Grokability in Five Inequalities: When AI Discovers New Math

Paper: 2605.05193 Authors: Paata Ivanisvili, Xinyuan Xie Categories: math.PR, cs.AI, math.AP, math.CA, math.FA

The Gap

Mathematical inequalities are the workhorses of analysis—they bound errors, control approximations, and reveal structure. But finding sharp constants and optimal forms is notoriously hard. Existing bounds for Gaussian perimeter, moment comparisons on the Hamming cube, autoconvolution inequalities, Sidon set sizes, and Szarek’s inequality all had room for improvement, but the techniques to tighten them weren’t obvious. The gap wasn’t conceptual—it was computational and exploratory. Researchers knew better bounds existed but lacked the pattern-matching horsepower to find them.

This paper reports five improvements discovered through human-AI collaboration with Grok. The authors posed problems, Grok suggested candidate inequalities, and the authors verified the proofs. The workflow: conjecture generation by AI, rigorous verification by humans.

Known bounds (suboptimal)
         |
         v
    [Grok explores]  <--- Human poses problem
         |
         v
  Candidate inequality
         |
         v
  [Human verifies] ---> New sharp bound

The Increment

One sentence: Before this paper, five classical inequalities had loose constants; after, we have tighter bounds with human-verified proofs, showing AI can accelerate conjecture generation in pure math.

Core Mechanism

The method isn’t a single algorithm—it’s a collaboration protocol. The human researcher identifies a problem where the current bound feels improvable. They frame it for Grok: “Can we improve the constant in this inequality?” Grok, trained on vast mathematical corpora, pattern-matches against similar structures and proposes candidate improvements. The human then takes the candidate and either proves it rigorously or finds a counterexample.

Human: Problem framing
   |
   v
Grok: Pattern matching + conjecture generation
   |
   v
Human: Rigorous proof or counterexample
   |
   v
Verified result (or back to Grok)

Think of it like a chess engine partnership. The engine (Grok) evaluates millions of positions and suggests promising moves. The grandmaster (mathematician) understands why a move works, sees the deeper strategy, and decides whether to commit. The engine doesn’t understand chess—it pattern-matches. The grandmaster doesn’t have infinite compute—they rely on intuition. Together, they reach positions neither could alone. Here, Grok doesn’t “understand” why an inequality holds, but it recognizes structural patterns from training data. The mathematician doesn’t have time to test thousands of candidate constants, but they know how to verify a proof once they see it.

Key Concepts

• Gaussian perimeter: Imagine a blob in high-dimensional space. The Gaussian perimeter measures how much “surface area” it has when you weight points by a Gaussian distribution (bell curve). Convex sets (no dents) have special properties—this paper improves the lower bound on how much perimeter they must have. Concretely: if you draw a convex shape in 100 dimensions and measure its boundary under Gaussian weighting, this paper tells you it can’t be too small.

• Hamming cube moment comparison: The Hamming cube {-1,1}^n is the set of all n-bit strings (think: coin flips). A function on this cube assigns a number to each string. Moment comparison asks: how do different ways of averaging (L_2 vs L_1 norms) relate? This paper sharpens the inequality connecting these averages, which matters for analyzing Boolean functions and noise stability.

• Sidon sets: A Sidon set is a collection of numbers where all pairwise sums are distinct. Example: {1, 2, 5, 7} is Sidon because 1+2=3, 1+5=6, 1+7=8, 2+5=7… wait, 1+7=8 and no collision. Finding the largest Sidon set in {1,…,n} is a classic combinatorics problem. This paper improves the asymptotic bound—we now know these sets can’t grow faster than a certain rate.

Framework Shift

Before (human-only):                After (human-AI):
                                    
Mathematician                       Mathematician
    |                                   |
    v                                   v
Intuition + experience          Problem framing
    |                                   |
    v                                   v
Try candidate bounds            Grok explores space
    |                                   |
    v                                   v
Manual verification             Candidate inequalities
    |                                   |
    v                                   v
New result (slow)               Human verifies (fast)
                                        |
                                        v
                                  New result (faster)

From solo exploration to collaborative search, the core shift is offloading conjecture generation to AI while keeping verification human.

Expert Assessment

Problem choice: These are real gaps—classical inequalities with known slack. Not manufactured problems, but also not the deepest mysteries in their fields. They sit in the “known to be improvable but tedious to improve” zone, which is exactly where AI assistance shines.

Method maturity: This isn’t a method paper—it’s a report of results. The “method” (ask Grok for conjectures) is more workflow than algorithm. The value is in the verified results, not the process. That said, the process is underspecified: we don’t know how many failed conjectures Grok produced, how the authors prompted it, or whether the improvements are near-optimal or just incremental.

Experimental integrity: The results are mathematical proofs, so “experimental integrity” means proof correctness. The authors claim verification, and given their expertise (Ivanisvili is a respected analyst), I trust the proofs are sound. But the paper would be stronger if it included the proofs inline or in an appendix, not just “we verified them.”

Writing quality: This is a short note, not a full paper. It reads like a progress report: “Here are five things we found.” That’s fine for a rapid communication, but it leaves questions unanswered. How much human effort went into each result? How many dead ends did Grok produce? What’s the hit rate? A longer version with proof sketches and workflow details would be more valuable.

Verdict: weak accept — Solid results in a novel collaboration mode, but the paper is more “look what we did” than “here’s how to do it.” The inequalities themselves are contributions; the AI angle is interesting but underexplored.

Takeaways

If you’re working on optimization, approximation theory, or any field where you’re hunting for sharp constants: use AI as a conjecture engine. Frame your problem clearly, let the model explore the space, then verify rigorously. The hit rate might be low, but even one good conjecture can save weeks of manual search.

For AI researchers: this is what “AI for science” looks like when it works—not replacing human insight, but accelerating the search through a vast space of possibilities. The bottleneck shifts from “finding candidates” to “verifying candidates,” which is often faster.

For mathematicians skeptical of AI: this paper shows AI doesn’t need to “understand” math to be useful. Pattern matching over a large corpus can surface non-obvious candidates. Your job—proof and insight—remains irreplaceable.