How many random points does a disk need?

Draw N points uniformly at random from a disk and bin them into a grid of voxels. Each voxel's count is a Poisson-ish random variable with mean μ = points per voxel, so its relative noise is 1/√μ — independent of how big the disk is. Sweep N and the resolution below to see when the sampled disk stops looking speckled and starts looking like a disk.

Ideal

Exact density: each voxel shaded by the fraction of it inside the disk.

00.8×1.6× ideal

Sampled

Counts from random points, on the same color scale.

00.8×1.6× ideal

Noise

Error in units of the expected standard deviation, (count − ideal) / √ideal.

−3σ0+3σ
Points per voxel μ
Measured noise
Predicted 1/√μ
Interior voxels
Count range (interior)
Voxel width
Points for noise
Sample time

Why this matters for MRI simulation. Discretizing a continuous object into isochromats is exactly this experiment: the signal in a voxel is a sum over the isochromats that landed in it, so random placement injects noise of order 1/√μ on top of the physics. Statistics are computed only over voxels lying entirely inside the disk — boundary voxels have a smaller expected count and would otherwise inflate the measured spread. Note that the noise panel looks like white noise at every setting: what improves with more points is not its character but its size relative to the signal.