mathliblemma-folklore-lemma-generation

name

mathliblemma-folklore-lemma-generation

description

Multi-agent system for discovering and formalizing missing 'folklore' lemmas in Lean 4 / Mathlib. Identifies gaps in formal math libraries, generates Lean 4 statements, type-checks them, and iterates until verified. Trigger phrases: 'find missing lemmas in Mathlib', 'generate folklore lemma', 'formalize lemma in Lean 4', 'Mathlib gap analysis', 'discover missing Lean theorems', 'automate Lean formalization'.

MathlibLemma: Folklore Lemma Discovery and Formalization for Lean 4

This skill enables Claude to operate as a multi-agent pipeline that discovers missing "folklore" lemmas -- well-known mathematical results that mathematicians use routinely but that have not yet been formalized in Lean 4's Mathlib library. Drawing from the MathlibLemma framework (Liu et al., 2026), the technique decomposes the problem into four coordinated phases: gap identification, natural-language statement drafting, Lean 4 formalization, and iterative type-check verification. The result is a stream of verified, merge-ready Lean 4 lemma declarations that fill the connective tissue between existing Mathlib theorems.

When to Use

• When a user is working in Lean 4 and hits a missing lemma that "should obviously exist" in Mathlib but doesn't.
• When a user wants to systematically audit a Mathlib module (e.g., Mathlib.Topology.Basic) for commonly-needed but absent intermediate results.
• When a user asks to formalize a well-known mathematical fact (from a textbook, paper, or informal proof) into a type-checked Lean 4 statement with proof.
• When building a verified library of supplementary lemmas for a specific mathematical domain (algebra, analysis, topology, combinatorics, number theory).
• When a user wants to contribute new lemmas to Mathlib and needs help identifying what is actually missing and how to formalize it to Mathlib style standards.
• When a user has a Lean 4 proof that fails because an intermediate step has no Mathlib support, and they need that bridge lemma generated.

Key Technique

Gap-Finding via Proof Dependency Analysis. The core insight of MathlibLemma is that missing lemmas can be discovered systematically rather than stumbled upon. The Gap Finder agent analyzes the dependency graph of existing Mathlib theorems and identifies structural holes: pairs of theorems where a plausible intermediate result is missing, API patterns where one direction of an iff exists but not the other, coercions or simp lemmas that are conspicuously absent for a given type, and theorems that exist for one algebraic structure (e.g., Group) but not an analogous one (e.g., AddGroup). The agent also cross-references informal mathematical sources (textbooks, MathOverflow, Wikipedia) against the Mathlib namespace to spot well-known results that lack formal counterparts.

Formalization with Iterative Type-Check Feedback. Once a candidate lemma is identified in natural language, the Formalizer agent translates it into a Lean 4 theorem or lemma declaration using Mathlib's existing type universe, notation, and naming conventions. This is not a one-shot process. The Verifier agent runs lean --run (or lake build) to type-check the statement. If it fails, the error message -- including unknown identifiers, type mismatches, or universe issues -- is fed back to the Formalizer for correction. This loop typically converges in 2-5 iterations. The key is that Lean's type checker provides precise, actionable error signals that guide the LLM toward correct formalization far more effectively than natural-language feedback alone.

Quality Gate: Mathlib Style Compliance. A generated lemma is not merely type-correct; it must also follow Mathlib's strict style conventions -- correct namespace, dot-notation compatibility, @[simp] annotations where appropriate, and proofs that use the preferred tactic style (e.g., exact, simp, ring, linarith over manual term-mode). The Refinement agent checks these conventions and rewrites proofs to be idiomatic before declaring a lemma ready for contribution.

Step-by-Step Workflow

1. Scope the target domain. Identify the Mathlib module or mathematical area to audit (e.g., Mathlib.Analysis.SpecificLimits.Basic or "basic properties of cyclic groups"). Use lake env printPaths and grep through .lean files to inventory existing declarations in the target namespace.

