Does the Central Limit Theorem Survive Skew and Heavy Tails?

Open with the classic CLT statement as taught, then immediately pose the tension: textbooks use dice and coin flips, but real data — income, claim sizes, wait times — is skewed and heavy-tailed. Frame the driving question visually.

• The classical CLT: averages of independent draws approach a normal distribution
• Real-world distributions are often skewed or have extreme outliers
• Does the bell curve still emerge, or does the theorem quietly fail?

Ask the learner to commit to an answer before seeing the evidence: when the source distribution is highly skewed or heavy-tailed, what happens to the distribution of the sample mean as n grows?

• Forces a concrete hypothesis: still normal, stays skewed, or something in between
• Tests the common belief that the CLT requires 'nice' data

Simulation widget: draw repeated samples of size n from a uniform, an exponential (right-skewed), and a Pareto distribution (heavy-tailed), and watch the histogram of sample means evolve as n increases. Lets the learner see normality emerging — or not — for each source side by side.

• Increase n via a slider and resample repeatedly
• Compare the three source distributions' sampling distributions of the mean at the same n
• Observe that skew persists longer than uniform, and heavy tails persist longer still

Show Q-Q plots of the standardized sample means against a true normal, at matched n, for the uniform, exponential, and Pareto sources. The uniform matches almost perfectly, the exponential shows systematic deviation in the tails, and the Pareto shows strong tail deviation even at large n.

• Q-Q plot deviations are concentrated in the tails for skewed sources
• Heavy-tailed sources show the largest and most persistent tail deviations
• Center looks roughly normal long before the tails do

Explain the mechanism: the CLT requires only finite mean and finite variance, not symmetry. The standardized mean's distribution is normal in the limit because characteristic functions factor and cancel — but how fast the characteristic function of the standardized sum approaches exp(-t^2/2) depends on the third and fourth moments, which are dominated by the tail.

• CLT assumptions: i.i.d. samples, finite mean, finite variance — nothing about shape
• Skewness controls the third-moment contribution; kurtosis controls the fourth
• Heavy tails inflate higher moments, so convergence is slow

Simulation widget that plots a tail-deviation metric (e.g., max absolute Q-Q deviation) against sample size n for several Pareto tail parameters. Demonstrates that as the tail gets heavier (alpha approaches 2), the n needed to reach a given accuracy grows sharply.

• Heavier tail = more samples needed for the same approximation quality
• The 'n≈30 rule' is a uniform-distribution artifact
• For tail parameter alpha near 2, you may need thousands of samples

Show the boundary: if the source distribution has infinite variance (e.g., Pareto with alpha ≤ 2, Cauchy), the CLT for the mean does not apply. The sample mean does not converge to a normal — it converges to a stable distribution, or does not converge at all. Distinguish this from the slow-convergence case.

• Finite variance is a non-negotiable CLT assumption
• Infinite-variance stable laws absorb extreme observations too slowly
• Heavy but finite-variance tails: slow CLT. Truly infinite tails: no CLT for the mean.

Present a transfer scenario: an insurance portfolio of claim sizes is right-skewed with a heavy tail. The learner must choose a sample size n, simulate repeated sampling, and decide whether a normal-based 95% confidence interval is trustworthy. The widget reports empirical coverage vs. nominal 95%.

• Apply the convergence lesson to a realistic decision
• Small n gives under-coverage because the sampling distribution is still skewed
• Confidence interval reliability depends on the tail, not just on having 'enough' data

Close by directly answering the driving question. The CLT still gives an approximately normal sampling distribution for the mean of skewed or heavy-tailed data, provided variance is finite. The price is convergence rate: skew and especially heavy tails force much larger n than the textbook picture suggests, and infinite-variance sources are outside the theorem's reach entirely.

• Yes, the CLT holds under skew and heavy tails — with finite variance
• Convergence rate is governed by the tail, not the shape of the bulk
• Practical n must be chosen with the tail in mind, and infinite-variance cases need a different tool