Exponential state complexity of one-way liveness #
OpenAI, An Exponential Two-Way Deterministic State Lower Bound for One-Way Liveness (25 September 2026): https://github.com/openai/math/tree/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/An-exponential-two-way-deterministic-state-lower-bound-for-one-way-liveness-September-25-2026
The language consists of words of binary relations whose product is nonempty.
It has an h + 3-state one-way nondeterministic endmarker automaton, while every
equivalent two-way deterministic automaton has exponentially many states.
Both the zero-length-run and positive-run acceptance conventions are covered.
The source proof also rules out polynomial simulation even when its input
machines are required to be one-way.
The NFA bridge identifies the same language with a standard h-state Mathlib
NFA. This alphabet consists of relations on Fin h; it is finite but grows
with h. The imported Binary module gives a new fixed-binary-alphabet
consequence using a linear-size truth-table scanner.
One-way liveness is recognized by the ordinary h-state Mathlib NFA.
The one-way upper bound in the endmarker model, under both acceptance conventions.
Every two-way deterministic recognizer for one-way liveness has exponential state cost.
The lower bound applies to a language represented by an ordinary Mathlib NFA.
No alphabet-independent polynomial simulation of one-way NFAs by two-way DFAs exists.
Consequently, unrestricted two-way NFAs admit no such polynomial deterministic simulation.