Skip to content

Canonical NFAs as generators (RFSA, prime átomaton) and as shift presentations; generator-search seeds - #64

Merged
Autoplectic merged 1 commit into
mainfrom
canonical-generators
Nov 16, 2026
Merged

Autoplectic merged 1 commit into
mainfrom
canonical-generators

Conversation

@Autoplectic

Copy link
Copy Markdown
Member

Summary

Generators (sofic.generators.canonical_generators, also EpsilonMachine / BidirectionalEpsilonMachine methods), for a process with a finite bidirectional ε-machine:

canonical NFA constructor states
átomaton atomaton_hmm (reverse ε-machine run forward) S-
canonical RFSA canonical_rfsa_hmm forward causal states whose reverse beliefs R[s+] are vertices of conv{R[s+]}
maximized prime átomaton maximized_prime_atomaton_hmm reverse causal states whose forward beliefs Q[s-] are vertices of conv{Q[s-]}, built in reverse time and reversed back

A causal state's future morph is a mixture of the futures given the reverse states (disjoint supports), so "not a mixture of other morphs" is "a vertex of the belief hull", the convex analogue of a prime residual (Denis 2002; Esposito et al. 2002). Every other state is a mixture of the vertices, and the generator on the vertices has kernel K[E] C. The prime-átomaton analogue is the same on the reversed process (Maarand & Tamm 2022). Both are linear programs and generally not unifilar. The identification is an analogy, documented as not from the literature. The RFSA analogue is the bidirectional-coordinates form of the experimental canonical_residual_hmm being recovered in #63; once that merges, one could delegate to the other.

Generator search. _GeneratorProblem.residual_seed() / prime_atomaton_seed() give the seeds with completed beliefs. generative_model offers them in the machines stage (canonical RFSA generator, maximized prime átomaton generator), and coupled_search starts from them at every k. On the binary 3-state dataset the prime-átomaton analogue compresses the reverse machine on 112 of 620 processes and is the best generator known on 14; with this PR the pipeline itself lowers the best upper bound on three: n3-t088-d4 1.2809 → 1.0994, n3-t088-d0 1.3655 → 1.3301, n3-t081-d3 1.4331 → 1.4282.

Shifts (sofic.shifts.canonical_presentations, SoficShift methods): presentations by the canonical RFSA, maximized prime átomaton and átomaton of the factor language, trimmed to essential states. An automaton for the factorial, extendable factor language with every state accepting presents the shift once trimmed.

Test plan

  • All three constructors generate the source process exactly (difference-automaton test) on the golden mean, even process and a 3-causal-state mixture; two-state ε-machines are fixed points; the RFSA analogue drops the mixture state; the prime-átomaton analogue compresses a 7-state reverse machine to 3 states with entropy 1.0994 < C_mu = 1.2809
  • The pipeline's machines stage returns the prime-átomaton generator on that process
  • Hypothesis: on random sofic shifts all three presentations have the same factor language
  • Generator and shift suites: 130 passed (one test updated: the zero-budget upper bound may now come from either canonical generator)
  • ruff, sphinx-build -W; ty has no new diagnostics

Made with Cursor

…e generator search

canonical_rfsa_hmm, maximized_prime_atomaton_hmm, atomaton_hmm: stochastic analogues of the canonical
NFAs on extreme points of the causal-state belief polytopes. generative_model offers the two new ones
and coupled_search starts from them. SoficShift gains canonical RFSA, prime atomaton and atomaton
presentations of its factor language.

Co-authored-by: Cursor <cursoragent@cursor.com>
@Autoplectic
Autoplectic merged commit 9e6edac into main Nov 16, 2026
10 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant