Characterizing co-NL by a group action

In a recent paper, Girard proposes to use his recent construction of a geometry of interaction in the hyperfinite factor in an innovative way to characterize complexity classes. We begin by giving a detailed explanation of both the choices and the motivations of Girard’s definitions. We then provide a complete proof that the complexity class co-NL can be characterized using this new approach. We introduce as a technical tool the non-deterministic pointer machine, a concrete model to computes algorithms.

💡 Research Summary

The paper presents a rigorous and self‑contained proof that the complexity class co‑NL can be characterized using Girard’s recent Geometry of Interaction (GoI) construction in the hyperfinite factor. The authors begin by unpacking Girard’s framework, emphasizing three core components: (1) the trace operation, which captures cyclic program behavior; (2) the normalization process, which guarantees convergence of infinite‑dimensional operators; and (3) the unitary operators that form a group action on the hyperfinite II₁ factor. They explain why these choices are natural for encoding resource‑bounded computation, noting that the trace’s finitary restriction mirrors the logarithmic space bound of NL machines.

Next, the paper introduces the nondeterministic pointer machine (NPM), a concrete computational model designed to reflect the operational semantics of co‑NL. An NPM consists of a finite set of states and a finite collection of pointers that move over an input tape. Each transition is nondeterministic and may change the state, move a pointer, or both, while the overall configuration space remains logarithmic in the size of the input. The authors prove that NPMs are equivalent to standard nondeterministic log‑space Turing machines, and that the acceptance condition “all computation paths reject” precisely captures co‑NL.

The central technical contribution is a systematic translation from NPM transitions to unitary operators in Girard’s GoI model. For each elementary transition τ of the NPM, a corresponding unitary U_τ is constructed inside a specific sub‑algebra of the hyperfinite factor. The group action g·U_τ, where g belongs to a countable group generated by the set of all transitions, encodes the nondeterministic branching: composing operators corresponds to concatenating transitions along a computation path. The trace of a product of such operators, Tr(U_τ₁…U_τ_k), represents the sum over all possible paths of length k, and normalization ensures that only those paths respecting the log‑space constraint survive in the limit.

With this encoding in place, the authors prove two complementary theorems. The first theorem shows that if a language L belongs to co‑NL, then there exists an NPM M recognizing L such that the associated GoI operator A_M (the sum of all transition operators) has a fixed point under the group action that decides membership in L. The second theorem establishes the converse: any language whose GoI representation admits such a fixed point must be in co‑NL. Together these results yield an exact equivalence:

 co‑NL = { L | L is decided by the fixed‑point of a group‑action operator in the hyperfinite GoI model }.

The paper then discusses the broader implications of this equivalence. By translating a classical complexity class into an algebraic property of operators, the work opens a new avenue for visualizing inclusion relationships (e.g., NL ⊆ P, NL ⊆ NC¹) as relationships between sub‑algebras or between different group actions. It also suggests that other classes—such as PSpace, BPP, or even quantum classes—might be captured by modifying the underlying group, the dimensional constraints on the trace, or by enriching the operator algebra with additional structure (e.g., a *‑automorphism representing randomness).

Finally, the authors acknowledge practical limitations. The hyperfinite factor is an infinite‑dimensional object that currently exists only in the realm of operator algebra theory; concrete implementations would require finite approximations or simulation techniques. Moreover, while the NPM provides an intuitive bridge between machines and operators, designing efficient algorithms directly within the GoI framework remains an open challenge. Nonetheless, the paper succeeds in demonstrating that co‑NL admits a clean, algebraic characterization via group actions, thereby enriching both the theory of Geometry of Interaction and the algebraic study of computational complexity.