A correspondence problem for mathematical proof

Simon DeDeo, Eamon DuedearXiv 2026 · DOI · arXiv

Abstract

Mathematical proofs are often said to justify their conclusions by indicating the existence of a corresponding formal derivation. We argue that this widespread view relies on an under-examined notion of correspondence, or what it means for a particular derivation to "correspond" to a particular proof. Mere existence of a formalization is not enough, and a substantive account of the required correspondence resolves into two criteria—adequate representation (of the original theorem) and tracking (of the steps in the original proof). An examination of the actually-existing formalization systems we have today shows the variety of quasi-empirical ways we establish these criteria, and points towards new burdens that may be placed on the future evolution of mathematics itself.

Introduction: The Standard View

In contemporary mathematics, there exists a significant gap between the "informal" proofs found in textbooks and the "formal" derivations checkable by machines. Textbook proofs are written in natural language, often containing gaps, tacit assumptions, and intuitive leaps. In contrast, formal derivations are line-by-line logical sequences executed within a proof assistant like Lean or Coq.

Philosophers typically bridge this gap using the "Standard View": an informal proof is considered valid if there exists a formal derivation "underneath" it. However, this view treats the relationship between the two—the "correspondence"—as a black box. As mathematics moves toward large-scale formalization projects and AI-assisted discovery, this black box is beginning to crack, revealing what we call the "Correspondence Problem."

The Correspondence Problem

The core of the problem lies in the distinction between "thin" and "thick" correspondence.

  • Thin Correspondence: The mere existence of some formal proof for a given theorem.
  • Thick Correspondence: The existence of a specific formal proof that mirrors the human’s logic and inferential steps.

We argue that thin correspondence is insufficient for epistemic justification. If a computer finds a valid proof of a theorem using a method entirely alien to the human proposer, it may confirm the theorem's truth, but it does not necessarily justify the human's original reasoning. For a formal proof to truly validate a human one, it must satisfy two primary criteria: Adequate Representation and Tracking.

1. Adequate Representation

This criterion asks: Does the formal code actually mean what the natural language theorem says? For example, when a mathematician defines a "prime number" in a code library, we must ensure that the logical constraints in the code perfectly capture the mathematical concept. Because we cannot deductively prove that a piece of code is a perfect translation of a mental concept, mathematicians have begun to treat formalization as a form of "unit testing." They run the code against known cases to see if it behaves as expected, importing methods from software engineering into pure mathematics.

2. Tracking

Tracking asks: Does the formal derivation follow the same "proof-ideas" as the human version? In practice, formal proofs often take massive, unintelligible detours due to the constraints of the software library being used. This "library lock-in" or path dependency means that the way a previous mathematician coded a definition forces future mathematicians to structure their proofs in specific, perhaps non-intuitive, ways. This can lead to formal proofs that are "semantically detached" from the human ideas they were meant to represent.

Empirical Realities of Formalization

Our analysis of high-profile formalization projects—such as the Liquid Tensor Experiment and the Polynomial Freiman–Ruzsa (PFR) Conjecture—highlights the scale of this challenge. In the PFR project, led by Timothy Gowers and Terence Tao, the informal proof consisted of approximately 17 major statements. However, the formal version in Lean required over 910 "declarations." This 10x increase in complexity, or "declaration gap," illustrates the immense amount of hidden work and translation required to achieve even a semblance of correspondence.

The AI Challenge: AlphaProof and Opacity

The rise of AI tools like AlphaProof, which generate valid code for complex problems, further complicates the issue. When an AI produces a formal proof that humans find "mysterious" or "unintelligible," the causal and explanatory links between human thought and formal verification are weakened. If we cannot track the AI's reasoning, the "thick" correspondence required for traditional mathematical certainty begins to dissolve.

Conclusion: A Quasi-Empirical Future

Recognizing the Correspondence Problem suggests that mathematical justification is becoming increasingly "defeasible" and "empirical." We establish correspondence through three routes: structural similarity between code and text, the causal history of the translation, and the explanatory power the formal proof provides back to the human.

As mathematics continues to integrate with digital libraries and AI, the field may become more rigid and less "evolvable" than the informal tradition. The future of mathematics may depend not just on the discovery of new truths, but on our ability to maintain the bridge between human intuition and machine logic.

Summary updated

Abstract

Mathematical proofs are often said to justify their conclusions by indicating the existence of a corresponding formal derivation. We argue that this widespread view relies on an under-examined notion of correspondence, or what it means for a particular derivation to ''correspond'' to a particular proof. Mere existence of a formalization is not enough, and a substantive account of the required correspondence resolves into two criteria -- adequate representation (of the original theorem) and tracking (of the steps in the original proof). An examination of the actually-existing formalization systems we have today shows the variety of quasi-empirical ways we establish these criteria, and points towards new burdens that may be placed on the future evolution of mathematics itself.
BibTeX
@misc{dedeo2026correspondence,
  author = {Simon DeDeo and Eamon Duede},
  title = {{A correspondence problem for mathematical proof}},
  year = {2026},
  eprint = {2603.13680},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.13680},
}

Loading related content...

Figure