The Gaussian soap bubble

My favorite stats fun fact is that high-dimensional isotropic Gaussian distributions have almost all of their probability mass concentrated in a thin spherical shell centered on the mean, like a soap bubble, even though their density is bell-shaped and maximized at the center of that shell.

To see it, draw a point from a standard Gaussian in $d$ dimensions:

$$X = (X_1,\ldots,X_d), \qquad X_i \sim N(0,1) \text{ independently.}$$

This is the standard multivariate Gaussian, usually written $X \sim N(0,I_d)$: independent standard-normal coordinates give a jointly normal vector with mean zero and identity covariance.

The point is not usually close to the origin: its typical distance is about $\sqrt d$. Although the pointwise Gaussian density is highest at the origin and decreases outward, the radial density of $R=\lVert X\rVert$ behaves differently. For $d>1$ it initially rises, because thin shells farther from the origin contain more volume. In high dimensions it peaks near $\sqrt d$, and most of the mass lies in a shell that is thin relative to that radius.

The proof is a trivial application of the central limit theorem. Let

$$R = \lVert X \rVert, \qquad \text{so} \qquad R^2 = \sum_{i=1}^d X_i^2.$$

The $X_i^2$ are iid, so the central limit theorem says $R^2$ is approximately normal, centered around $d$, with fluctuations of order $\sqrt d$. Equivalently, $R^2/d$ fluctuates around $1$ on the scale $1/\sqrt d$. In short,

$$R^2 = d + O_p(\sqrt d),$$

and therefore

$$R = \sqrt d + O_p(1).$$

Thus $R$ has constant-order fluctuations around $\sqrt d$. A central band containing, say, 90% of the radii has constant-order absolute width. But $R/\sqrt d=1+O_p(1/\sqrt d)$, so that band becomes arbitrarily thin relative to its radius. This relative thinning is the soap bubble phenomenon.

The simulation below draws Gaussian points and plots the distribution of $R=\lVert X\rVert$. Move the dimension slider and watch the distribution march outward while its relative width shrinks.

Each sample is a draw from N(0, I_d). As d grows, the radius moves outward to sqrt(d) while R / sqrt(d) concentrates near 1.

50 dimensions

dimensions

50

2d10d50d200d1000d

simulated radius large-d normal approximation sqrt(d)

sqrt(d)

7.07

mean radius

7.02

middle 90%

5.89 to 8.20

relative width

32.8%

This is one of the reasons high-dimensional probability can feel geometrically strange. In one or two dimensions, “near the origin” feels like a plausible place for Gaussian mass to be. In many dimensions, the origin is still the point of highest density, but any small neighborhood around it contains almost no probability mass.

The tension is between pointwise density and available volume. At radius $r$, the Gaussian density supplies a factor proportional to $\exp(-r^2/2)$, while the surface area of a thin spherical shell supplies a factor proportional to $r^{d-1}$. The radial density is therefore exactly proportional to $r^{d-1}\exp(-r^2/2)$. For $d>1$ it starts at zero, rises, and peaks at $\sqrt{d-1}$, where volume growth and density decay balance exactly; after that, density decay dominates and the radial density falls. Since $\sqrt{d-1}\sim\sqrt d$, the balance occurs at the same radius predicted by the CLT. This is a concrete example of concentration of measure: in high dimensions, probability mass can be confined to a region that is narrow relative to its scale even when the underlying density looks smooth and bell-shaped.

This is not only a really neat fact, it is also relevant in practice. Pairwise distances between independent Gaussian points concentrate around $\sqrt{2d}$, so distance-based methods see less relative contrast as dimension grows. Likelihood-based anomaly detectors can mistake a high-density but atypical point, such as the mean, for an ordinary observation. And in Bayesian computation, mode-based summaries can miss the typical set where most posterior mass and representative draws are found.