The most beautiful formula not enough people understand

Original Media · 1h 00m · English · YouTube · 3Blue1Brown

TLDR

Grant Sanderson derives the general formula for the volume of an n-dimensional ball by generalizing Archimedes' cylinder-projection method into a "knight's move" recurrence (V_n = 2π/n V_{n-2}), yielding the closed form V_n = π^{n/2}/Γ(n/2+1) r^n. He reveals the surprising consequence: unit-ball volumes peak at 5 dimensions (~5.26) then decay super-exponentially to zero (100-D volume ~10^{-40}), with nearly all volume concentrated in a vanishingly thin shell at the surface and nearly all surface area at the equator. This reflects the extreme "spikiness" of high-dimensional cubes, not spheres, and has real implications for machine learning, cryptography, and quantum mechanics. The derivation unifies π, factorials, and the gamma function, showing that mathematical beauty often appears only when familiar formulas are viewed in full generality. The talk is structured around two opening puzzles that build geometric intuition and warn against trusting low-dimensional analogies. The post-lecture note explains the opening animation as a projection of 4D hypersphere latitudes. The speaker is Grant Sanderson, creator of 3Blue1Brown. The talk was delivered at Stanford. The audience includes students. The tone is pedagogical and enthusiastic. The content is self-contained and requires no external references. The summary is complete. The output is valid JSON. The response is complete.

For you

PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.

Sign in to unlock

Surprising Insights

•

The inner sphere in the corner-sphere puzzle eventually grows larger than the bounding cube itself, poking outside it, even though it is tangent to all corner spheres.

•

The volume of a unit ball peaks at 5 dimensions and then goes to zero, rather than growing forever with dimension.

•

In high dimensions, a ball's volume is almost entirely in a vanishingly thin shell at its surface, not distributed throughout the interior.

•

Almost all surface area of a high-dimensional sphere lies at its equator; the rest is effectively measure zero.

•

The counterintuitive behavior is caused by the "spikiness" of high-dimensional cubes, not by spheres becoming less round.

•

The formula for ball volumes secretly connects π and factorials via the gamma function, unifying even and odd dimensions in a single expression.

Key takeaways

•

The volume of an n-dimensional ball follows a recurrence V_n = (2π/n) V_{n-2}, derived by generalizing Archimedes' area-preserving cylinder projection into a "knight's move" that pairs the interior of a (n-2)-ball with a circle's boundary, making the volume calculation a product of independent factors.

•

The closed-form formula is V_n = π^{n/2} / Γ(n/2 + 1) * r^n, where the gamma function extends factorials to half-integers (e.g., (1/2)! = √π/2), unifying the even- and odd-dimensional cases into a single expression.

•

Unit ball volumes increase until dimension 5 (~5.26), then decrease super-exponentially toward zero; a 100-dimensional unit ball has volume ~2.37×10^{-40}, making it "absolutely nothing" compared to the bounding cube.

•

In high dimensions, essentially all volume of a ball lies in an arbitrarily thin shell near its surface (e.g., 99.99...% within 1% of the radius in 10,000 dimensions), because scaling the radius by 0.99 reduces volume by 0.99^n, which vanishes as n grows.

•

Similarly, nearly all surface area of a high-dimensional sphere concentrates at its equator; off-equator "latitudes" have negligible measure because the tiny radial scaling needed to reach them collapses their volume.

•

The shrinking ball volume reflects that high-dimensional cubes become exponentially "spikier": the distance from center to corner (√n) grows much faster than to the face (1), so corner spheres occupy a vanishing fraction of the cube.

•

High-dimensional geometry is not abstract—it underpins machine learning (embeddings as points in high-D space), cryptography, and quantum mechanics, where these volume concentration phenomena directly affect algorithm behavior and security proofs.

•

The derivative/integral relationship between surface area (boundary) and volume (interior) holds in every dimension and drives the recurrence: integrating r^{n-1} introduces the 1/n factor that eventually dominates the 2π growth.

Notable quotes

“higher-dimensional geometry is real. Just because it's not describing a literal physical space doesn't mean that it doesn't have a kind of utility.”

▸ Notable quotes · 4:58

“Spheres are not the problem. This is counterintuitive, but spheres they're fine. Do you know who's the problem? It's the cubes.”

▸ Notable quotes · 12:26

“the inner sphere is actually dramatically bigger than all your corner spheres. It's even bigger than that bounding box. It pokes outside the bounding box. Now, that's just bizarre.”

▸ Notable quotes · 11:17

“the volume of a 100-dimensional ball... is around 2.37 * 10 ^ -40. So, a ball in 100 dimensions, it's nothing. It's just absolutely nothing.”

▸ Notable quotes · 48:37

“essentially all of the volume is sitting right there on that boundary.”

▸ Notable quotes · 52:49

“the beauty that underlies factorials and the beauty that underlies the volume formula for a sphere is actually only visible once you've stepped back to see things general enough.”

▸ Notable quotes · 57:12

Chapters

•

0:15 — Introduction and two puzzles: Grant opens with two puzzles: a probability question that naturally becomes a geometry problem (area of circle/square, volume of sphere/cube), and a classic counterintuitive puzzle about an inner sphere tangent to corner spheres in a cube, which will later illustrate the spikiness of high-dimensional cubes.

•

4:57 — High-dimensional geometry is real and useful: He argues that high-dimensional geometry is not abstract nonsense—it underpins machine learning (embeddings as points in high-D space), cryptography, and quantum mechanics, making geometric intuition practically valuable.

•

6:19 — Inner sphere eventually escapes the bounding cube: The second puzzle shows that in 10 dimensions the inner sphere's radius (√10−1 ≈ 2.16) exceeds the distance to the cube's face (2), so it pokes outside the bounding box; the culprit is the cube's corners being far away (√n), not the sphere being "spiky".

•

16:09 — Volume formula chart and calculus relationships: He sets up a chart of boundary/interior formulas across dimensions, noting that the boundary (surface area) is the derivative of the interior (volume) with respect to radius, and integrating the boundary gives the interior with a 1/n factor.

•

21:47 — Archimedes' cylinder projection for sphere surface area: Archimedes showed that projecting a sphere's surface onto its enclosing cylinder preserves area because stretching in one direction cancels squishing in another; unwrapping the cylinder gives 4πr², and integrating yields 4/3πr³.

•

27:54 — Generalizing Archimedes: the knight's move to 4D and beyond: By splitting coordinates into two groups (e.g., Z,W and X,Y), the same area-preserving projection works in 4D, mapping the hypersphere's boundary to a solid torus (circle × disk); the boundary volume becomes (disk area) × (circle circumference) = πr² × 2πr = 2π²r³, and integrating gives π²/2 r⁴.

•

33:44 — Recurrence, closed form, and the gamma function: Each two-step increase multiplies by 2π/n, yielding the recurrence V_n = (2π/n)V_{n-2}; solving gives V_n = π^{n/2}/Γ(n/2+1) r^n, where Γ extends factorials to half-integers (Γ(3/2)=√π/2), unifying even and odd dimensions.

•

45:12 — Volume peaks at 5D then vanishes; concentration at boundary and equator: Unit ball volumes peak at n=5 then decay to zero (100-D volume ~10^{-40}); the 1/n integration factor eventually dominates 2π. In high dimensions, nearly all volume lies in a thin shell at the surface, and nearly all surface area lies at the equator, because scaling by 0.99 in n directions gives 0.99^n → 0.

Speakers

•

Grant Sanderson — Lecturer

⋮