3Blue1Brown — This open problem taught me what topology is

Why this is in the vault

27-minute Sanderson “second edition” original (December 2024) on the inscribed-rectangle proof, a piece he names as top-5 favorite math in his career. The load-bearing move: to prove every closed continuous loop contains an inscribed rectangle, reframe “find four points forming a rectangle” as “find two unordered pairs of loop points sharing midpoint and length,” then lift that search into a 3D surface whose self-intersections ARE the rectangles. A parallel construction represents “unordered pairs of loop points” as a Möbius strip. The proof hinges on the topological impossibility of embedding a Klein bottle into 3D without self-intersection. The vault keeps it because (1) this is a fourth high-quality source for CA-023 “coordinate change as the core problem-solving move” — promotes it further beyond the already-canon-tier 3-source bar, with Sanderson explicitly naming why the technique works (“topology is a game of understanding continuous associations between things, and understanding what is or is not possible under those associations”); (2) the Möbius-strip-is-not-one-surface admission (a snail-shell embedding by Dan Asimov is just as valid) is a clean lesson in object-identity-under-continuous-equivalence that maps directly to how LMs handle “the same concept in different embeddings”; (3) the Green-Lobb 2020 extension (every smooth curve has inscribed rectangles of every aspect ratio, proven by embedding Möbius strips in 4D) is a live open-problem pointer and an example of the lift-further-when-stuck pattern — if 3D isn’t enough, lift to 4D.

Episode summary

Sanderson proves that every closed continuous curve in the plane contains four points forming a rectangle. The proof: (1) reframe rectangles as “two unordered pairs of points sharing midpoint and length,” (2) construct a 3D surface from the data (x, y, d) for each pair — the surface’s self-intersections correspond exactly to inscribed rectangles, (3) observe that “unordered pairs of points on a loop” is a Möbius strip with the diagonal (x=x) as its edge, (4) observe that glueing the Möbius strip to its reflection across the plane produces a Klein bottle, (5) invoke the known topological fact that Klein bottles cannot be embedded in 3D without self-intersection — QED, rectangles exist. Closes with the still-unsolved inscribed-SQUARE problem (Toeplitz 1911), the Green-Lobb 2020 smooth-curve result (all aspect ratios, proven via 4D Möbius embedding), and a meta-lesson: topology isn’t the study of bizarre shapes for their own sake — shapes like Möbius strips and Klein bottles arise naturally as problem-solving tools, and their “impossibility properties” are fuel for logical proofs by contradiction.

