Status: closed (2026-10-25). Found while reviewing
Retracted(L-SPEC-004 Algebraic Factors Toric and Grassmannian §2.2). Related: factors, Everything is a Factor, grassmann. Code:lib/Mycelium.jl/src/factors.jl(Momentum,Nesterov),lib/Mycelium.jl/test/runtests.jl.
Problem. Mycelium’s Momentum(η, γ) and the put of Nesterov(η, γ) computed
, then . Cruttwell et al., Categorical Foundations of
Gradient-Based Learning, Example 3.16 (arXiv:2103.01931v1), define momentum by
and ; with Mycelium’s state this is
. Example 3.17’s Nesterov look-ahead was
already consistent with ; only the update had the wrong sign. With the old
rule damped each step by the previous one instead of accelerating it. The Mycelium test
hard-coded the wrong value, and two tables (factors.md, Everything is a Factor) repeated
it.
Resolution. The sign is corrected in both rules, the test now expects
, and the tables are fixed. The Grassmannian descent test of
lib/AlgebraicFactors.jl runs Retracted(Momentum(0.04, 0.6), M) with the corrected rule.