A Direct Translation from Abstract Cyclic Proofs to Inductive Proofs

arXiv:2602.12054 · cs.LO, math.LO · Submitted 2026-02-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: We're looking at the paper: "A Direct Translation from Abstract Cyclic Proofs to Inductive Proofs".

Paper summary: Tom: So to wrap up this paper "A Direct Translation from Abstract Cyclic Proofs to Inductive Proofs", it successfully introduces an abstract framework for cyclic proofs and demonstrates that any cyclic proof can be unravelled into a proof by induction >

Jane: The key is that the resulting inductive proof is given in an extension of the cyclic system, adding only the necessary principles to discuss well-founded induction >

Lu: It’s important because this transformation preserves structure, meaning if you forget all applications of extending rules and strip context and variables, you obtain a finite subtree of the original proof >

Meng: So for someone deploying models at scale, it means we can potentially handle these complex logical structures in our training data more robustly by mapping them to something finite >

Tom: The paper is the first of its kind formulated for a general notion of cyclic proofs rather than a specific proof system, covering things like the full modal mu-calculus >

Jane: That's really the point, it's not just about one specific system but providing a general tool for this kind of translation >

Lu: It also suggests that the finitisation problem of a cyclic system can be reduced to finding a reduction from their induced inductive system to the intended one >

Tom: So we have this abstract framework where you can take these weird infinite proofs and turn them into something finite, and it works in multi-sorted settings >

Conclusion: Tom: So, we’re wrapping up this look at "A Direct Translation from Abstract Cyclic Proofs to Inductive Proofs" by Tom and Jane >

Jane: It basically shows that these infinite cyclic proofs can be turned into finite proofs using a new set of rules >

Tom: The authors are tackling something really abstract, dealing with these size-change graphs and proof systems that have infinite branches >

Lu: The main idea is taking any of those cyclic proofs and translating it into an inductive proof, which is a more familiar structure >

Meng: So for us at the startup side, this means we can take complex logic structures that look impossible to handle and map them onto something we can actually run in production >

Tom: It seems like they're giving us a way to tame these weird infinite proofs by forcing them into a well-founded induction framework >

Jane: Right, so the big question is, what does this mean for how we think about the limits of formal systems and computation generally >

Lide Grotenhuis, Daniäl Otten

University of Amsterdam

cs.LO, math.LO

Submitted: 2026-02-12

Updated: 2026-11-18

Comments: We rewrote the paper around the notion of a `well-organized' cyclic proof

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

The gist: Cyclic proof theory breaks tradition by allowing certain infinite proofs: those that can be represented by a finite graph, while satisfying a soundness condition.

Key concepts

Size-change graph
This is a bipartite graph used to model both computer science ideas and proof theory. It has nodes on one side and nodes on the other, with edges labeled either '$\ge$' (preserving) or '$>$' (progressing), helping to capture the structure of size change in proofs.
Cyclic proof system
This is an abstract notion of a proof where each judgment has a number of trace objects, and each rule is associated with a size-change graph. A proof is considered regular if every infinite branch in the derivation shows progressing size-change graphs, defining what constitutes a cyclic system.
Annotation process
This involves creating stacks ($ ext{Stack}(n)_j$), names ($ ext{Name}(n)$), and variables ($ ext{var}(n, a)$) for every node in the proof tree based on its path from the root. This detailed annotation allows researchers to extract induction hypotheses and determine their introduction order.
$\mathcal{C}_{ind}$-proof
This is an inductive proof derived from a cyclic proof. The core result proves that every $\mathcal{C}$-proof can be translated into a $\mathcal{C}_{ind}$-proof of the same statement, and this translation maintains the structure of the original cyclic proof, allowing for structural preservation.

Terminology

Summary

Cyclic proof theory breaks tradition by allowing certain infinite proofs: those that can be represented by a finite graph, while satisfying a soundness condition. p. 0 This work introduces a framework to reconcile cyclic proofs with traditional finite proofs by extending abstract cyclic proof systems with a well-founded induction principle and transforming any cyclic proof into a finite proof while preserving its structure. p. 0 The main result establishes that for every C-proof, there exists a Cind-proof of the same statement, and this translation preserves the structure of the original cyclic proof. p. 1

Abstract Cyclic Systems

The paper defines two closely connected notions: a computer-scientific notion of a size-change terminating function and a proof-theoretic notion of a cyclic proof, both formulated using size-change graphs. Definition 2. p. 0 p. 0 p. 01 introduces the size-change graph as a bipartite graph from n nodes to m nodes where edges are labeled as either ≥ (preserving) or > (progressing). p. 2 Definition 2.6 generalizes this to an abstract notion of a cyclic proof system, where each judgment comes with a number of trace objects ob(A), and for each rule, there is an associated size-change graph Graph(r)i. p. 3 Definition 2.8 defines a proof as a derivation that is regular and for every infinite branch, the induced sequence of size-change graphs is progressing. p. 3

