Section 3, “Circuit complexity and neural networks”Related results on parity, threshold circuits and attention models
External cited result A published result used as background or input; not proved in Lean here.
1 Paper text
Circuit complexity and neural networks. Parity is outside (Furst et al., 1984; Håstad, 1986), while constant-depth threshold circuits () can compute it. Attention results depend on the precise model: Hao, Angluin and Frank analyse hard-attention variants, including restrictions for unique hard attention (Hao et al., 2022). Merrill and Sabharwal study the additional power of intermediate decoding over a baseline (Merrill and Sabharwal, 2024); that containment does not rule out parity. Their sequential-computation perspective is relevant to our recurrent bypass, although the architectures and formal assumptions differ. Our contribution applies these distinctions to NIM policies and a restricted rollout-measurement interface.
Generated from paper_jmlr.tex, lines 380–383.
2 Explain this
Related work on circuits and neural models: parity is outside AC0 but inside ; results on attention depend on the exact model (hard attention can be in AC0; other variants have bounds that do not rule out parity). The sequential-computation view of chain-of-thought work is mentioned because it resembles the recurrent bypass, though the formal settings differ.
3 How this is formalised
Cited literature, not formalised here. The paper's own use of these distinctions is in Theorem 2.4 and Proposition 5.13.
| Paper anchor | par-circuit-complexity-and-neural-networks-p1 (website anchor; the manuscript has no label here) |
|---|---|
| Displayed as | Section 3, “Circuit complexity and neural networks” |
| Claim | Related results on parity, threshold circuits and attention models |
| Coverage category | External cited result |
| Statement source | paper_jmlr.tex lines 380–383; SHA-256 of the normalised block source e47d558f65a93f01… |
| Explanations | explain/c-rw-circuits, formal/c-rw-circuits |
| Build evidence | fresh isolated rebuild, audit report |
4 Lean statement and source
No Lean statement corresponds to this claim. This is a published classical result. It is cited, not proved in this Lean development.
Build evidence: fresh isolated rebuild of all 38 bundled modules with Lean (version 4.26.0, started 2026-10-10T15:21:56Z, finished 2026-10-10T15:27:00Z: 38/38 modules compiled, 1588 axiom reports, no errors. See Build for the commands and the independent audit of 2026-10-10 (dated evidence).