🔢 mathematics

Almost-Orthogonality in Lp Spaces: A Case Study with Grok

This article refutes a proposed sharpened triangle inequality for $L^p$ spaces with $p>2$ , establishes the optimal exponent for such estimates in the critical case $c=p'$ for integer $p\ge 2$ , and derives a sharp three-function estimate with an optimal exponent that improves upon earlier results, employing the large-language model Grok to explore intermediate lemmas.

Original authors: Ziang Chen, Jaume de Dios Pont, Paata Ivanisvili, Jose Madrid, Haozhu Wang

Published 2026-05-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ziang Chen, Jaume de Dios Pont, Paata Ivanisvili, Jose Madrid, Haozhu Wang

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 measure the entire "size" or "weight" of a pile of various objects. In the world of mathematics, specifically in $L^p$ spaces, these objects are functions (visualize them as wavy lines or shapes on a graph), and the "size" is a specific method for calculating their magnitude.

Normally, when you have a collection of things, the size of the entire pile is less than or equal to the sum of the sizes of the individual parts. This is the famous triangle inequality. It is like saying: if you walk 3 miles north and 4 miles east, your total distance traveled is 7 miles, but your direct distance home is only 5 miles. The straight line is always shorter than or equal to the sum of the parts.

However, mathematicians are always seeking sharper, more precise rules. They asked: What if these "objects" (functions) are, in a sense, independent of each other, like people walking in different directions? If they are truly independent (orthogonal), the mathematics becomes much simpler. But what if they are only almost independent?

This article, titled "Almost-Orthogonality in $L^p$ Spaces," is a detective story about the search for the perfect rule to measure these "almost independent" piles. Here is the breakdown of their journey:

1. The Broken Rule (The Counterexample)

A mathematician named Carbery proposed a very elegant, sharpened rule. He conjectured that if you know how strongly two functions "overlap" (a bit like the overlapping area of two shadows on a wall), you could predict the size of their sum with a specific formula.

The authors of this article decided, with some help from an AI named Grok, to test this rule. They built a specific, tricky mathematical "house of cards" (a counterexample).

The Discovery: They found that Carbery's proposed rule fails for almost all cases where the exponent $p$ is greater than 2. It is like trying to use a ruler designed for straight lines to measure a spiral staircase; it simply does not work.
The Lesson: They proved that for a rule of this kind to work, the "power" used in the formula must be smaller than a certain threshold. If one attempts to use a power that is too high, the mathematics collapses.

2. Repairing the Rule (The Correction)

Once they knew the original rule was broken, they asked: What is the highest power we can use before it fails?

They found the "tipping point" or the critical exponent. They proved that if you use exactly this specific power (which they call $p'$ ), the rule works perfectly for all integer values of $p \ge 2$ .
The Analogy: Imagine trying to balance a stack of plates. The original rule claimed you could stack them as high as you wanted if they were slightly tilted. The authors showed that the stack falls if you tilt them too much. But they found the exact angle at which the stack is perfectly stable.

3. The Three-Function Puzzle (The Special Case)

The article then zooms in on a specific, difficult scenario: What happens when you have exactly three functions?

Previous mathematicians had a rough estimate for this, but it was not the best possible answer.
The authors used a combination of human intuition and AI assistance to solve this puzzle. They derived a sharp, perfect bound.
The "Orthogonality" Measure: They introduced a variable called $\Gamma$ $Γ$ (Gamma). Visualize this as a "friendship meter" or an "overlap meter."
- If the functions are completely independent (no overlap), the meter reads 0.
- If they are identical, it reads 1.
- The new formula uses this meter to provide a much more accurate prediction of the total size than any previous formula. It is like upgrading from a general weather forecast to a hyper-local, minute-by-minute prediction.

4. The Role of AI (Grok)

A unique part of this story is how it was written. The authors explicitly credit the AI model Grok with the role of a co-pilot in their discovery.

The "Aha!" Moment: The authors knew a counterexample must exist, but could not find a clean one. They had found one through brute force, but it was messy and random. Grok, however, constructed a counterexample that revealed a clear, structural pattern. This pattern led them directly to the correct mathematical boundary.
The Heavy Lifting: The proofs involved checking dozens of complex polynomial inequalities (mathematical statements about shapes and curves). Doing this by hand would have required months of tedious calculation. Grok helped generate diagrams to visualize the behavior of these functions and verified about 30 different inequalities.
The Human Touch: The authors emphasize that although Grok took over the heavy computational work and visualization, the core idea – the strategy of counting the roots of complicated functions to prove the inequality – came from the human mathematicians. Grok was the tool that enabled them to execute their complex plan efficiently.

Summary

Simply put, this article is about repairing a broken mathematical ruler.

They showed that an old, proposed ruler was inaccurate.
They found the exact settings needed to create a new, perfect ruler.
They solved a specific, tricky 3-element version of the problem with a much sharper formula than before.
They demonstrated a new way of working where human mathematicians provide the strategy and intuition, while AI handles the massive, tedious calculations and visualizations required to prove the theory.

The result is a more precise understanding of how "almost independent" mathematical objects behave when combined, achieved through a successful partnership between human insight and artificial intelligence.