Key arguments / segments

[00:00:00] Problem statement. Toeplitz’s 1911 inscribed-square problem: does every closed continuous curve contain four points forming a square? Still open. The rectangle version (a square with two possibly-different side lengths) is solvable, and the proof is the entire video.
[00:02:30] Sanderson’s thesis about topology. Topology is usually taught as bizarre-shape tourism (Möbius strips, non-orientable surfaces, rubber-sheet geometry). His load-bearing claim: that framing misses the point. Topology is “a game of understanding continuous associations between things, and understanding what is or is not possible under those associations.” The bizarre shapes are tools for proofs; their impossibility properties are the fuel.
[00:03:00] Reframe step 1: rectangles as equal-midpoint-equal-length pairs. Instead of “find four points forming a rectangle,” search for two distinct pairs of points whose connecting line segments share (a) the same midpoint and (b) the same length. The four endpoints of two such segments form a rectangle. This converts a 4-point geometric question into a 2-pair collision question — exactly the kind of move CA-023 flags as the core problem-solving pattern.
[00:04:00] Lift to a 3D surface. For each unordered pair of points on the loop, record (x, y) = midpoint coordinates and d = distance. Package as a 3D point above the loop’s plane. Every pair of loop points → one 3D point. The full set of outputs is a surface. Inscribed rectangles = self-intersections of this surface. (The surface LOOKS Frank-Gehry-weird in 3D, but what matters is that self-intersection has a concrete geometric meaning.)
[00:07:00] The circle special case — sanity check. For a circle, the surface is a dome, and it LOOKS like it has no self-intersection. But actually: all inscribed rectangles in a circle share the center as midpoint and the diameter as length, so infinitely many pairs map to the same 3D point (the top of the dome). The collision is still there — it’s just degenerate. Perturb the circle to an ellipse and the single-point collision becomes a vertical line of self-intersections.
[00:09:30] Chekhov’s gun hanging on the wall. Pair-of-same-point (x, x) maps to (x_loop_coords, 0) — a point on the loop itself, d=0. This detail is load-bearing for the final step but is introduced quietly here. “If in the first act you have hung a pistol on the wall, then in the following one it should be fired.”
[00:10:30] Reframe step 2: unordered pairs of loop points = Möbius strip. Assign each loop point a coordinate in [0, 1] with 0=1 (glued). Pairs of points live in a unit square [0,1] × [0,1]. Glue the boundaries to respect the loop-identification → torus. But we want unordered pairs, so identify (x,y) ≡ (y,x), which folds the square along the diagonal. The diagonal becomes the “both-points-same” locus. After resolving the awkward edge-gluing that remains, the result is a Möbius strip — and the edge of the Möbius strip is precisely the x=x diagonal, corresponding to pairs that map to the loop itself in the 3D surface.
[00:17:00] Combining the two lifts. The 3D surface (from the midpoint-distance map) and the Möbius strip (from the unordered-pairs representation) are both natural ways to organize pair-of-loop-points data. There’s a continuous map from the Möbius strip to the 3D surface. The edge of the Möbius strip must land on the loop itself (because edge pairs have d=0). The interior must land strictly above the plane (d > 0 for distinct points).
[00:18:30] Dan Asimov’s counterexample-that-isn’t. Mathematician Dan Asimov showed Sanderson a construction where a Möbius strip IS embedded in 3D with its edge as a flat circle — the “snail-shell” embedding. So the naive claim (“Möbius strip’s edge can’t be confined to a plane in 3D”) is FALSE. Important detail: Asimov’s embedding has the interior going both above AND below the plane. For inscribed-rectangle purposes we need the interior strictly above. That restricted claim IS impossible.
[00:20:00] The Klein bottle reduction. Take the Möbius strip + its reflection across the plane, glued along the edge. What do you get? A Klein bottle. (Gluing two Möbius strips along their edges is the textbook construction of a Klein bottle.) Klein bottles cannot be embedded in 3D without self-intersection — a famous impossibility result from algebraic topology about closed non-orientable surfaces. Therefore the combined object self-intersects, therefore the Möbius strip hits its reflection somewhere, therefore two distinct unordered pairs have the same midpoint and length, therefore a rectangle exists. QED.
[00:22:30] The inscribed-square extension and 4D. For squares (not just rectangles), also track the ANGLE of the pair’s segment. That’s 4 numbers of data, so the natural surface lives in 4D. Sanderson frames this as the instinct for Möbius-strips-in-4D. Joshua Greene and Andrew Lobb (2020) proved every SMOOTH curve has inscribed rectangles of every aspect ratio using exactly this approach (Möbius strips and Klein bottles in 4D). The rough-curve (fractal-style) case remains open.
[00:24:00] Why smoothness matters: clean limiting behavior of the angle. For smooth curves, the angle of a pair’s segment approaches the tangent-line angle as the pair converges. That’s a well-defined edge-behavior for the 4D surface. For rough curves, no such limit exists — that’s exactly why the problem is hard.
[00:25:30] Closing thesis. Möbius strips and Klein bottles are not shapes — they are topological spaces, infinite families of shapes connected by equivalence under continuous maps. The Asimov snail-shell and the familiar half-twist strip are both the same Möbius strip. Unordered-pairs-on-a-loop is also the same Möbius strip. Topology is about continuous associations and what’s possible under them; impossibilities (like Klein-bottle-embeddings) are fuel for proofs.

