Low-Rank Schemes for 4x4 Matrix Multiplication: Accuracy Records, Rank-48 Schemes Without 1/2, and Verified Negative Results
Status: preprint, not peer reviewed, assisted by AI. The author directed the work. AI systems (Claude models through Claude Code, GPT models through Codex) wrote and ran most of the code, did most of the computations, checked each other's results and drafted the paper. Checks by AI referees are not independent human confirmation. We report the results of a one-month computer search for bilinear schemes for the product of two 4×4 matrices (decompositions of the tensor ⟨4,4,4⟩). The main open targets (rank 47 over Q or C, rank 46 over GF(2), and rank 48 over Z without division by 2) are not reached. On the secondary target, numerical accuracy of rank-48 schemes, we give exact rational rank-48 schemes with Euclidean-input growth factor γ∞,2 ≈ 21.40734 (best public value we know: 22.627) and maximum-entry growth factor γ∞,∞ ≈ 117.689 (public value: 132). We also give rank-48 schemes with all entries in Z[ω] and in Z[φ] (no 1/2 is needed), two new exact rank-48 schemes over Q and Q(i), a new flip-free class of rank-47 schemes over GF(2), about 30,000 new classes of rank-48 schemes over GF(2), and a complex border radius of at least 11 for the Li–Wang–Hu scheme D(2). Finally, we list exact negative results that close large regions around all known schemes. Evidence (schemes, checker, certificates, ledger): github.com/jeffchen006/tensor-rank-4x4
Authors
- Zhiyang Chen (ORCID: https://orcid.org/0000-0002-2315-397X)
Institutions
- University of Toronto (CA)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-20
- DOI
- https://doi.org/10.5281/zenodo.24404331
- Primary Topic
- Complexity and Algorithms in Graphs
- Type
- preprint