Revisiting average case complexity of multilevel syllogistic: From the 1995 Courant Technical Report to Lean 4 Formalization
This paper presents a Lean 4 formalization of the 1995 Courant Technical Report on the average-case complexity of Multilevel Syllogistic, encoding its semantics, decision procedures, and complexity results to establish conditional NP-average completeness and non-AvP hardness corollaries.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are trying to solve a massive, complex puzzle. In the world of computer science, some puzzles are known to be incredibly hard. If you pick the absolute worst possible puzzle piece arrangement, it might take a supercomputer the age of the universe to solve it. This is called the "worst-case" scenario.
However, in the real world, we rarely encounter the absolute worst-case scenario. Most puzzles we face are "average" puzzles. The big question this paper asks is: Are these "average" puzzles actually easy to solve, or are they still secretly hard?
The Old Report (1995)
Back in 1995, a team of researchers (Cox, Ericson, and Mishra) wrote a technical report. They looked at a specific type of logic puzzle called Multilevel Syllogistic (MLS). Think of MLS as a language for describing how sets of things relate to each other (e.g., "The set of cats is inside the set of animals").
The researchers suspected that while these puzzles are theoretically "hard" in the worst case, they might be "easy" on average. They used a mathematical framework called Average-Case Complexity to try to prove this. They claimed that if you pick a random MLS puzzle, it's actually just as hard as the hardest possible puzzles in the universe, unless a massive, unlikely mathematical miracle happens (specifically, that two huge classes of computing power turn out to be the same thing).
The New Project (2026)
Fast forward to 2026. The author of this paper, Lars Ericson, decided to revisit that 1995 report. But instead of just reading it and nodding along, he did something much stricter: he translated the entire report into Lean 4.
What is Lean 4?
Think of Lean 4 as a super-strict, robotic math teacher. You cannot just say "it seems obvious" or "trust me." You have to write down every single logical step, and the robot checks it to make sure it is 100% true. If you make a tiny mistake, the robot says, "No, that doesn't follow."
The Mission: "Grind Out the Truth"
The author's goal was to take the 1995 claims and force them through this robotic teacher. The plan had a few possible outcomes:
- The Proofs Check: The 1995 math was perfect, and the robot agrees.
- The Paper is Wrong: The 1995 authors made a mistake, and the robot finds the exact spot where the logic breaks.
- The Tool is Too Weak: The 1995 math is correct, but Lean 4 isn't powerful enough yet to prove it.
- The Definitions are Shaky: The concepts used in 1995 were too vague to be programmed into a robot.
What They Actually Did
The paper is essentially a "construction log" of building a digital fortress. Here is what they built, using simple analogies:
- Building the Dictionary (Phase 1): They taught the robot what "Average-Case Complexity" means. They defined what a "puzzle" is, what a "random distribution" of puzzles looks like, and how to measure if a puzzle is "hard" on average.
- Translating the Language (Phase 2): They taught the robot the language of MLS (Multilevel Syllogistic). They created a way for the robot to read set-theory sentences and understand what they mean.
- The Solver (Phase 3 & 4): They built a "solver" (a program) that tries to solve these puzzles. They proved that this solver works correctly for a specific, safe subset of puzzles.
- The Hardness Test (Phase 5): This is the climax. They tried to prove the 1995 claim: "These puzzles are hard on average."
The Results: "Proofs Check" (With Caveats)
The paper concludes that the 1995 report was largely correct.
- The Good News: The robot successfully verified the definitions and the logic for the parts of the 1995 report that were fully formalized. The core idea that "MLS puzzles are hard on average" holds up under the strict scrutiny of Lean 4.
- The "But": The author didn't just copy-paste the 1995 math. He had to make some choices where the original report was vague. For example, the 1995 report assumed a specific way to translate a computer program into an MLS puzzle. The 1995 authors didn't write out the code for this translation; they just said it exists.
- In the Lean 4 version, the author had to axiomatize this missing piece. This means he told the robot, "Assume this translation exists and works perfectly."
- Because of this, the final proof relies on a few "assumptions" (axioms) rather than being a 100% closed loop from first principles.
The "Nose" Diagram
The paper mentions a famous diagram from the 1995 report called "The Nose."
- Imagine a graph where the vertical axis is "How hard is the worst puzzle?" and the horizontal axis is "How hard is the average puzzle?"
- There is a "nose" shape at the bottom left. This is the "sweet spot" where puzzles are easy to solve on average.
- The 1995 report (and this new paper) argues that MLS puzzles do not live in this sweet spot. They live outside the nose, meaning they are hard even on average.
Why This Matters (According to the Paper)
The paper doesn't claim this will fix your software tomorrow. Instead, it's a historical and mathematical audit.
- It confirms that the 1995 researchers were right to be skeptical about "easy average cases" for this type of logic.
- It highlights that the field of "Average-Case Complexity" has moved on. In the 1990s, people tried to prove specific logic languages were hard on average. Today, the field focuses more on cryptography (making sure keys are hard to break) and Smoothed Analysis (looking at how algorithms handle slightly messy real-world data).
- The specific "marriage" of average-case theory to set-theory solvers (MLS) was largely abandoned by the industry because real-world software isn't random; it's structured. Modern solvers use clever tricks (heuristics) to solve these problems quickly, regardless of the theoretical "average" hardness.
Summary
This paper is a rigorous audit. The author took a 30-year-old mathematical claim, rebuilt it inside a robot-proof environment, and found that the original claim stands: Multilevel Syllogistic puzzles are indeed hard to solve on average. However, the audit also revealed that the original authors relied on some "hand-waving" steps that had to be explicitly assumed as true for the modern robot to accept the proof. It's a victory for the old math, but with a reminder that even brilliant 1995 papers can have gaps that only a 2026 robot can spot.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.