The most beautiful formula not enough people understand
A mathematician presents what he considers one of the most beautiful yet underappreciated formulas in mathematics - the formula for the volume of high-dimensional spheres. He derives this formula using an extension of Archimedes' method and explores the surprising behavior of spheres in higher dimensions.
Summary
The lecture begins with two probability puzzles to motivate the study of high-dimensional geometry. The first involves choosing random points in a square and determining if they fall within a circle, demonstrating how geometric intuition helps solve analytical problems. This extends to higher dimensions, showing the practical relevance of high-dimensional geometry in machine learning and language models like ChatGPT.
The second puzzle reveals how counterintuitive high-dimensional geometry can be. By placing unit spheres at the corners of a cube and finding the largest sphere that fits in the center, the speaker shows that in 10 dimensions, this inner sphere actually extends outside the bounding cube - a seemingly impossible result that highlights how our 3D intuitions fail in higher dimensions.
The core of the lecture focuses on deriving the formula for volumes of high-dimensional spheres using an extension of Archimedes' cylinder method. Starting with familiar 2D and 3D formulas, the speaker builds a systematic approach using what he calls "knight's moves" - jumping two dimensions at a time in a chart relating boundary areas to interior volumes. This method reveals that each step up in dimension involves multiplying by 2π and dividing by the dimension number.
The resulting general formula involves π raised to half the dimension divided by the factorial of half the dimension, requiring an extension of factorial to non-integers through the gamma function. Remarkably, when examining actual numerical values, sphere volumes initially increase with dimension but peak at 5 dimensions, then decrease rapidly. By 100 dimensions, the volume becomes infinitesimally small (around 10^-40), meaning that choosing 100 random numbers and having their squares sum to less than 1 is extraordinarily unlikely.
The lecture concludes by exploring why this happens and what it means geometrically - essentially all volume in high-dimensional spheres concentrates near the boundary, and spheres become negligible compared to cubes. This has practical implications in machine learning, cryptography, and quantum mechanics.
Key Insights
- The speaker argues that high-dimensional geometry is real and useful, not just abstract nonsense, because machine learning represents data as points in high-dimensional space
- Large language models work by converting text chunks into long lists of numbers that researchers interpret as points in high-dimensional space to understand model behavior
- The speaker claims that spheres are not the problem in high-dimensional geometry - cubes are, because spheres are perfectly round by definition while cubes become counterintuitively spiky
- In high-dimensional cubes, corners are much farther from the origin than the edges, creating a situation where unit spheres at corners are 'way, way, way farther away than the edge'
- Archimedes' cylinder method works because two competing effects cancel out perfectly - stretching in one direction and shadow-casting compression in another direction
- The speaker demonstrates that the relationship between sphere volume and surface area follows calculus principles, where the derivative of volume gives surface area
- The volume formula for n-dimensional spheres follows a recursive pattern where each dimension multiplies by 2π and divides by the dimension number, creating 'knight's moves' in a dimensional chart
- Sphere volumes peak at 5 dimensions because that's where the numerator factor 2π and denominator factors are closest, with 2π being slightly larger than 5
- A 100-dimensional unit sphere has volume around 2.37 × 10^-40, making it 'absolutely nothing' despite intuitions about containing lower-dimensional spheres
- In high dimensions, essentially all volume sits right at the boundary - scaling a 10,000-dimensional sphere by 99% reduces its volume by factor 0.99^10,000, which is effectively zero
- The speaker reveals that factorial can be extended to non-integers through the gamma function, where one-half factorial equals square root of π divided by 2
- Mathematical beauty often requires seeing familiar concepts like factorials or circle areas in more general contexts, where hidden connections like π appearing in factorial calculations become visible
Topics
Full transcript available for MurmurCast members
Sign Up to Access