Dependency Graphs of Mathematical Arguments: Representation, Structural Measurement, and Proof Review

Eunji ParkarXiv 2026 · arXiv

This paper, authored by Eunji Park from Seoul National University, introduces a formal framework for mapping the logical "DNA" of mathematical arguments. By representing informal mathematical papers as structured dependency graphs, the author provides a way to measure the complexity of proofs, navigate dense manuscripts, and potentially catch errors in both human-written and AI-generated mathematics.

TL;DR

The paper proposes Paper Graphs and Math Graphs: a framework that uses Large Language Models (LLMs) to extract the underlying logical structure of mathematical arguments into directed acyclic graphs (DAGs). It introduces standardized metrics for "Depth" and "Bulk" to quantify proof complexity and a rigorous prompt protocol (PGPP 0.1) to ensure these measurements are reproducible across different AI systems.

What Problem Does This Paper Address?

Mathematical knowledge is traditionally organized for exposition, not just logic. A paper might present a calculation in an appendix that supports a theorem in the introduction, or multiple results might reuse a single complex construction. This non-linear structure makes it difficult for readers to navigate prerequisites and for researchers to analyze how different fields of math interconnect.

Existing tools are either too coarse (like bibliometric networks that only look at document-level citations) or too high-effort (like formal verification in Lean or Coq, which requires translating math into code). There is a "missing middle": a way to represent the fine-grained logic of informal mathematical prose in a way that is machine-readable, measurable, and verifiable.

Methods

The core of the proposal is the Paper Graph, a source-grounded representation of a manuscript.

  1. The Representation: The graph includes not just numbered theorems, but also unnumbered equations, local assumptions, and prose claims. It distinguishes between the "source occurrence" (where it appears in the text) and the "computational state" (the actual logical claim).
  2. The Measurement Contract: To make measurements comparable, the author defines a contract $C = (M, B, P, W, H)$ that fixes the manuscript scope, boundary assumptions, granularity of steps, proof selection, and "macro anchors" (key results).
  3. Key Metrics:
    • Depth ($D$): The length of the longest path of prerequisites leading to a result.
  4. Bulk ($S$): The total number of distinct edges in the ancestor subgraph. Crucially, the paper uses a "union recurrence" to ensure that shared sub-arguments are only counted once.
  5. PGPP 0.1 (Paper Graph Prompt Protocol): A standardized, versioned protocol for prompting LLMs to perform the extraction. It separates the process into three modes: DISCOVER (proposing new annotations), RECOMPUTE (calculating metrics on a fixed graph), and LINK (connecting different papers).
  6. Operation Registry: To standardize what counts as a "step," the paper provides a registry of operations like APPLY (using a theorem), SUB (substitution), and OPAQUE (a block that isn't further decomposed).
Table 1: An illustrative operation registry. Each entry must be specialized by its allowed inputs, outputs, and context.

Datasets

The paper does not report empirical findings on a specific dataset. Instead, it provides a comprehensive evaluation design for future research. It calls for the creation of a benchmark involving:

  • Multiple mathematical fields (calculation-heavy vs. conceptual).
  • Independent annotation by at least two qualified experts to create a "gold standard" reference graph.
  • Comparison of LLM extraction against these expert baselines.

Key Insights and Ideas

  • Structure $\neq$ Difficulty: The author emphasizes that these metrics describe the representation of the argument, not the intrinsic mathematical difficulty or importance. A short, elegant proof might have low depth but involve a "profound idea" that the graph doesn't capture.
  • Stability via Anchors: The paper proves that while fine-grained steps might vary between annotators, the "Macro Skeleton" (the graph of relationships between major results) remains stable if the reachability between those anchors is preserved.
  • Proof Review via Co-use: By looking at "Math Graphs" (collections of linked Paper Graphs), we can identify "theorem co-use patterns." If a new proof uses Theorem A and Theorem B but omits Theorem C—which is almost always used alongside them in the corpus—the system can flag this as a potential omission for review.

Key Results

As a theoretical and methodological paper, the "results" are the established definitions and properties of the framework:

  • Proposition 4.3: Proves that Depth is determined by fixed-vertex reachability, meaning transitive reduction preserves depth.
  • Proposition 5.2: Establishes that the macro skeleton preserves the reachability relation of the full graph restricted to the chosen anchors.
  • Algorithm 1: Provides a deterministic procedure for contract-relative graph measurement, ensuring that once a graph is accepted, its metrics are computed identically every time.

Limitations and Future Work

  • Interpretation Ambiguity: The biggest hurdle is the interpretation of informal math. Different mathematicians might disagree on where one "step" ends and another begins.
  • LLM Reliability: The framework relies on LLMs for discovery, but the author notes that "model agreement does not establish mathematical validity." Human review remains essential.
  • Future Directions: The author suggests using this framework to study "argument shapes" associated with specific authors or collaborations and to build interactive reading interfaces that allow students to "unfold" proofs layer by layer.

The paper concludes that while the framework turns a reading map into a research dataset, its ultimate success depends on developing fixed annotation policies and conducting the independent empirical work outlined in the evaluation design.

Summary updated

Abstract

We propose a source-grounded framework for representing and measuring mathematical arguments with assistance from large language models. It includes numbered and unnumbered statements, equations, assumptions, and constructions, distinguishing source occurrences from contextual states and inference records. Proposed dependencies retain source evidence and review status. A measurement contract fixes manuscript versions, boundaries, granularity, proof choices, and macro anchors. On a selected finite directed acyclic graph, depth is the longest prerequisite-path length, and bulk counts distinct ancestor edges. We give a union recurrence that accounts for shared subarguments and identify reachability conditions preserving depth and the macro skeleton; unrestricted refinement need not preserve depth or bulk. A standardized, versioned prompt protocol separates annotation discovery, deterministic recomputation, and review assistance. Cross-paper links support a research agenda on theorem co-use, rare combinations, and connections between dependency communities. These patterns motivate probabilistic premise suggestions and review of possible omissions or unchecked application conditions in human-written and AI-generated proofs. Further applications include source-linked reading support and comparison of individual and collaborative argument shapes. We provide an evaluation design rather than empirical evidence of these benefits. Suggested premises require mathematical validation, and the measures describe a specified representation rather than intrinsic mathematical difficulty, originality, or importance.
BibTeX
@misc{park2026dependency,
  author = {Eunji Park},
  title = {{Dependency Graphs of Mathematical Arguments: Representation, Structural Measurement, and Proof Review}},
  year = {2026},
  eprint = {2610.36776},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.36776},
}
Figure