The Cauchy Distribution

Averaging more data should make an estimate better. That's the Law of Large Numbers, and it's one of the most trusted intuitions in statistics. The Cauchy distribution is the standard example that breaks it completely: average a thousand Cauchy draws and the sample average is no more reliable than a single draw — it doesn't converge to anything. The distribution looks perfectly bell-shaped and innocent at a glance; the danger is entirely in the tails.

The analogy: a spotlight on an infinite wall

Stand a spotlight one meter from an infinitely long wall, aim it at a random angle (uniformly between $-90°$ and $90°$), and mark where the beam hits the wall. That mark's position, over many random aim angles, follows a Cauchy distribution. Most beams land near the point directly opposite the spotlight — but every so often, the angle is close enough to grazing the wall that the beam lands astronomically far away. Those rare, extremely far hits happen often enough, and are extreme enough, that no amount of averaging tames them: one grazing shot can outweigh a thousand centered ones.

Writing it down

The standard Cauchy distribution has density

$$f(x) = \frac{1}{\pi(1+x^2)}.$$

In words: the density falls off like $1/x^2$ for large $|x|$ — much slower than the normal distribution's $e^{-x^2/2}$ decay. That single difference — polynomial tails instead of exponential tails — is the entire story. It's symmetric and bell-shaped, with a well-defined median (0) and mode (0), but

$$E[X] = \int_{-\infty}^{\infty} \frac{x}{\pi(1+x^2)}\,dx \quad \text{does not converge.}$$

In words: the integral defining the mean doesn't settle to a finite number — the positive and negative tails both contribute infinite area, and they don't cancel in any meaningful sense. Variance, having no mean to be measured around, is likewise undefined (formally infinite). This breaks the ordinary CLT and LLN outright: for i.i.d. Cauchy draws, the sample average $\bar{X}_n$ has exactly the same Cauchy distribution for every $n$ — averaging a million draws is no better than looking at one.

This normal curve looks similar to a Cauchy near the center — same rough bell shape — but the resemblance is deceptive: stretch the tails of this picture out and imagine them decaying far more slowly, never approaching zero fast enough for the tails to stop mattering. That's the entire visual difference, and it changes everything about the math.

Worked example 1: the sample mean refuses to converge

Simulate $n=10$ standard Cauchy draws, e.g.: $0.3, -1.2, 0.8, 15.7, -0.5, 0.2, -2.1, 0.6, -0.9, 1.1$. Sample mean: $1.4$ — already dragged noticeably by the single outlier 15.7. Draw 10 more; suppose one lands at $-340$ this time (entirely plausible for Cauchy — grazing shots happen). The running average lurches again, wildly. Contrast with a normal(0,1) sample: after $n=1000$, the sample mean is reliably within about $\pm0.06$ of 0 (by the usual $\sigma/\sqrt{n}$ shrinkage). For Cauchy, the sample mean at $n=1000$ has the exact same spread as a single draw — no shrinkage at all, ever.

Worked example 2: as a ratio of normals

If $Z_1, Z_2$ are independent standard normals, $Z_1/Z_2$ is exactly standard Cauchy. Concretely, suppose two independent price shocks $Z_1=0.02, Z_2=0.001$ (a tiny denominator) — their ratio is 20, a huge, unremarkable-looking Cauchy draw. This is precisely how Cauchy-like tails sneak into finance: any ratio of two roughly-normal quantities (a spread divided by a small, noisy liquidity measure; a P&L divided by a near-zero volume) inherits Cauchy-style fat tails, even if neither input alone looks dangerous.

Set a negative-leaning exponent here to see how slowly a power-law tail decays compared to the normal curve above — that slow, "always still there" decay far from zero is the visual signature of the $1/x^2$ Cauchy tail, in contrast to a normal curve's tail which is essentially invisible past a few standard deviations.

What this means in practice

• Ratios are dangerous. Sharpe ratios, hedge ratios, and any statistic dividing by a quantity that can be near zero can inherit Cauchy-like behavior — huge, destabilizing outliers with no averaging protection.
• Robust statistics. When a return series shows Cauchy-like tail behavior, use the median and interquartile range instead of the mean and standard deviation — both remain well-defined and stable even though the mean and variance formally do not exist.
• A cautionary limit case. The Cauchy distribution is the $\alpha=1$ member of the Lévy stable family — a useful mental anchor for how much worse "fat-tailed" can get than a normal or even a Pareto distribution with a finite mean. See Lévy Stable Distributions.