Theoretical Foundations, Method of Moments, and Quantum Learning: An Interview with Ankur Moitra

An interview with Ankur Moitra, recipient of the 2026 W. Wallace McDowell Award.

Ankur Moitra is the Rockwell International Career Development Professor at MIT and Director of the Statistics and Data Science Center, whose work in theoretical computer science has provided foundational algorithmic frameworks for understanding high-dimensional data and machine learning.

We connected with Dr. Moitra to discuss the transition from mastering material to generating research, the necessity of provable guarantees in an era of "black box" AI, and the multidisciplinary metaphors driving the future of quantum learning.

When you started your PhD at MIT, what was the biggest mental shift required to move from "mastering known material" to "generating original research" in theoretical computer science?

The biggest mental shift was in raising my own bar for what it means to understand something. When I first started working with my advisor, Tom Leighton, I would explain papers I’d read. He hadn’t read them. But he had this extraordinary ability to distill the proofs down to their essence. It impressed upon me the importance of building intuition and world models for how things work.

In a field often dominated by empirical "black box" models, you advocate for algorithms with provable guarantees. For a junior machine learning engineer used to tuning hyperparameters by trial and error, why is a theoretical foundation critical for the long-term reliability of AI?

When you play around with AI models, they work amazingly well. It’s like an alien technology. But when you ask why it works, you often get fuzzy answers about inductive biases, overparameterization and scaling laws. For me, theory is about building abstractions that help us interrogate our understanding of what’s really going on.

You’ve done extensive work on the Method of Moments for learning latent variable models. Could you explain, in a way that an early-career data scientist could grasp, why this classical statistical technique is suddenly so powerful for modern high-dimensional data?

The method of moments is powerful because it balances statistical and algorithmic considerations. It is strong enough to be able to identify the parameters of many sorts of latent variable models from a small number of samples. And yet it can often be computed efficiently, even in high-dimensions, through a growing war chest comprised of spectral and tensor methods and semidefinite programming hierarchies.

One of your newer research areas is quantum learning. For someone just starting their career in it today, what advice would you give them? For example, do you view quantum computing as a tool that will replace classical machine learning algorithms, or as a specialized niche for specific problem sets? Might that change in the coming years?

Quantum learning is blessed with problems that require truly multidisciplinary perspectives to make progress on. For me, I was drawn to it because I felt that there were missing metaphors with classical learning that helped me see different paths forward. I still think that’s true and maybe we’re just getting started. In that spirit, I would advise young researchers in the field to learn to speak and translate between multiple languages, from physics to computing to mathematics, fluently.

You have been heavily involved in summer research programs for high school and undergraduate students. What is the one thing you wish every young student knew about the "joy of research" before they decide whether or not to pursue a PhD?

Most students get frustrated working on their first research project. The answers just don’t come as quickly as they’re used to. It takes some time and experience to realize that that’s the way it’s supposed to be. When good problems fight back, sometimes you end up discovering more beautiful things than you would have ever dreamed of at the outset.