computation: Optimal Pairwise-Max ReLU Representations

DOI: 10.5281/zenodo.24051030

Abstract

The maximum function offers a concrete benchmark for the cost of exact neural computation. For five and six inputs in the pairwise-max model of side size two (PM2), we prove the exact directional widths W*(5,2)=6 and W*(6,2)=33. The features are all pairwise maxima, including the input coordinates; each hinge applies a ReLU to the difference between two sums of two features, with repetitions allowed. We count distinct nonparallel hinge directions and allow a free feature-linear term. Explicit rational identities attain both bounds on all real inputs. Optimality holds for arbitrary coefficient signs and magnitudes, without symmetry assumptions. The lower bounds exploit cancellation: every kink introduced where the maximum is linear must disappear from the sum. This condition yields exact linear equations. Small dependencies among binary kink profiles settle the five-input case; a canonical reduction that never increases support and a complete analysis of local support costs settle the six-input case. Both rational optimality theorems are fully formalized in Lean. A separate rationalization argument gives the same minima for real coefficients in the fixed dictionary.

Paper and Supplement

The paper presents the model, exact identities, lower-bound strategy and neural-network constructions. The supplement gives the full case analysis, computational certificates and correspondence with the Lean proofs.

Lean Proofs and Verification Materials

The compact release includes both PDFs, their LaTeX sources, direct exact checkers, the complete Lean source compendium, dependency pins, verification records and integrity checksums. Every proof case, hypothesis and finite proof datum is preserved.

Start with README.txt in the extracted package. With Python 3.10 or newer, run python3 -I -B verify.py to check file integrity. To run the included direct computational checks and reconstruct all 26,201 Lean files, use python3 -I -B verify.py --reproduce --work ../max5-max6-check, choosing a new work directory outside the extracted package.

These commands check the release and reconstruct its sources; they do not execute Lean. The nested compendium contains sources/READING_GUIDE.txt and source audit and expansion commands. Compiling the proofs requires Lean 4.32.2 and the exact external dependencies listed in DEPENDENCIES.json, installed separately. Compiler binaries, external libraries and compiled objects are omitted from this compact download.

Citation and file integrity

Schroeder, Michael. Neural Computation: Optimal Pairwise-Max ReLU Representations. Version 1.0, Zenodo, 2026. DOI: 10.5281/zenodo.24051030. Download the BibTeX citation.

The paper, supplement and ZIP on this site are byte-identical to the published Zenodo files. Download the SHA-256 checksums. The archive contains its own manifests and verification records.

The Result and Its Formalization

The exact directional widths are W*(5,2)=6 and W*(6,2)=33. The Lean theorems concern rational coefficients and identities on every real input, using the fixed pairwise-max feature dictionary. Each hinge compares two sums of two features, with repetitions allowed. Distinct nonparallel hinge directions are counted, and a feature-linear term is free.

Both rational optimality theorems are fully formalized in Lean. The accepted MAX6 verification of 31 October 2026 rebuilt all 16,057 project dependency modules from source, followed by independent kernel replay and final scope and completion audits. MAX5 has its separate accepted kernel verification. The final results use only the standard Lean foundations propext, Classical.choice and Quot.sound.

The real-coefficient extension is a separate written rationalization argument. The exact minima apply to PM2 directional width; they do not establish neuron optimality among unrestricted ReLU networks. The compact release preserves the accepted proof sources and records without claiming a new complete Lean rebuild.

Licences

The papers, manuscript sources and publication documentation are released under Creative Commons Attribution 4.0 International (CC BY 4.0). The author’s Lean proofs, code and supporting software data are released under the Apache License 2.0. The licences apply to different components, as specified in LICENSES.txt in the ZIP. Existing third-party licences and attributions are preserved.

Acknowledgments

The preparation of this manuscript and its computational companion benefited from assistance by AI systems developed by OpenAI and Anthropic. Responsibility for the mathematical claims, implementation, and presentation rests with the author.