“How (and why) to take a logarithm of an image” — 3Blue1Brown

Episode summary

Grant’s full-length companion to the 102-second Escher’s Print Gallery teaser filed under 2026-03-27-3blue1brown-eschers-print-gallery. The video is a 45-minute tour of complex analysis framed as reverse-engineering Escher’s 1956 lithograph “The Print Gallery” — the self-referential piece where a man looks at a picture containing the gallery that contains the man. The de Smit–Lenstra (2003) paper showed that Escher’s distortion is, formally, the image you get by applying the complex logarithm to an unwrapped undistorted version — i.e. the distortion is a conformal map. Grant builds the full machinery from scratch: what a conformal map is (Ch. 13:04), the complex exponential (21:41), the complex logarithm (25:56), and then the explicit construction of the key function that reproduces Escher’s distortion (33:14). The closing section (40:16) argues this is the pattern Escher intuited without the math — and that the “blank spot” in the center of the lithograph is the unavoidable fixed point of the recursive map. Co-written with Paul Dancstep; artwork by Talia Gershon, Mitchell Zemil, and Anna Fedczuk.

Key arguments / segments

[0:00] The Print Gallery lithograph — the puzzle, the blank spot, the man-contains-gallery-contains-man recursion
[13:04] Conformal maps: transformations that preserve angles locally, even when they distort distances and shapes globally
[21:41] The complex exponential z → e^z: what it does to horizontal strips, vertical lines, and why the output “wraps”
[25:56] The complex logarithm z → log(z) as the inverse — takes an annulus to a strip, and takes the plane punctured at zero to a multi-valued infinite strip
[33:14] Constructing the key function: unwrap the scene into a strip, then apply the exponential with a specific complex scaling factor to get exactly Escher’s distortion
[40:16] The blank spot is the fixed point of the recursion — no finite image can exist there because the map is infinitely zooming into itself at that point
Continuous closing meditation: Escher didn’t have the math, he had the intuition; the math came 47 years later and confirmed the structure was precisely a complex logarithm

Notable claims

Escher’s Print Gallery distortion is mathematically equivalent to applying the complex logarithm to a straightened version of the scene — this is a theorem, not an analogy (de Smit & Lenstra 2003)
Conformal maps are the “right” class of transformations for this kind of problem because they preserve local angular structure even while globally deforming — the eye perceives them as “curvy but coherent”
The blank spot at the center of the lithograph is a topological/analytical necessity, not an artistic choice — the map has a fixed point there and any attempt to fill it would be inconsistent with the surrounding structure
Intuition-first artists can sometimes arrive at mathematical structures decades before the mathematicians formalize them; the inverse direction — math-first derivation of artistic structure — is much rarer

Transcript availability caveat

YouTube did not expose an English auto-caption track for this video at ingest time (cycle 41, 2026-04-20). The assessment above was built from the chapter list, description, the de Smit/Lenstra paper reference, and the companion 102s teaser already in the vault. A Spanish manual VTT was available and could be retrieved if a deeper citation-level pass is needed.

Why this is in the vault

The vault’s position on representation learning and feature engineering leans hard on the claim that the choice of coordinate system is the whole game — pick coordinates where the problem looks like something you already know how to solve, and the work collapses. This video is the cleanest pure-math instance of that principle in the vault. “Take the logarithm of an image” is, literally, a feature transformation: the log takes a multiplicative/recursive/rotational structure and makes it additive and linear. That’s exactly what log-transforms do to skewed distributions, what SVD does to correlated features, and — at the conceptual level — what embeddings do to tokens. Escher-via-log is a more memorable hook than any of those, and it works as the on-ramp for a Sanity Check piece on “when a transformation makes a hard problem trivial.” Second reason: the newsletter frequently argues that domain experts’ intuition precedes formalization; the Escher-anticipates-de-Smit-and-Lenstra story is a perfect illustrative anecdote.

Mapping against Ray Data Co

Novel data-transformation technique as rhetorical hook for feature engineering. The “log of an image” framing is precisely the kind of shock-of-recognition opener Sanity Check uses for technical pieces. A reader who understands why the complex log unwraps Escher’s recursion has implicitly understood why log-transforming a price series unwraps multiplicative dynamics, why polar coordinates unwrap rotation, and why embedding spaces unwrap semantic relationships. Mapping strength: strong / direct rhetorical reuse.
Representation-transforms as the right level of abstraction for “what is an embedding.” When the newsletter next needs to explain embeddings, vector databases, or why dimensionality reduction isn’t “just compression,” this video’s treatment of how a well-chosen map takes a hostile geometry (Escher’s self-reference) into a friendly one (a flat strip) is the cleanest available metaphor. Mapping strength: medium / conceptual-reuse.
Conformal maps as a frame for “preserve what matters, distort what doesn’t.” In product/metrics work, this maps onto the choice of which invariants to enforce in a transformation pipeline. Mapping strength: light / speculative.

2026-03-27-3blue1brown-eschers-print-gallery
2026-04-20-3blue1brown-why-colliding-blocks-compute-pi
2026-04-20-3blue1brown-exploration-epiphany-paul-dancstep
2026-04-20-3blue1brown-but-how-do-ai-images-and-videos-actually-work