2. Run gap-finding heuristics. For the target namespace, apply these concrete checks:

   • Symmetry gaps: For each theorem foo_bar, check whether bar_foo exists. For each A → B implication, check whether B → A or A ↔ B exists.
   • Algebraic analogy gaps: For each lemma about Mul/Group/Ring, check whether the additive counterpart (Add/AddGroup/AddRing) exists (and vice versa via the @[to_additive] attribute).
   • Simp completeness: Identify terms that appear in simp lemma LHS patterns and check that all natural rewrite directions are covered.
   • Coercion chains: Check for missing coercion lemmas between related types (e.g., Nat → Int → Rat → Real).
   • Textbook cross-reference: Compare the namespace's coverage against a standard reference (e.g., Bourbaki, Lang's Algebra, Rudin's Principles) and flag missing standard results.

3. Draft candidate lemma statements in natural language. For each identified gap, write a precise English statement: "For any commutative ring R and elements a, b in R, if a divides b and b divides a, then a and b are associates." Include the expected Lean 4 types and the Mathlib namespace where it belongs.

4. Formalize into Lean 4. Translate each natural language statement into a Lean 4 declaration. Use Mathlib's existing definitions, type classes, and naming conventions. Skeleton:

   import Mathlib.RingTheory.Associated
   
   theorem dvd_dvd_iff_associated {R : Type*} [CommMonoid R]
       (a b : R) (h1 : a ∣ b) (h2 : b ∣ a) : Associated a b :=
     ⟨IsUnit.mk0 _ (by sorry), by sorry⟩
   

   Use sorry for proof bodies initially -- the priority is a type-correct statement.

5. Type-check the statement. Run the Lean 4 type checker on the file. Parse error output for:

Concrete Examples

Example 1: Finding a missing iff lemma in order theory

User: "I'm working in Mathlib.Order.Basic and I noticed there's le_antisymm but no convenient iff connecting a ≤ b ∧ b ≤ a with a = b as a single simp lemma. Can you generate it?"

Approach:

1. Search Mathlib.Order.Basic for existing le_antisymm variants and any iff form.
2. Confirm that le_antisymm_iff does not exist in current Mathlib.
3. Draft: "For a partial order, a = b ↔ a ≤ b ∧ b ≤ a."
4. Formalize:

import Mathlib.Order.Basic

@[simp]
theorem eq_iff_le_and_ge {α : Type*} [PartialOrder α] {a b : α} :
    a = b ↔ a ≤ b ∧ b ≤ a :=
  ⟨fun h => h ▸ ⟨le_refl _, le_refl _⟩, fun ⟨h1, h2⟩ => le_antisymm h1 h2⟩

5. Type-check passes. Verify @[simp] orientation is correct (LHS is the more complex side).

Output: A verified, simp-annotated Lean 4 lemma ready for Mathlib contribution.

---

Example 2: Generating additive counterparts for a multiplicative lemma

User: "Mathlib has mul_left_cancel for groups but I can't find add_left_cancel stated as a standalone lemma in Mathlib.Algebra.Group.Basic. Can you generate the missing additive version?"

Approach:

1. Locate mul_left_cancel in Mathlib and check its signature and attributes.
2. Check whether @[to_additive] was already applied (if so, the additive version exists automatically).
3. If not auto-generated, formalize the additive version:

import Mathlib.Algebra.Group.Basic

@[to_additive]
theorem mul_left_cancel_of_eq {G : Type*} [LeftCancelMonoid G]
    {a b c : G} (h : a * b = a * c) : b = c :=
  LeftCancelMonoid.mul_left_cancel a b c h

4. Type-check. If LeftCancelMonoid doesn't exist under that name, search for the correct typeclass (LeftCancelSemigroup, CancelMonoid, etc.) and adjust.
5. Verify the @[to_additive] attribute automatically produces add_left_cancel_of_eq.

Output: A single declaration that produces both multiplicative and additive versions via Mathlib's to_additive machinery.

---

Example 3: Bridging a proof gap in analysis

User: "My Lean proof needs the fact that the sum of two continuous functions on a metric space is continuous, but I'm getting errors. Can you find and formalize whatever intermediate lemma I'm missing?"

