Gaussian propeller — the best partition of space (3D)
About this app
A Gaussian hill, glass walls that split it into sectors, and arrows for the Gaussian centroids of the cells: drag the walls, 2 to 6 sectors, shift the centre, compare half-space, tetrahedron, cube and octants in space. The measuring ring closes only for the propeller of three 120° sectors with value 9/(8π). Result from openai/math (family 096, formalised in Lean): no partition in any dimension gets above it. Exhibit-ready.
What it is about
Split space into parts and weight every point by the Gaussian bell. Each part A gets its centroid z(A) = ∫_A x dγ, without dividing by the part’s mass. How large can the sum Σ‖z(A)‖² be? Khot and Naor met this question in 2008 in their work on kernel clustering and conjectured the propeller: three sectors of 120° each, with value 9/(8π) ≈ 0.3581. For three-dimensional space this has been proved since 2013. An AI-generated preprint in github.com/openai/math (family 096) claims that no partition in any dimension, with any number of parts, does better. The catalogue lists the bound and the equality case as formalised in Lean; the preprint is not peer-reviewed.
What the app shows
Glass walls stand over a Gaussian hill and divide the plane into sectors. Arrows above the peak show the centroids z of the parts, and a gauge ring at the base shows Σ‖z‖² as a fraction of the bound; it only closes completely for the propeller. You can choose 2 to 6 sectors, drag walls by the spheres on top and move the centre. In space the app compares a half-space, tetrahedron, cube and octants. Values are computed for individual partitions; that none exceeds the bound is shown only by the proof.
Try it yourself
Start with three sectors and twist one wall: as soon as the angles move away from 120°, the ring opens. Try four, five or six sectors and move the centre off the peak: more parts do not help. In space, look at the tetrahedron at 0.3532, which stays just under the bound; cases like this made the proof hard. The app can run as an exhibit.
Frequently asked questions
- What is the Gaussian propeller?
- The Gaussian propeller is the partition of the plane into three sectors of 120° each, measured with the Gaussian distribution. For it, the sum of squared Gaussian centroids reaches 9/(8π) ≈ 0.3581. In higher dimensions one takes the product of this partition with the remaining coordinates.
- What does the propeller conjecture have to do with clustering?
- Khot and Naor came across the question while studying kernel clustering, an optimisation problem from data analysis. The largest possible value of the centroid sum determines how well that problem can be approximated. The conjecture that the propeller is optimal is therefore tied to the computational complexity of the problem.
- Has the propeller conjecture been proved?
- Heilman, Jagannath and Naor proved it for three-dimensional space in 2013. An AI-generated preprint in the catalogue github.com/openai/math (family 096) claims the bound 9/(8π) for every dimension and every finite number of parts. The catalogue lists the bound and equality case as formalised in Lean, but not the clustering consequence; the preprint has not yet been peer-reviewed.
- What can I try out in this app?
- You drag glass walls over a Gaussian hill, choose 2 to 6 sectors and move the centre. Arrows show the centroids and a gauge ring shows the value against the bound. In space you compare a half-space, tetrahedron, cube and octants.
Subject: Open problems, solved by AI
More from Open problems, solved by AI
- Anderson localisation: 2D vs. 3D (3D)
- Barker sequences and circulant Hadamard matrices
- Barnette's conjecture: a round trip over every polyhedron (3D)
- Birkhoff billiards: only the ellipse has no gaps
- Cardy's formula — percolation in a random tiling
- The plane is not five-colourable (Hadwiger–Nelson)
- Double-dimer loops and CLE₄
- Triangular billiards — irrational angles mix (3D)
A VisuApp by heyprof: interactive, free in your browser, no sign-up. What is a VisuApp? · All apps