Iso-Density Tail Geometry

Statement¶

Iso-density tail geometry studies the level sets of the iid joint density

near a large-deviation constraint such as $x_1+x_2=s$. The diagnostic asks whether the most likely configurations along the constraint sit near the diagonal, where both coordinates are moderately large, or near the axes, where one coordinate is large and the other is ordinary.

Here $X_1,X_2$ are independent copies of a continuous random variable with density $f$, and $x_1,x_2$ are its coordinate values. The sum level $s$ is fixed for each comparison. Shared notation is planned.

Normal versus Cauchy geometry¶

For two iid standard normal variables, the joint density is proportional to

On the line $x_1+x_2=s$,

so the joint density is maximized at the equal split $(s/2,s/2)$. A large Gaussian sum is geometrically a many-moderate-deviation event.

For two iid standard Cauchy variables,

Compare the equal split $(s/2,s/2)$ with an axial split $(0,s)$. The axial to equal density ratio is

Along a large-sum line, the Cauchy geometry increasingly favors one large coordinate rather than two equal coordinates. This is the density-contour version of the one-big-jump intuition in Subexponentiality and the Survival Tail Ratio.

Static density comparison¶

For the same sum level $s$, the normal and Cauchy axial-to-equal ratios are

The table evaluates exact density ratios, without estimating probabilities. Normal equal splits dominate axial splits more strongly as $s$ increases; Cauchy axial splits eventually dominate equal splits.

The axial points are comparison points, not the exact Cauchy maximizers for finite positive $s$. To locate those maximizers, minimize the denominator $D(x)=(1+x^2)(1+(s-x)^2)$ along the line. Its derivative factors as

For $s>2$, the two minima occur at

so the maximizing pairs approach $(1/s,s-1/s)$ and its swap as $s$ grows. They sit near the axes, while the equal split becomes a local density minimum. For $0\le s\le2$, the equal split maximizes the constrained density.

This algebra describes density along a constraint of probability zero. It does not prove a tail probability asymptotic. The Cauchy example is two-sided; applying the nonnegative-summand subexponential theorem to it requires a separate right-tail argument.

Caveats¶

• Iso-density geometry requires a density. Discrete, singular, censored, or heavily rounded data need a different diagnostic.

• Density at a point is not probability mass. The contour picture explains local geometry; probability statements require integration over regions.

• The iid product-density formula fails under dependence. Volatility clustering, common factors, and contagion can rotate or bend the contours.

• The examples here are symmetric and centered. Skewed losses or one-sided positive variables should first be put into a problem-specific coordinate system.

References¶

• Taleb, Statistical Consequences of Fat Tails 2020.

• Embrechts, Klueppelberg, and Mikosch, Modelling Extremal Events 1997.

Backlinks¶

• Depends on: Survival Tail Ratio and Subexponentiality.

• Related diagnostic: Max-to-Sum Ratio.

• Related density geometry: Body, Shoulders, and Tails.

Source and adaptation¶

Adapted from incerto-wiki, content/concepts/methods/iso-density-tail-geometry.md, revision 9717c9c (2026-09-13 import). Copyright (c) 2023 xshi19. Licensed under MIT. Links and notation were adapted for this site; executable figures and simulations were replaced with static calculations. No upstream execution or formal-proof verification is claimed for this adaptation.