Applied Pure Mathematics

Some thoughts about mathematics as a cultural and social technology.

Hi there, argmin readers! As the fall semester picks up, posting volume will, too. So I’m going to commit to writing short descriptive headers to help you sort through the different threads.

Many readers have asked me to write about AI companies’ conquest of mathematics. Today’s post is a first, but by no means final, attempt at grappling with our new mathematical condition.

---

Early in my career, I was fortunate to get caught up in a fascinating research frenzy at the intersection of pure and applied math, the compressed sensing gold rush. Compressed sensing asked whether signals could be compressed at the time of measurement. Rather than sampling an image with a high-resolution camera and then compressing it to a JPEG, could we collect a number of samples equal to the number of bytes in the JPEG? Compressed sensing rested on deep mathematics from geometric functional analysis, convex geometry, and probability theory. It yielded multiple engineering artifacts, from faster MRI capture times to better systems for content recommendation.

Though we can see the influence of the field across many applied domains, the math of compressed sensing was never decidedly prescriptive. The theorems always assumed things about reality that couldn’t be verified or required measurement systems that were too costly or impractical. Yet the math of compressed sensing helped us focus on a shared narrative of design principles. It helped us design new algorithms. It helped us construct new measurement schemes that were robust to noise. It helped us map out which other system structures were amenable to compressive techniques. Pure math gave us a frame to see what was possible.

While this mathematical formalism was unreasonably effective, it came with a decidedly unhealthy downside. Shahar Mendelson best described this general problem of applied pure mathematics in a talk he gave at COLT 2014. Applied mathematicians often need to build a giant scaffolding of mathematical modeling to solve a problem. This scaffolding creates new mathematical puzzles that aren’t directly connected to the original problem of interest, but that entice problem solvers. You’ll then see dozens of follow-up papers solving the puzzles but forgetting the problem we cared about in the first place.

This is open problem culture, and it’s corrosive. It leads to trophy hunting, where people race to scoop each other, consult expert friends for secret insights, or steamroll each other with ever more complicated math.

This fetishization of puzzle-solving as genius has long been a destructive tendency in mathematics more broadly. It’s easy to get caught up in the thrill of it. Mathematics is arguably the most meritocratic academic discipline. There are set problems, and the people who solve them are the smart ones. Everyone forgets that the only reason problems confer status is that (a) they are currently unsolved and (b) enough mathematicians have decided these are worth solving. That (b) part is not meritocratic.

This is why many are confused and angry at the practicing mathematicians who try to explain that the discipline of mathematics is about understanding, not proving stuff. To many observers, even those who strive to become mathematicians, math seems set up as a competition from the get-go. It’s rote testing all the way up through college. Ace the SAT as a 7-year-old. Win the IMO gold as a 14-year-old. Max the Putnam Exam as a 19-year-old.

Your reward is the permission to work on whatever puzzles you want, without questions, for the rest of your life. There is no requirement for the winners to explain anything. Maybe they have to teach calculus, but they don’t have to do a good job at it.

From the outside, you can see why people think mathematics is just about winning those competitions and proving what is true. Math doesn’t send many outward signals that “understanding” is a core part of the pursuit. Most people see math as a quiz show culture. Math culture is ruthlessly competitive, and it makes a lot of people feel stupid.

The actions of many notable mathematicians have only lent credibility to their critics. Wars over credit and who gets there first have now ruined two of Clay’s Millennium Problems. This will have to change in light of recent events with AI companies solving math problems few thought they’d be able to. When computers do something we think they wouldn’t, the reaction should not be writing insanely long posts about how Eliezer Yudkowsky was right and the machines are going to kill everyone. Instead, we have to adjust our reference narrative about what we thought was true.

Indeed, I didn’t learn anything about fluid dynamics from OpenAI’s proposed solution to the Clay Millennium Prize Navier-Stokes problem. This problem is exactly the sort of puzzle artifact that I lamented above. The resolution of the Navier-Stokes problem itself tells us nothing about the dynamics of fluids that the equations attempt to model.

That said, I’ve learned a lot from the supposed resolution. I learned that the jump from rote IMO solving to the Millennium Prizes was much shorter than I expected. If you build an algorithm that’s good at solving IMO problems, and you present it with the right ingredients and computational resources, you can solve hard math problems too. That is, a lot of mathematics is training people to benchmaxx. We have already created a battery of tests, carefully tuned with the best psychometrics to find mathematical genius. Training computers to maximize those benchmarks ends up solving the benchmarks. What are millennium problems other than humanity’s final math exam?

This unfortunately makes a lot of sense with the benefit of hindsight!

If this is the lesson, there’s a funny takeaway. While it feels like you need to be an IMO prodigy to set foot in the mathematical arena, being a great IMO solver doesn’t mean you’ll become a great mathematician. For that, you need to bring other talents to bear. Despite the efforts of many smart and caring people, those talents remain ineffable. They certainly aren’t benchmarkable.

In an age of the decidedly anti-intellectual culture of artificial intelligence, mathematicians, both pure and applied, need to keep working to articulate what on earth those talents are. The statements so far, describing how mathematical programs are more than the truth values of their associated theorems, are a good start even if they are not met with universal acclaim. More need to chime in with stories about how mathematics, even the very pure variety, is valuable for scientists, engineers, and everyone else.

I can describe my own experience. Though I’m much less concerned with proving theorems than I was earlier in my career, I still consider myself an applied pure mathematician. Applied mathematics is a formal language that bridges two unbridgeable worlds. Mathematics is a deductive practice that combines axioms via a set of well-specified rules to generate lemmas, theorems, and corollaries. Empirical science and engineering are inductive. We confirm theories when they make correct predictions, willfully committing the logical fallacy of affirming the consequent. This does not make science wrong. It just means, as David Hume told us three hundred years ago, that mathematics can’t justify science.¹

Applied mathematics is thus a logical language for describing inductive processes. It’s, um, unreasonably effective at this task. As captured above in my discussion of compressed sensing, it can never perfectly specify what you should do in practice. Instead, it acts as a form of linguistic technical drawing, allowing communities of scientists to build complex theories and engineers to build complex systems. Pure mathematics gives applied mathematicians new pens and brushes for those drawings.

This is why I like (and have been using throughout) Jordan Ellenberg’s term applied pure mathematics. Applied mathematics often just means the mathematics of partial differential equations. Applied pure mathematics is any application of any mathematics to anything outside of the closed world of mathematics itself. You never know which weird corner of the vast libraries of “apparently useless” mathematics will help you make sense of reality.

Let me give an example of unexpected brushwork from my time in the compressed sensing gold rush. Did I need to learn p-adic analysis as an undergrad? Maybe not, but it fixed a set of regularities and patterns in my head. I remembered Bochner’s theorem on locally compact abelian groups when Ali Rahimi and I were trying to make sense of our code generating random features. This turned into a very cool paper with a lot of practical impact. The web of facts I had gathered sitting through weird courses and reading esoteric math books shaped how I saw this applied machine learning problem. AI could likely make that connection today, but my personal education is still needed to create the prompt.

In the first lecture of my first college math course, the legendary Chicago Professor Paul Sally (IYKYK) barked that he wasn’t there to teach us facts, but to fix our brains. Sally dedicated his career to mathematics education, passionately broadening the conception of who could be a mathematician. Math wasn’t a competition for Sally. It was a way of seeing. It still can be, even if our computers now outcompete us.

1. A popular argument on social media is that once mathematics falls to AI, all the sciences will follow. This may end up being true eventually. Mathematics has certainly been disrupted in a shocking way this summer, but science has not (yet). However, it can’t follow logically.