Why does the Central Limit Theorem fail for Cauchy-distributed data?

## The condition the theorem actually needs The classical Central Limit Theorem assumes i.i.d. draws with a finite mean `mu` and a **finite variance** `sigma^2`. The finite-variance requirement is not decoration. The theorem's conclusion — that `sqrt(n) * (Xbar - mu) / sigma` tends to a standard normal — mentions `sigma` explicitly, so if `sigma` is infinite the statement has nothing to say. ## What makes the Cauchy pathological The standard Cauchy has density `1 / (pi * (1 + x^2))`. Its tails decay like `1 / x^2`, which is slow. Slow enough that computing the expectation requires integrating something behaving like `1 / x` out to infinity, and that integral diverges. So the Cauchy has **no finite mean** — not a large mean, not an unknown one; the defining integral simply does not converge. With no finite mean, there is no finite variance either. The Cauchy also has a memorable construction: the ratio of two independent standard normal variables is standard Cauchy. Two of the most tractable random quantities in statistics divide into one of the least tractable. ## The stability property The striking consequence is that the average of n i.i.d. standard Cauchy variables is *itself exactly standard Cauchy*, for every n. Not approximately, not asymptotically — exactly, at n = 1, n = 100 and n = 10,000,000. Averaging accomplishes nothing at all: the distribution of the average is identical to the distribution of a single observation. This is the cleanest possible refutation of the intuition that 'more data always makes an average more reliable'. That intuition rests on the variance of the average shrinking as `sigma^2 / n`, and if `sigma^2` is infinite the argument evaporates. ## The behaviour you would actually see Compute a running average of Cauchy draws and plot it against the number of observations. It does not converge. It wanders at some level for a while, then a single enormous observation arrives and yanks it to a completely new level, where it wanders again until the next one. The plot has visible jump discontinuities that keep occurring no matter how far out you go, because the tail keeps producing observations larger than everything previously seen. ## The general picture beyond the Cauchy The Cauchy is one member of a family. A power-law (Pareto-type) tail with index `alpha` — meaning the probability of exceeding x decays like `x^(-alpha)` — behaves as follows: - `alpha > 2`: finite mean and finite variance; the classical CLT applies, though convergence can be slow. - `1 < alpha <= 2`: the mean exists but the variance is infinite. The classical CLT fails. Suitably scaled sums converge to a non-normal **stable** distribution, and the correct scaling is `n^(1/alpha)` rather than `sqrt(n)`. - `alpha <= 1`: not even the mean exists. The Cauchy sits at `alpha = 1`. So 'the CLT fails' does not always mean 'nothing converges' — often something converges, just to a heavy-tailed stable law under a different scaling, and treating it as normal understates tail risk severely. ## Diagnosing it in real data You cannot observe an infinite variance directly; every finite dataset has a finite sample variance. What you look for is instability. 1. **Running average.** Plot it against sample size. Repeated level shifts driven by single observations, rather than a settling trend, are the signature. 2. **Running variance.** If the sample variance keeps climbing as you add data instead of stabilising, the population variance may not exist. 3. **Maximum-to-sum ratio.** Track the largest observation as a fraction of the running total. Under finite variance this ratio drifts toward zero; when it stays substantial no matter how much data arrives, the tail is dominating. 4. **Tail plots.** Rank the observations and plot the tail on log-log axes. A straight line indicates a power law, and its slope estimates `alpha`, which tells you which regime you are in. ## What to do instead If the variance genuinely does not exist, the mean may be the wrong summary rather than a hard-to-estimate one. Medians and other quantiles remain well defined and stable for any distribution, including the Cauchy, and typically become the honest target. Where the mean must be reported for a business reason, the right response is to be explicit that no finite-sample average is trustworthy, and to quantify uncertainty by a method that does not assume a finite variance rather than by normal-curve arithmetic. ## Why this comes up at senior level Genuinely infinite variance is rare in practice; what is common is data heavy-tailed enough that the CLT is effectively unusable at any n you will ever have. Interviewers use the Cauchy as the clean extreme case to test whether a candidate treats the finite-variance condition as a real assumption to check, or as boilerplate to recite.