Chapter 1 — Prologue

Reading the Prologue¶

The Prologue asks what changes in statistical reasoning when extreme outcomes matter. Its terminology is deliberately broader than a single asymptotic tail class. We use that question to separate three properties: a large fourth moment relative to variance, the existence of exponential moments, and the scaling of a survival function.

The source is Taleb (2025), Statistical Consequences of Fat Tails, third edition, arXiv:2001.10488v4, chapter 1, printed pp. 3–6. This discussion chapter states no numbered mathematical theorem. The calculations below and in the tail class catalog are our own teaching examples; the Prologue is their motivation, not their proof source.

The source uses “thick” in a Gaussian-relative, kurtosis-based sense, and “fat” more broadly than a strict power-law definition (ch. 1, pp. 4–5). Here we name the mathematical property explicitly. The Cramér condition asks for a finite positive exponential moment; regular variation asks for a limiting multiplier ratio; subexponentiality asks about the tail of an iid sum. None is defined by the word “fat” alone.

High kurtosis with bounded support¶

For a variable with finite fourth moment and positive variance, define its Pearson kurtosis, locally denoted by $K$, as

This is kurtosis, not excess kurtosis $K-3$. A standard Gaussian has $K=3$: integration by parts with $\phi'(z)=-z\phi(z)$ gives $\int z^4\phi(z)\,dz=3\int z^2\phi(z)\,dz=3$, where both integrals run over $\mathbb R$ and $\phi$ is the standard normal density.

We prove that arbitrarily large kurtosis is compatible with bounded support and finite exponential moments. Fix a local probability $0<\varepsilon\le1$ and give $X$ the three-point law

Symmetry gives $\mathbb E X=0$. Summing over the three atoms gives

For every real $\theta$, the same finite sum gives

For each fixed $\varepsilon$ the right tail ends at $\varepsilon^{-1/2}$. It therefore cannot have a regularly varying tail with positive exponent, or belong to the unbounded-support subexponential class used here. Choosing smaller $\varepsilon$ produces a different bounded law with larger kurtosis. It does not turn any one of these laws into a power-tailed distribution. For example, $\varepsilon=0.01$ gives variance 1, kurtosis 100, and support $\{-10,0,10\}$.

Ordinary moments and tail classes¶

The catalog supplies a second contrast: a lognormal variable has all positive power moments finite but has no finite positive exponential moment. It is subexponential, yet not regularly varying. An exact Pareto variable has a positive power moment of order $q$ precisely when $q<\alpha$. The catalog proves the lognormal moment calculations and separates the cited classifications; the Pareto moment note owns the exact Pareto integral proof.

These examples answer different questions. A finite population variance requires a finite second moment. The kurtosis comparison requires a finite fourth moment as well. A finite Chernoff bound requires an exponential moment at the chosen positive parameter. A multiplier-ratio claim requires information arbitrarily far into the tail. A finite empirical fourth moment alone establishes none of those population assumptions.

Reproduce the model comparisons¶

From the installed repository environment, run:

Use bounded_spike_summary.

from xmath.incerto.tail_classes import bounded_spike_summary

for epsilon in (0.1, 0.01, 0.001):
    print(f"epsilon={epsilon:.4g}")
    for name, value in bounded_spike_summary(epsilon).items():
        print(f"  {name}: {value:.4g}")

epsilon=0.1
  support_radius: 3.162
  mean: 0
  variance: 1
  kurtosis: 10
epsilon=0.01
  support_radius: 10
  mean: 0
  variance: 1
  kurtosis: 100
epsilon=0.001
  support_radius: 31.62
  mean: 0
  variance: 1
  kurtosis: 1000

Every law here has bounded support, so large kurtosis alone does not establish a heavy asymptotic tail.

Figure regeneration: uv run python demos/incerto/prologue_tail_classes.py.

The script evaluates specified probability laws in dimensionless units. There are no observed data, random draws, fitted parameters, or confidence intervals. It prints the three-point calculation for $\varepsilon\in\{0.1,0.01,0.001\}$ and the multiplier ratios for the catalog’s five unbounded models. A fixed 241-point grid spans $1\le u\le10^4$, with multiplier $t=2$; runtime is bounded by this grid and two static figure exports, normally a few seconds.

The catalog figure shows why a vanishing multiplier ratio cannot by itself distinguish lognormal and Weibull subexponential tails from exponential and gamma tails. The code uses log survival functions to avoid dividing two underflowed probabilities; it stops on non-finite values rather than dropping them. Finite curves illustrate the models and do not prove their limiting classes.

References and next reading¶

• Taleb, Statistical Consequences of Fat Tails, third edition (2025), arXiv:2001.10488v4, ch. 1, pp. 3–6; terminology on pp. 4–5. Source context only.

• Tail class catalog: proofs, exact classification locators, and the model-ratio comparison.

• SCoFT reading map: available chapter guides.

Symbols follow the shared notation. The Prologue supplies no equation-level notation to translate; $K$ and $\varepsilon$ above are local to our counterexample. The book’s text and figures are not reproduced here.