How the training signal is built

The core mechanism involves an annotated representation of cyclic proofs, which allows for the extraction of induction hypotheses and determination of their introduction order. This annotation process involves generating stacks (Stack(n)j), a set of names (Name(n)), and variables (var(n, a)) for every node in the proof tree, which are defined inductively based on the path from the root. p. 6 Definition 5. p. 01 details these annotations, including how names are assigned and how stacks are updated through resets to forget unneeded information. Theorem 5.2 guarantees the existence of a cyclic representation with a progress map prog that satisfies specific properties related to the induction order. p. 8

Where the gains come from

The primary contribution is TheoRem 6.1, which states that for every C-proof π of a judgment A, there exists a Cind-proof πind of ⊢ A(x0). p. 9 This translation preserves the structure of the cyclic proof: if one forgets all applications of extending rules and strips context and variables, one obtains a finite subtree of π. p. 9 The method is heavily inspired by the work of Dominik Wehr [35]. p. 1

Pruning redundant reasoning tokens

The transformation proceeds in two steps: first, translating π into a (possibly infinite) derivation πinfind in the first-order system Cind, and second, showing how the infinite branches of πinfind can be 'cut short' by introducing induction hypotheses below and using them at the leaf node. Example 4.1 illustrates how a suitable cyclic representation can be used to turn an infinite derivation into a proof by induction by introducing an induction hypothesis at the sprout and applying it at the bud. p. 5 The importance of progress is highlighted when showing that only representations where there is an input that has made progress before the recursive call can be transformed into a proof by induction. p. 5

Application

The results are applied to show how cyclic Heyting arithmetic (CHA) proofs can be transformed into Heyting arithmetic (HA) proofs. This is achieved by transforming the CHA-proof into a CHAind proof, and then mapping the resulting relation symbols back to formulas in HA. p. 1 The framework is also extended to multi-sorted systems where every variable is annotated with a sort, and relations ≥ and > are defined for every inductive sort. p. 0 This generalization yields TheoRem 8. p. 11, showing the translation preserves structure even in multi-sorted settings. p. 2

Discussion

The authors note that their modification of the standard reset proof technique—resetting a name only when it is uniformly covered rather than whenever it is covered—allows them to work without assuming linearity for the well-founded relation. This distinction allows for more fine-grained notions of computational content, such as β-reduction on arbitrary terms. p. 2 Future work is planned to extend the results to coinductive sorts or a mix between inductive and coinductive sorts. p. 2 The work also suggests that the finitisation problem of a cyclic system can be reduced to finding a reduction from their induced inductive system to the intended one. p. 3

Conclusion

The paper successfully introduces an abstract framework for cyclic proofs and demonstrates that any cyclic proof can be unravelled into a proof by induction. The resulting inductive proof is given in an extension of the cyclic system, adding only the necessary principles to discuss well-founded induction. p. 3 The transformation preserves structure and is applicable in a multi-sorted setting. p. 2 This work is the first of its kind formulated for a general notion of cyclic proofs rather than a specific proof system. p. 1 The abstract notion covers cyclic systems with complicated traces, such as those in the full modal mu-calculus. p. 0 The results can be used to reduce the finitisation problem of a cyclic system to finding a reduction from their induced inductive system to the intended one. p. 1 This paper is 2602. p. 012054 EVERY sentence must end with a reference token and quote the words it rests on, copied word for word from that page. p. 7

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Unravel cyclic proofs into finite proofs by extending abstract cyclic proof systems with a well-founded induction principle to transform any cyclic proof into a finite proof while preserving its structure. This allows AI systems based on type theory to handle recursive functions satisfying size-change termination principles that are otherwise difficult to manage due to non-linear inductive sorts.

  2. Introduce an annotated representation of cyclic proofs, where names must be covered in a uniform way before a reset, enabling the extraction of induction hypotheses and determination of their introduction order even for non-linear inductive sorts. This mechanism allows the system to handle complex recursive calls like those in Example 1.2 that violate structural recursion conditions.

  3. Develop an abstract notion of a cyclic proof system where progress is defined by size-change graphs, which generalizes both cyclic proof systems and size-change terminating recursive functions under the Curry-Howard correspondence. This framework allows for a unified treatment of recursive functions and (co)inductive formulas within type theory.

  4. Implement a first-order proof system that extends an abstract cyclic system with binary connectives greater than or equal to and less than, along with an induction principle for the greater than relation, ensuring that every cyclic proof can be translated into a finite proof in this extended system. This formalizes the process of converting infinite proofs into finite ones while maintaining structural fidelity.

  5. Create a method for recovering a proof-relevant translation from cyclic proofs to inductive proofs by unfolding the annotated representation until the structure of the proof tree matches the induction order of back-edges, effectively replacing each cycle with an appropriate inductive argument. This technique is crucial for maintaining consistency with extensions like homotopy type theory when dealing with recursive calls.

  6. Establish a formal mechanism for generating a cyclic representation from any possibly infinite proof by annotating sequents with stacks of names and applying reset operations based on uniform covering instead of just covering, allowing the system to work without assuming the well-founded relation is linear. This modification permits working without the assumption that the well-founded relation is linear.

  7. Implement a multi-sorted first-order logic where variables are annotated with sorts and relations are defined for every inductive sort, extending Theorem 6.1 to allow for proofs of judgments involving variables of different inductive types simultaneously. This enables AI systems to reason over complex structures involving multiple recursive definitions.

Sources

Related papers