Notable claims

[00:02:30] Topology is not “the study of bizarre shapes.” It is “a game of understanding continuous associations between things, and understanding what is or is not possible under those associations.” This reframe is load-bearing editorial for any RDCO writing about abstract mathematical objects in AI (embeddings, manifolds, topology-of-loss-landscapes).
[00:16:00] The Möbius strip’s edge corresponds exactly to the x=x diagonal of unordered loop pairs — i.e., pairs that aren’t really pairs. This identification is the “gun hanging on the wall” that fires at [00:17:00] when the edge must land on the loop itself.
[00:18:30] Dan Asimov’s snail-shell embedding: the claim “Möbius strip’s edge can’t be planar in 3D” is FALSE in full generality. The correct restricted claim — edge planar AND interior strictly on one side — IS impossible, and that’s what the proof needs. A clean example of why a first-pass proof needs stress-testing against degenerate configurations (pairs with CA-022-adjacent “proofs with hidden position-dependence” pattern and with the Five-Puzzles Slovakian-student correction on Monge’s theorem).
[00:20:00] Gluing two Möbius strips along their edges produces a Klein bottle. This is a textbook algebraic-topology fact, but Sanderson derives it visually with cut-and-glue, which is the same technique used to produce the Möbius strip in the first place. Shows how one topology technique scales across proof steps.
[00:22:30] Greene-Lobb 2020: every smooth curve has inscribed rectangles of every aspect ratio, proven by embedding Möbius strips and Klein bottles in 4D. The problem-solving move of “lift to higher dimension when the current dimension isn’t enough” generalizes from 2D→3D (rectangle proof) to 3D→4D (Greene-Lobb), and the technique is the same.
[00:26:00] Möbius strip is not a shape — it is an infinite family of shapes connected by continuous equivalence. Half-twist paper strip, Asimov snail-shell, unordered-pairs-on-a-loop, and “every possible musical interval” are all the same Möbius strip. Topological equivalence is object identity under continuous deformation.

Guests

Grant Sanderson (3Blue1Brown) — solo host; no guests.
Named attributions only: Otto Toeplitz (1911 problem statement); Herbert Vaughan (original rectangle proof); Dan Asimov (snail-shell Möbius embedding counterexample-that-isn’t); Joshua Greene and Andrew Lobb (2020 smooth-curve extension); Anton Chekhov (the gun-on-the-wall quote as a pedagogical device).

Mapping against Ray Data Co

CA-023 “coordinate change as core problem-solving move” now has a fourth canon-tier source. Already promoted to canon-tier at the Five-Puzzles video (3 sources); this video adds a fourth independent example from a DIFFERENT technique family — not a “lift geometric problem to higher-dimensional geometric space” but a “lift combinatorial / geometric problem to a surface in a topological space whose self-intersections encode the answer.” The core move is the same (change what you’re looking at until the answer falls out), but the target is a topological space rather than a higher-dimensional Euclidean space. This generalizes CA-023 one level up: coordinate change isn’t only about picking better axes — it’s about picking a different object entirely (topological space, manifold, category, representation) where the problem becomes visible. Worth folding into the CA-023 concept page (when drafted) as “not just coordinate change, but representation change — pick the mathematical object where the problem’s structure becomes manifest.”
The Möbius-strip-is-not-a-shape admission is the editorial frame for LLM embedding writing. Sanderson’s closing thesis — that Möbius strips are infinite families of shapes connected by continuous equivalence — is exactly the right way to talk about how LMs represent concepts. A token’s embedding vector isn’t the concept; it’s one point in an infinite family of equivalent vectors (under rotation, under compositional context, under paraphrase). RDCO AI writing has historically defaulted to “the embedding IS the concept” language, which is wrong at the same category-level that “the half-twist paper strip IS the Möbius strip” is wrong. Worth writing as a candidate CA-025 “representation-is-not-the-object” concept: one vault-internal source here (this video), one in the five-puzzles video (rhombic dodecahedron = 4D hypercube projection = multiple equivalent instantiations), one from diffusion literature (the image manifold as an equivalence class), and a fourth from LM literature (token embeddings equivalent under orthogonal rotation of the full embedding space). Proposed addition to CANDIDATES.md.
The Asimov snail-shell and the Slovakian-student correction are the SAME pattern — stress-test claimed proofs against degenerate configurations. Both this video and the Five-Puzzles video have a “your intuitive proof is almost right, but here’s the configuration where it breaks” beat. In both cases, the repair is to weaken the intermediate object (Asimov showed “planar edge” isn’t enough — need “planar edge AND interior-above-plane”; the Slovakian coach showed “sphere with tangent plane” doesn’t work — need “cone with tangent plane” which generalizes to center-of-similarity). This is the exact pattern CA-022 is tracking at the generality-under-degenerate-configurations sub-pattern level. Both corrections lean the SAME way: find the configuration that breaks the proof, then find the weakening that still supports the conclusion. For RDCO skills this maps to: when an audit invariant passes for all observed data, stress-test against the configuration you haven’t observed (zero-length input, all-identical input, adversarial input). Specifically add a CANDIDATES.md entry CA-024 (if not already filed from Five Puzzles) for this meta-pattern.
The lift-to-higher-dimension-when-stuck generalization to 4D (Greene-Lobb) is the “if 3D isn’t enough, go to 4D” rule. For RDCO data work the analog is: if a 2-feature scatter plot doesn’t separate classes, go to a 3D embedding; if 3D doesn’t, go to UMAP / t-SNE (which lift to thousand-D then project back); if retrieval doesn’t separate semantically-similar from semantically-different documents, go to a richer embedding model. The rule is don’t give up when the current dimension isn’t enough — first ask whether a higher-dimensional representation has the structure you need.
Sanity Check angle: “Topology Is How You Prove a Rectangle Exists In Every Curve You’ve Never Seen.” Open with Toeplitz’s 1911 problem (“does every closed curve contain a square?” — still open after 115 years). Walk through the rectangle proof at Sanderson’s pace (pairs → midpoint-distance surface → Möbius strip → Klein bottle impossibility). Closing: this is the shape of every impossibility-driven proof in computer science — show that some configuration is topologically ruled out, and the answer falls out by contradiction. Tie to RDCO’s audit layer: audit-newsletter-outputs.py is a deterministic verifier that proves impossibility (a frontmatter format can’t be malformed) and by doing so, proves correctness. ~1,800 words. Strong candidate because the Klein-bottle-self-intersection image is as memorable as any narrative device Sanity Check has used this quarter.
Pairs with the Euclid backfill notes for a broader “impossibility-as-proof” cluster. Euclid-Proposition-1 was about “the construction may fail at a hidden intersection that wasn’t proven to exist” (gap in the proof of general-position configurations). This video is about “the construction of a non-self-intersecting embedding is topologically impossible” (gap between 3D and what you need). Both are impossibility-driven-proof patterns. Worth cross-linking in any future CA-023 or CA-025 concept page.

