Eliciting implicit assumptions of proofs in the MIZAR Mathematical Library by property omission

When formalizing proofs with interactive theorem provers, it often happens that extra background knowledge (declarative or procedural) about mathematical concepts is employed without the formalizer explicitly invoking it, to help the formalizer focus on the relevant details of the proof. In the contexts of producing and studying a formalized mathematical argument, such mechanisms are clearly valuable. But we may not always wish to suppress background knowledge. For certain purposes, it is important to know, as far as possible, precisely what background knowledge was implicitly employed in a formal proof. In this note we describe an experiment conducted on the MIZAR Mathematical Library of formal mathematical proofs to elicit one such class of implicitly employed background knowledge: properties of functions and relations (e.g., commutativity, asymmetry, etc.).

💡 Research Summary

The paper investigates how implicit background knowledge—specifically the properties attached to functions and relations in the MIZAR Mathematical Library (MML)—is used in formal proofs without being explicitly cited by the author. MIZAR allows “constructor properties” such as reflexivity, symmetry, asymmetry, connectedness, irreflexivity for relations, and projectivity, involutiveness, idempotence, commutativity for functions. These properties are automatically available to the verifier and act as hidden premises.

The authors devise an experiment that systematically removes a selected property from a constructor’s environment (stored as XML) and then re‑runs the MIZAR verifier on an individual article item (definition, theorem, etc.). If verification fails, the property is deemed necessary for that item; if it succeeds, the property is considered unnecessary. By applying this procedure to every micro‑article in the MML, they obtain a complete map of direct and indirect dependencies on each property. Direct dependence means the item fails immediately when the property is removed; indirect dependence means the item depends on another item that itself needs the property.

Statistical results reveal striking patterns. Reflexivity is directly needed by over half of the library’s items and indirectly by almost all, reflecting its foundational role in equality and subset inclusion. Irreflexivity, while directly required by only a few items, becomes indirectly essential for roughly two‑thirds of the library because the proper‑subset relation carries this attribute. Asymmetry is attached to only five constructors, yet the asymmetry of the membership relation (∈) accounts for the indirect need of more than 80 000 items, underscoring its logical significance (the weak form of the foundation axiom). Projectivity, involutiveness, and idempotence appear in specialized contexts such as closure operators in topology or involutive functions, yet they support tens of thousands of items indirectly. Commutativity of addition is directly needed by about fourteen thousand items, illustrating its frequent but not universal role in algebraic reasoning.

Two concrete examples illustrate the methodology. The first shows how the proper‑subset relation’s irrefl exivity and asymmetry are essential for a proof by contradiction; removing irrefl exivity would invalidate the contradiction step. The second demonstrates that the commutativity of addition is silently used when swapping terms in an algebraic manipulation, even though the proof does not mention the property explicitly.

Beyond the statistical insight, the authors discuss several applications. In premise selection for automated theorem proving, knowing which properties are required can prune irrelevant premises, dramatically shrinking the search space. In automated generalization, the absence of a property suggests that a theorem may hold in a broader class of structures; a system could automatically propose such generalizations. The data also support reverse‑mathematics‑style analyses, allowing reconstruction of minimal axiom sets needed for each theorem, and can guide the creation of lightweight sub‑libraries. Finally, the property‑dependency vectors can serve as features for machine‑learning models that predict proof steps or assess the difficulty of new conjectures.

In summary, the paper presents a practical, language‑agnostic technique for exposing hidden assumptions in a large formal corpus. By leveraging MIZAR’s explicit property mechanism and its XML representation, the authors produce a detailed dependency landscape that is both of theoretical interest (understanding the structure of formal mathematics) and of practical value (enhancing AI‑driven proof assistance, library maintenance, and theory exploration). The methodology is readily transferable to other proof assistants that support similar attribute systems, opening a path toward richer meta‑analyses of formalized mathematics.