The Shape of a Proof

Two Languages, No Dictionary

Since the 1800s, mathematicians have studied knots — closed loops in three-dimensional space that cannot be untangled without cutting the string. Picture a trefoil, the simplest nontrivial knot: it looks like a three-leaf clover made of rope, with the ends fused together. You can rotate it, stretch it, push it around in space — but you can never transform it into a simple circle, no matter how hard you try. That impossibility is the whole point.

What makes knot theory so strange is that mathematicians have developed two fundamentally different languages to study the very same objects. The first is geometric. When you look at a knot floating in three-dimensional space, you can measure things about the complement — the space around the knot. This space has a natural hyperbolic geometry, a kind of negative curvature. You can calculate its volume, trace out geodesics (the straight lines in hyperbolic space), measure angles and rotations. These are the geometric invariants: the hyperbolic volume, the natural slope, the curvatures.

The second language is algebraic. Instead of thinking about the knot's shape in space, you think about its crossing pattern — where the rope crosses over and under itself. You can encode this crossover information into abstract polynomials, equations that exist only on paper. The most famous is the Alexander polynomial, discovered in the 1920s. Another is the signature — a single integer that captures essential crossing information. These are the algebraic invariants: symbols and numbers that encode topology as pure algebra.

For decades, topologists suspected that these two languages must be connected. They study the same knot, after all. How could there be no relationship between geometric measurements and algebraic properties? Yet the datasets were too vast and too complex. With hundreds of knot invariants scattered across both domains, finding patterns by hand was like searching for a specific grain of sand on an entire beach — and not knowing if that grain even exists.

A Machine That Reads Patterns

In 2021, researchers at DeepMind collaborated with topologists at the University of Oxford — Professor Marc Lackenby and Professor Andras Juhasz — and researchers at the University of Sydney to ask a bold question: could machine learning help mathematicians find patterns in invariant data that human intuition could not? The resulting paper, titled "Advancing mathematics by guiding human intuition with AI," was published in Nature in December 2021.

The approach was elegant in its simplicity. The team assembled a large dataset of known knots — roughly 3,000 knot types with up to 12 crossings — and computed both their geometric and algebraic invariants. They then trained a supervised machine learning model on this data. The task given to the model was straightforward: given the geometric invariants of a knot, predict its algebraic signature. The model learned the association. When tested on new knots it had never seen, it predicted the signature better than chance — a genuinely non-trivial achievement, since no such relationship had been known before.

But here's where the method became genuinely clever. The model's success alone proves nothing mathematically. A neural network can recognize patterns without understanding why those patterns exist. The team needed to know which geometric features the model was actually using to make its predictions. To answer this, they applied attribution techniques borrowed from machine vision research — the same methods used to identify which pixels matter in an image recognition task. They essentially asked the model: "Show us your work. Which features are you paying attention to?"

The model's answer was clear and unambiguous. Among all the geometric invariants available, one stood out. The model was focusing on a specific measurement: something that would soon be known as the natural slope. This is a real number derived from the hyperbolic geometry of the knot's complement — specifically, it measures how much a geodesic on the boundary rotates as it travels around the knot once. Before this analysis, the natural slope was not considered particularly important. It was a known quantity, but a quiet one, buried in the mathematical literature. The AI's attention mechanism had revealed it as the key.

The Bridge That Was Always There

The natural slope is not an intuitive quantity. It exists in the hyperbolic geometry of the knot complement — a space with negative curvature where parallel lines diverge and triangles have angles that sum to less than 180 degrees. In this space, you can trace geodesics: the analog of straight lines in hyperbolic geometry. The natural slope measures a specific rotation property of these geodesics. It is a real number, sometimes irrational, derived entirely from the geometry of space surrounding the knot.

What the AI revealed was that this geometric quantity — this number buried in the curvature and topology of hyperbolic space — was the strongest predictor of the algebraic signature. The signature is an integer encoding the knot's crossing pattern. Geometry and algebra, two entirely separate mathematical languages, were connected through a single bridge: the natural slope. But the AI finding this correlation was only the beginning. A correlation is a pattern; a theorem is a proof.

Armed with this insight — this mathematical compass needle pointing at the natural slope — Professor Lackenby and his colleagues returned to traditional mathematics. They asked: why would the natural slope predict the signature? What logical chain connects these two invariants? And crucially, they set out to prove it. The result was a rigorous mathematical theorem, proven in the classical sense with formal deduction, lemmas, and logical necessity. The AI had found the question. The humans had found the answer.

This distinction is crucial. The machine learning model did not discover a theorem. It discovered a pattern. The pattern then served as a compass for human mathematicians, pointing them toward a truth that, once found, could be proven with complete certainty. This is not automation replacing human mathematics. It is a new kind of scientific instrument — a pattern-detection microscope for datasets too large and too complex for human intuition to navigate alone.

A New Kind of Mathematical Instrument

Pure mathematics has long seemed like the last fortress against artificial intelligence. It is not pattern recognition; it requires rigorous logical deduction. Mathematical truth is absolute — you cannot convince a theorem to be different, and no amount of training data can change what is provably true. Yet DeepMind's collaboration with Oxford topologists demonstrated something unexpected: AI can function as a mathematical microscope, revealing structures that human mathematicians have collected data about but cannot intuitively process.

The philosophical significance runs deep. For centuries, mathematics was understood as an entirely human endeavor: creative, intuitive, the product of genius and insight. Computers were thought to be mere calculators, tools for verification but not discovery. What the knot theory result shows is that there is a middle ground. The machine cannot replace the rigor of proof, but it can navigate high-dimensional spaces of mathematical objects in ways no human mind can match. It can ask: across 3,000 knots and hundreds of invariants, which geometric features correlate most strongly with which algebraic properties?

The paper published in Nature in December 2021 did not stop with knots. The same team applied their method to representation theory — an entirely different area of mathematics dealing with abstract algebraic structures. They discovered a second connection, a second bridge between domains that had seemed disconnected. This matters because it suggests the approach is not a one-off curiosity. It is generalizable. The method works across mathematical fields.

Consider the implications. Every mathematical field produces invariants — properties that capture essential structure. Number theory has invariants. Geometry has invariants. Algebra has invariants. Probability theory has invariants. These invariants often sit in isolated silos, studied by specialists within each field. But what if there are hundreds or thousands of hidden connections between these silos? What if there are bridges between number theory and geometry, between algebra and topology, that no human has yet discovered because the datasets are too large, the patterns too subtle, the dimensional space too vast for human intuition alone?

This is the deeper implication of the 2021 discovery. The question is no longer whether AI can help discover mathematics. It already has, rigorously, in Nature, with full peer review. The real question is about infrastructure: What training datasets should we prepare? Which attribution techniques are most effective for mathematical discovery? How do we systematically apply machine learning to guide human mathematicians toward theorems waiting to be proven? The answer to these questions could reshape how mathematics develops over the coming decades.