Elliptical Distributions

Mean-variance portfolio optimization ranks portfolios using only their mean and covariance matrix. That's a huge simplification — it implicitly assumes the joint distribution of returns is fully described by those two things, ignoring skewness, tail dependence, anything else about the shape. When is that simplification actually valid? Precisely when returns follow an elliptical distribution — the family for which mean and covariance really are the whole story, whatever the individual assets' tail behavior looks like.

The analogy: shadows of a stretched, spun sphere

Take a perfect sphere of probability density (like independent standard normals in many dimensions), then stretch it along some axes and rotate it — you get an ellipsoid. Every distribution built this way, no matter which starting "sphere" you began with (a normal sphere, a heavier-tailed one, anything with the right kind of round symmetry), ends up with density contours that are perfect ellipses, just oriented and sized differently. The multivariate normal is the most familiar member; the multivariate t-distribution, with fatter tails, is another, built from the exact same "stretch and rotate" recipe applied to a rounder, heavier-tailed starting shape.

Writing it down

A random vector $X$ has an elliptical distribution with location $\mu$ and dispersion matrix $\Sigma$ if its density (when it exists) has the form

$$f(x) = c \, g\big((x-\mu)^\top \Sigma^{-1} (x-\mu)\big)$$

for some non-negative function $g$ (the "generator") and normalizing constant $c$. In words: the density depends on $x$ only through the single Mahalanobis-distance-like quantity $(x-\mu)^\top\Sigma^{-1}(x-\mu)$ — which is exactly why the contours of equal density are ellipses defined by $\Sigma$: any two points at the same value of that quadratic form have the same density, regardless of direction. Choosing $g(u) = e^{-u/2}$ recovers the multivariate normal exactly; other choices of $g$ give the same elliptical shape of dependence but with fatter or thinner tails, like the multivariate t. Crucially, for any elliptical distribution, linear combinations of the components (like a portfolio's return, $w^\top X$) are always well-summarized by $w^\top\mu$ and $w^\top \Sigma w$ alone — which is exactly the mathematical justification mean-variance optimization needs, and normally never states out loud.

Drag the transform here and watch a circle become an ellipse — that's literally the geometric operation defining this whole family: start with a "spherical" distribution (density depending only on distance from center, contours are circles) and apply a linear transform $\Sigma^{1/2}$; the circles become the ellipses that give elliptical distributions their name and their tractable correlation structure.

Worked example 1: two assets, elliptical contours

Two assets with $\mu = (0, 0)$, variances $\sigma_1^2=4$, $\sigma_2^2=1$, and correlation $\rho=0.6$, so $\Sigma = \begin{pmatrix}4 & 1.2\\ 1.2 & 1\end{pmatrix}$. If returns are jointly normal (elliptical with $g(u)=e^{-u/2}$), the 95% confidence region is exactly the ellipse $(x-\mu)^\top\Sigma^{-1}(x-\mu) = \chi^2_{2,0.95} \approx 5.99$. That ellipse is tilted along the direction of positive correlation and stretched more along asset 1's axis (larger variance) — a portfolio $w=(0.5,0.5)$ has variance $w^\top\Sigma w = 0.25(4)+0.25(1)+2(0.25)(1.2) = 1+0.25+0.6=1.85$, computed with nothing but $\mu$ and $\Sigma$, valid precisely because the joint distribution is elliptical.

Worked example 2: same ellipse, fatter tails

Keep the identical $\mu$ and $\Sigma$, but switch the generator to a multivariate t with 4 degrees of freedom (still elliptical — same-shaped ellipses, same $w^\top\Sigma w$ formula for portfolio variance) instead of normal. The 99% quantile of the portfolio $w^\top X$ under the normal model is $2.326 \times \sqrt{1.85} \approx 3.16$. Under the t(4) model, the corresponding quantile multiplier is about $3.75$ instead of $2.326$ (t-distribution tail multipliers are notably larger at low degrees of freedom), giving $3.75 \times \sqrt{1.85} \approx 5.10$ — a 99% loss estimate over 60% larger than the normal model's, using the exact same mean and covariance inputs. This is the payoff of the elliptical framework: you can swap in fatter tails without touching the correlation structure at all.

Each individual asset in an elliptical model still has its own one-dimensional marginal distribution, like this one — the elliptical property is about how those marginals are glued together by $\Sigma$, not about the shape of any single asset's returns on its own.

What this means in practice

• When mean-variance optimization is defensible. It's exactly justified when returns are (at least approximately) elliptically distributed — correlation and covariance genuinely capture the whole dependence structure. It is not automatically justified otherwise.
• Fat tails without abandoning covariance. Switching from a multivariate normal to a multivariate t keeps every covariance-based tool (portfolio variance, correlation, minimum-variance weights) valid while fixing the "tails are too thin" problem normal-based risk models have.
• What elliptical distributions can't fix. They assume all assets share the same tail-thickness parameter and a fundamentally symmetric dependence structure — real markets often show asymmetric tail dependence (correlations spike specifically in crashes), which needs tools like copulas. See Copulas.