From Gaussian shells to generation

A landscape you can climb

Spread noise around a center. Combine the scales into a mountain. Add more centers, then let the gradient guide the climb.

Equal draws occupy very different volumes

Try

Where the draws land

Radial schematic, not a two-dimensional projection. Faint rings show all 32 scales; colored dots show equal raw counts at three representative scales.

Density and probability tell different stories

The density peaks at the center. Volume moves radial probability outward. Both profiles describe individual Gaussian components.

The density and the two update rules

We use 32 logarithmically spaced noise scales and the exact weighted mixture pMN(x) ∝ ∑j,i σi² 𝒩d(x; μj, σi²I). Its score is s(x) = β(x)(μ̄(x) − x). The landscape remains fixed during a run.

An Euler update takes a finite straight step using the current field. Gradient ascent: xn+1 = xn + ηs(xn). A large η may overshoot and decrease density. Normalize by β: xn+1 = xn + η(μ̄(xn) − xn). This exact-mixture update increases density for 0 < η < 2. Normalizing preserves continuous integral curves; finite Euler steps need not land on identical paths.

Changing a parameter reuses the same initial points and recalculates the trajectories. Escaping the visible window is reported rather than corrected. Tracks that become nonfinite or exceed radius 20 stop, and remain counted as outside. The shell view uses individual Gaussian components; the mountain view includes the σ² mixture weights. At every positive noise scale, a Gaussian density is positive everywhere.

The far-field profile uses a single center and the same 32-scale mixture. The outer probe is at radius 1.8√d, and the profile extends to 2.6√d. Its effective scale is 1/√β at that position. The reference broad cutoff is max(σmin, 0.80), with the same lower cutoff. Each log-density curve subtracts its own center value; no change in the score or density is added for display.