Solving a Research Problem in Mathematical Statistics with AI Assistance

Edgar DobribanarXiv 2025 · DOI · arXiv

In a recent field report from the intersection of artificial intelligence and high-level mathematics, Edgar Dobriban (University of Pennsylvania) documents a significant breakthrough in robust mathematical statistics. By leveraging the reasoning capabilities of GPT-5 Pro, Dobriban solved a previously open problem that had remained elusive for over two years: deriving the minimax optimal error rate for density estimation under Wasserstein-bounded contamination.

The Challenge: Robust Density Estimation

The core of this research addresses the problem of robust density estimation. In classical statistics, one often assumes that data points are drawn directly from a true probability density function $f$. However, in real-world scenarios, data is frequently "contaminated." Unlike the classical Huber model, which focuses on a small percentage of extreme outliers, this paper examines Wasserstein contamination.

In this model, every data point may be slightly shifted. The total "work" required to move the true distribution to the contaminated one is bounded by a Wasserstein distance $\varepsilon$. The researcher's goal was to find the minimax optimal error rate—the best possible performance any estimator can achieve in the worst-case scenario of contamination.

Illustration of density estimation under contamination. The goal is to estimate the true density f (blue) from samples drawn from a perturbed density f-tilde (red dashed).

As shown in the figure above, the estimator must reconstruct the blue curve $f(x)$ using samples $X'_i$ that actually follow the red dashed curve $\tilde{f}(x)$. While previous preprints had established some bounds, the upper and lower bounds did not match, leaving the true rate of convergence unknown until this collaboration.

A Hybrid Human-AI Workflow

The breakthrough was not the result of a single prompt, but rather an iterative Human-AI Interaction Workflow. Dobriban utilized GPT-5 Pro, specifically taking advantage of its "extended test-time reasoning" capabilities. The process involved:

  1. Contextualization: Copy-pasting LaTeX drafts of problem statements and existing proofs into the model.
  2. Refinement: Prompting the AI to "sharpen" specific bounds or identify logical gaps in current arguments.
  3. Reasoning: Allowing the model significant compute time (up to 25 minutes per turn) to process complex logical chains.

Mathematical Breakthroughs via Cross-Pollination

The AI acted as a "creative cross-pollinator," suggesting advanced techniques from optimal transport that the researcher had not previously applied to this specific problem.

  • Upper Bound: The AI suggested using deterministic optimal transport maps (referencing the work of Gangbo and McCann) to better control the bias introduced by contamination.
  • Lower Bound: The critical "aha moment" came when the AI suggested the dynamic Benamou-Brenier formulation of optimal transport. This allowed the researcher to relate the Wasserstein distance to a "kinetic energy" integral, providing the mathematical precision needed to match the upper bound.

This connection was pivotal. While these tools might be standard for an optimal transport expert, they were niche within the context of nonparametric statistics, demonstrating the LLM's advantage in accessing a vast, multi-disciplinary memory of mathematical literature.

Key Results and Implications

The collaboration successfully established the optimal rate for $s$-Hölder smooth densities in $p$ dimensions under $\varepsilon$-bounded $W_{q,r}$ contamination. The minimax risk was found to scale as:

$$\max \left\{ n^{-\frac{2s}{2s+p}}, \varepsilon^{\frac{2s}{s+1+p/r}} \right\}$$

Beyond the formula itself, the research yielded two practical insights:

  1. Optimality of KDE: The results prove that standard Kernel Density Estimation (KDE), when paired with a carefully tuned bandwidth $h^*$, is minimax optimal even under Wasserstein contamination.
  2. Efficiency Gains: The author estimates that the AI saved several months of manual literature review and trial-and-error discovery.

The Future of AI-Augmented Research

Despite the success, the report highlights significant limitations. The AI frequently produced "hallucinations," such as incorrect references or skipped technical steps that required days of human effort to rectify. Furthermore, the AI exhibited a "persistence trap," where it would generate increasingly convoluted (and still incorrect) fixes when challenged on a mistake.

Ultimately, this work suggests that human expertise remains indispensable. The AI serves as a powerful augmentative tool for high-level researchers who possess the foundational knowledge to verify and refine its outputs. This case study serves as a template for a new era of collaborative mathematical discovery, where the ability to navigate AI-assisted search and verification becomes a core competency for the modern scientist.

Summary updated

Abstract

Over the last few months, AI models including large language models have improved greatly. There are now several documented examples where they have helped professional mathematical scientists prove new results, sometimes even helping resolve known open problems. In this short note, we add another example to the list, by documenting how we were able to solve a previously unsolved research problem in robust mathematical statistics with crucial help from GPT-5. Our problem concerns robust density estimation, where the observations are perturbed by Wasserstein-bounded contaminations. In a previous preprint (Chao and Dobriban, 2023, arxiv:2308.01853v2), we have obtained upper and lower bounds on the minimax optimal estimation error; which were, however, not sharp. Starting in October 2025, making significant use of GPT-5 Pro, we were able to derive the minimax optimal error rate (reported in version 3 of the above arxiv preprint). GPT-5 provided crucial help along the way, including by suggesting calculations that we did not think of, and techniques that were not familiar to us, such as the dynamic Benamou-Brenier formulation, for key steps in the analysis. Working with GPT-5 took a few weeks of effort, and we estimate that it could have taken several months to get the same results otherwise. At the same time, there are still areas where working with GPT-5 was challenging: it sometimes provided incorrect references, and glossed over details that sometimes took days of work to fill in. We outline our workflow and steps taken to mitigate issues. Overall, our work can serve as additional documentation for a new age of human-AI collaborative work in mathematical science.
BibTeX
@misc{dobriban2025solving,
  author = {Edgar Dobriban},
  title = {{Solving a Research Problem in Mathematical Statistics with AI Assistance}},
  year = {2025},
  eprint = {2511.18828},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.18828},
}
Figure