Approach:

1. Check what the user's proof looks like and identify the exact error.
2. Search Mathlib.Topology.ContinuousOn and Mathlib.Topology.Algebra.Ring.Basic for Continuous.add.
3. If Continuous.add exists, the issue is likely a missing import or typeclass. If it genuinely doesn't exist for the user's specific type:
4. Formalize:

import Mathlib.Topology.Algebra.Ring.Basic

theorem continuous_add_of_continuous {X : Type*} [TopologicalSpace X]
    {M : Type*} [TopologicalSpace M] [Add M] [ContinuousAdd M]
    {f g : X → M} (hf : Continuous f) (hg : Continuous g) :
    Continuous (fun x => f x + g x) :=
  hf.add hg

5. Type-check. If ContinuousAdd is the wrong class, iterate with the error to find the right one (TopologicalAddGroup, HasContinuousAdd, etc.).
6. If the lemma already exists under a different name, report the correct name and import to the user instead of generating a duplicate.

Output: Either a new verified bridge lemma or the identification of the correct existing Mathlib lemma with the right import path.

Best Practices

• Do: Always search Mathlib thoroughly before generating a "new" lemma. Use #check, #print, exact?, apply?, and grep. Many folklore lemmas exist under non-obvious names.
• Do: Use sorry strategically -- get the statement to type-check first, then fill proofs. A correct statement with sorry is more valuable than a wrong statement with a proof.
• Do: Follow Mathlib naming conventions strictly: theorem dvd_mul_of_dvd_left not theorem my_dvd_lemma. Names should read as the statement itself.
• Do: Apply @[to_additive] whenever a multiplicative lemma has a natural additive counterpart. This halves the work and keeps the library consistent.
• Avoid: Generating lemmas that are trivial consequences of simp -- if simp closes the goal, a standalone lemma adds clutter, not value.
• Avoid: Using decide or native_decide in proofs intended for Mathlib contribution -- these are brittle and not accepted upstream for non-trivial instances.

Error Handling

Error	Cause	Fix
unknown identifier 'Foo'	Wrong import or name changed in recent Mathlib	Search with grep -r "def Foo|theorem Foo|class Foo" ~/.elan/toolchains/ or use #check Foo in a Lean file with broad imports
type mismatch: expected α, got β	Wrong typeclass hierarchy or argument order	Check the actual types with #check @targetTheorem and align your signature
universe inconsistency	Mixing Type and Type* or Prop incorrectly	Add explicit universe variables: universe u v and annotate types
tactic 'simp' failed	Missing simp lemmas in context or wrong goal shape	Try simp? to see what lemmas simp tried, then either add @[simp] to prerequisites or switch to exact
maximum recursion depth exceeded	Typeclass search loop from conflicting instances	Minimize instances in scope; use @theorem (explicit mode) to bypass typeclass resolution
Proof compiles but Mathlib linter fails	Style violations (unused variables, wrong attributes)	Run lake exe lint and address each warning; common fixes include adding _ prefixes to unused variables and removing redundant hypotheses

Limitations

• Lean 4 / Mathlib only. This workflow targets the Lean 4 ecosystem. It does not apply to Coq, Isabelle, or Lean 3 without significant adaptation.
• Mathlib version sensitivity. Mathlib changes daily. A generated lemma may type-check against one nightly but fail against the next due to renamed definitions or restructured modules. Always pin your Mathlib version and re-check before submission.
• Proof search is bounded. While statement generation succeeds reliably (type-checking is a strong signal), proof generation is harder. Complex proofs requiring novel mathematical insight are beyond the scope of automated generation -- the system excels at "routine" proofs (1-5 tactic steps) not research-level arguments.
• False positives in gap finding. Not every "missing" lemma is actually useful. Some gaps exist intentionally because the result is trivially derivable or stylistically disfavored. Always ask: "Would a Mathlib maintainer accept this?"
• No informal-to-formal for ambiguous statements. If the natural language statement is mathematically ambiguous (e.g., "continuous" without specifying the topology), the formalization may target the wrong definition. Precision in the natural language input is critical.

Reference

Paper: Liu et al., "MathlibLemma: Folklore Lemma Generation and Benchmark for Formal Mathematics" (arXiv:2602.02561, 2026). Focus on Section 3 (multi-agent pipeline architecture) and Section 4 (gap-finding heuristics and formalization loop) for implementation details. The benchmark of 4,028 type-checked Lean statements provides a useful test set for evaluating formalization quality.