Back to search
2007.12141

Dimension reduction in recurrent networks by canonicalization

Lyudmila Grigoryeva, Juan-Pablo Ortega

correctmedium confidence
Category
Not specified
Journal tier
Strong Field
Processed
Sep 28, 2025, 12:55 AM

Audit review

The paper proves a canonical realization theorem (Theorem 3.2) for causal, time-invariant filters with the input-forgetting property via Nerode equivalence on semi-infinite pasts, constructing X = ZZ_- / ∼, F([z], a) = [za], and h([z]) = HU(z); it establishes well-definedness, echo state property (at the state level), canonicity (strong reachability + observability), and uniqueness up to system isomorphism (via Proposition 2.4 and Corollary 2.6) . The candidate solution mirrors this construction and all key logical steps. The only notable gap is that it proves output-level uniqueness (ESP at the output) rather than the paper’s stronger state-level ESP used to assert canonicity formally; however, the rest of the argument aligns and can be completed by the paper’s state-level uniqueness argument .

Referee report (LaTeX)

\textbf{Recommendation:} minor revisions

\textbf{Journal Tier:} strong field

\textbf{Justification:}

The manuscript offers a solid, general canonicalization framework for semi-infinite inputs under input-forgetting. The main theorems are correct, proofs are constructive, and the work connects meaningfully to reservoir computing. Minor clarifications around the precise notion of ESP used at each step and explicit reminders of assumptions in uniqueness arguments would enhance clarity without altering results.