Open follow-ups

Add CA-025 “representation-is-not-the-object” to CANDIDATES.md. Sources: this video (Möbius strip as infinite family), five-puzzles (rhombic dodecahedron = hypercube projection), pending diffusion/LM sources. Editorial reason: the category-error in “embedding IS concept” language.
Strengthen CA-024 “generality-under-degenerate-configurations” with this video’s Asimov-snail-shell correction. If CA-024 hasn’t been added to CANDIDATES.md yet (per the Five-Puzzles follow-up), this video doubles the source count on the first pass.
When the CA-023 concept page is drafted, incorporate this video as the “representation-change beyond coordinate-change” widener. Sanderson’s explicit “topology = continuous associations” framing is the rhetorical anchor.
Chase Greene-Lobb 2020 paper for the 4D Möbius/Klein construction. If filed in the vault, it would be a good companion source for CA-023 at the 4D level and a direct empirical case for “lift further when stuck.”

~/rdco-vault/06-reference/transcripts/2026-04-20-3blue1brown-topology-open-problem-transcript.md — full transcript
~/rdco-vault/06-reference/2026-04-20-3blue1brown-five-puzzles-thinking-outside-the-box — canon-tier CA-023 companion (3 2D→3D lifts + one 3D→4D lift + Slovakian-student correction that pairs with Asimov-snail-shell here)
~/rdco-vault/06-reference/2026-04-20-3blue1brown-why-colliding-blocks-compute-pi — CA-023 source 1 (physics→geometry sqrt-mass-rescaling)
~/rdco-vault/06-reference/2026-04-20-3blue1brown-how-and-why-to-take-a-logarithm-of-an-image — CA-023 source 2 (complex-log unwrap of Escher)
~/rdco-vault/06-reference/concepts/CANDIDATES.md — track CA-023 4th source + propose CA-025 “representation-is-not-the-object” + reinforce CA-024 “generality-under-degenerate-configurations”