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.
This Version 1.0 compact deposit contains the main paper (26 pages), its integral technical supplement (205 pages), editable manuscript sources, direct exact-arithmetic checkers and their inputs, and both complete Lean proof developments with selected verification records.
The included 17.3 MB Lean compendium preserves all 10,143 MAX5 modules and 16,057 MAX6 modules, plus the separate MAX6 scope projection. Expansion restores all 26,201 Lean source files. Every proof case, hypothesis and finite proof datum is retained. Compiler binaries, external libraries and compiled objects are excluded; Lean 4.32.2 and the exact external revisions in DEPENDENCIES.json must be installed separately for compilation.
The large historical research archives and the macOS offline verification environment are retained locally and are not part of this Zenodo deposit. The included verify.py checks the deposited files, runs the direct identity, network and geometry checks, checks manuscript consistency, and reconstructs the complete Lean sources without those archives. These checks do not rerun the archive-dependent historical lower-bound programs or execute Lean.
The accepted MAX6 verification checkpoint is 31 October 2026: the complete source rebuild of all 16,057 project dependency 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 preserved in the source compendium. The present licensing and packaging revision does not claim a new Lean replay or full source rebuild.
The formalized results concern rational coefficients, all-real input identities, the fixed PM2 dictionary, distinct nonparallel directions and a free feature-linear term. The paper's real-coefficient extension is a separate written rationalization argument. These results do not establish minimum neuron counts for unrestricted ReLU networks.
The papers and manuscript documentation are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). The author's Lean proofs and code are licensed under Apache License 2.0. These licenses cover different components, rather than offering a choice for the same work. Existing third-party licenses are preserved. LICENSES.txt specifies the scope, and licenses/ contains the full texts.
Start with README.txt. The deposit consists of one compact ZIP; no large archive is required for its documented file checks, direct mathematical checks or complete Lean source reconstruction.