Consequences of the ETH refutations for proof complexity and knowledge compilation
Description
The Exponential Time Hypothesis was recently refuted by two counting algorithms, one proposed by OpenAI and one by Anthropic. We describe consequences for proof complexity and for knowledge compilation. Both follow from one result: the weighted model count of every 3-CNF on n variables has an arithmetic circuit of size 2^(o(n)), which can be chosen smooth and decomposable. We obtain the circuits from the identities that the Anthropic preprint uses to prove its algorithm correct. We prove the result in Lean, building on the Lean development that Anthropic published with its preprint.
For proof complexity, every unsatisfiable 3-CNF on n variables has a refutation of size 2^(o(n)) in the Ideal Proof System (IPS) of Grochow and Pitassi, over every field of characteristic other than 2. We also prove this bound in Lean. If Extended Frege polynomially simulates IPS over F_p for an odd prime p, then it refutes these formulas in the same size. By lower bounds from the literature, no such bound holds for resolution, Res(k) for fixed k, polynomial calculus, or sum-of-squares.
For knowledge compilation, we show that a monotone circuit (with addition and multiplication gates and nonnegative real constants) for the weighted model count of a satisfiable formula yields a decomposable negation normal form (DNNF) of the formula with at most as many gates. Some 3-CNFs on n variables with O(n) clauses have only DNNFs of size 2^(Ω(n)). So, for large n, circuits of size 2^(o(n)) for their weighted model counts must contain negative constants.
Changes in version 2: In version 1, three statements of the Anthropic preprint were assumptions. We now prove them in Lean from Anthropic's public Lean development, so Propositions 1 and 2 have no hypothesis from the preprint and rest on that development, on Mathlib and on the Lean kernel. A new Corollary 6 shows that Tseitin formulas of expanders of maximum degree 3 separate circuits for weighted model counts from DNNFs. It replaces an open problem of version 1. Proposition 6 of version 1 is now Proposition 7. The Lean project moved from Lean 4.30.0 to Lean 4.33.1.
Files: eth_consequences_note.pdf (the note, 14 pages),
eth_consequences_companion.pdf (the Companion, 20 pages,
with the precise statements taken from the preprint, the Lean theorems
and the paper proofs), eth_consequences_supplement.zip (the
Lean 4 project and four scripts that test the identities and statements
on small instances) and
anthropic_formal-math_eth_16a5ce4.tar.gz (an unmodified
copy of the folder eth and of the root files
LICENSE and README.md of Anthropic's
repository formal-math at the commit that our Lean project depends on,
under the Apache License 2.0, for a build without access to that
repository).
Earlier records: https://doi.org/10.5281/zenodo.23688831 (the OpenAI thresholds), https://doi.org/10.5281/zenodo.23797196 (the k×k Clique Hypothesis) and https://doi.org/10.5281/zenodo.23981771 (the klam value).
Files
eth_consequences_note.pdf
Additional details
Related works
- Continues
- Technical note: 10.5281/zenodo.23688831 (DOI)
- Technical note: 10.5281/zenodo.23797196 (DOI)
- Technical note: 10.5281/zenodo.23981771 (DOI)