Technical Summary: Fast-Orthogonality in $L^p$ Spaces: A Case Study with Grok

Problem Statement
The article investigates refined forms of the triangle inequality for sums of functions in $L^p$ spaces ( $p \ge 2$ ). In particular, it addresses a question raised by Carbery regarding the conditions under which the sum of a sequence of functions $(f_j)$ belongs to $L^p$ , given that the sum of their $L^p$ norms is finite. Carbery proposed an inequality involving "Fast-Orthogonality" coefficients $\alpha_{jk}$ , defined as:
$\alpha_{jk} := \sqrt{\frac{\|f_j f_k\|_{p/2}}{\|f_j\|_p \|f_k\|_p}} \in [0, 1].$
The central question is whether an inequality of the form
$\left\| \sum_j f_j \right\|_p \le \left( \sup_j \sum_k \alpha_{jk}^c \right)^{1/p'} \left( \sum_j \|f_j\|_p^p \right)^{1/p}$
holds for some exponent $c$ , where $1/p + 1/p' = 1$ . Previous work by Carlen, Frank, and Lieb established related bounds for specific cases and integer $p$ , but the optimal exponent $c$ and the validity of the inequality for general $p$ remained open.

Methodology
The authors employ a hybrid approach combining rigorous mathematical analysis with support from the Large Language Model Grok. The methodology is divided into two main parts:

Construction of a Counterexample: To determine the limits of the proposed inequality, the authors construct a specific counterexample using a measure space with disjoint "private" sets and a single "common" set. This construction, initially generated by Grok and subsequently refined by the authors, demonstrates that the inequality fails for any exponent $c > p'$ .
Proof of Optimality and Validity:
- Integer $p$ : For the critical exponent $c = p'$ , the authors prove that the inequality holds for all integer $p \ge 2$ . The proof relies on a generalized Hölder inequality applied to an integral kernel, as well as a lemma (Lemma 2.3) largely based on suggestions from Grok.
- General $n$ and $p$ : The authors analyze the Carlen–Frank–Lieb inequality for $n$ functions. They reduce the problem of verifying the inequality for general functions to verifying it for indicator functions (Theorem 1.2). This reduction involves analyzing the local maxima of a sum of variables under constraints using Lagrange multipliers to show that extremal configurations involve at most three distinct values.
- Case with Three Functions ( $n=3$ ): For the specific case of three functions, the authors prove that the inequality holds with the optimal exponent $r = \beta(3, p)$ . This section includes extensive technical verification of polynomial inequalities. The authors used Grok to generate plots of auxiliary functions (e.g., $h(s)$ , $Q(s)$ ), verify sign changes of derivatives, and check approximately thirty specific polynomial inequalities. These technical calculations were delegated to appendices to preserve the logical flow of the main text.

Main Contributions and Results

Disproof of Carbery's Conjecture for $c=2$ : The article constructs a counterexample showing that the inequality proposed by Carbery (with exponent $c=2$ ) fails for every $p > 2$ .
Bound for the Optimal Exponent: It is proven that if an inequality of the proposed form holds, the exponent $c$ must satisfy the condition $c \le p'$ .
Validity at the Critical Exponent: Theorem 1.1 states that the inequality holds for all integer $p \ge 2$ when the exponent is set to the critical value $c = p'$ .
Reduction to Indicator Functions: Theorem 1.2 provides a necessary and sufficient condition for the Carlen–Frank–Lieb inequality to hold with the optimal power $r = \beta(n, p)$ : it must hold for indicator functions of a common set.
Sharp Bound for Three Functions: Theorem 1.4 proves the sharp bound for three functions ( $n=3$ ) with $p \ge 3$ , achieving the optimal power $r = \beta(3, p)$ . This improves upon the previously known power $r(p) = \frac{6}{5p-4}$ obtained by Carlen, Frank, and Lieb. The optimal exponent is given by $c(p) = \frac{2 \ln 2}{(p-2) \ln 3 + 2 \ln 2}$ .

Significance and Claims
The article claims to clarify the question of the optimal exponent in Fast-Orthogonality inequalities for $L^p$ spaces, specifically identifying $p'$ as the critical threshold. The authors emphasize that while the conceptual framework (such as the reduction to indicator functions and the analysis of the number of zeros of complex functions) arose from human insight, the computational verification of the numerous technical lemmas was made feasible through collaboration with the AI model Grok.

The authors explicitly state that without AI support, the extensive computational verification required for the case with three functions would likely have been too time-consuming to pursue. They note that Grok's construction of the counterexample revealed a structural pattern leading to the optimal exponent, whereas previous brute-force searches yielded no such insights. The article serves as a case study in human-AI collaboration, where the AI handles routine but tedious algebraic verifications and plot generation, enabling the human authors to focus on the overarching logical structure and conceptual proofs. The authors remain modest, noting that the arguments for $p \in [2, 3)$ are similar but would require the analysis of a larger number of subcases; therefore, the presentation is restricted to $p \ge 3$ .

1. The Broken Rule (The Counterexample)

2. Repairing the Rule (The Correction)

3. The Three-Function Puzzle (The Special Case)

4. The Role of AI (Grok)

Summary

More like this