Chapter 2 — Glossary, definitions, and notations

Reading a glossary through examples¶

A probability, a conditional mean, and a contribution to an expectation can all refer to observations above the same threshold. They answer different questions. This guide uses that distinction to connect the glossary to counting exceedances, the Pareto law, the extreme-value index, and the tail class catalog.

The source is Taleb (2025), Statistical Consequences of Fat Tails, third edition, arXiv:2001.10488v4, chapter 2, printed pp. 7–21. The tables give reading locations, not transcriptions of the glossary. Our examples and arguments are developed independently. The shared notation controls the symbols used here; source-specific conventions need translation.

Notation at the point of use¶

For a lognormal, $\mu$ is a log-location, not its ordinary mean. The catalog derivation gives $\mathbb E X=\exp(\mu+\sigma^2/2)$, so a prescribed mean $m>0$ requires $\mu=\log m-\sigma^2/2$. Likewise, the stable-law index is restricted to $(0,2]$, whereas a Pareto exponent can exceed 2; neither is silently interchangeable with the EVI.

Map of the glossary concepts¶

The core route is probability → tail size → tail class → normalized extremes. Other glossary entries point ahead to later chapters. A link below identifies a relevant existing note, not a claim that all of that topic is proved here.

The normalized expression in the first table fails the triangle inequality when $0<p<1$: for $(1,0)$ and $(0,1)$ it gives 1 for their sum but $2^{1-1/p}<1$ for the sum of their values. Independence can always be expressed by $P(X\in A,Y\in B)=P(X\in A)P(Y\in B)$ for measurable sets. A density ratio can express it only where the denominator is positive and the densities exist. An elliptical scatter matrix is a model parameter; identifying it with a covariance additionally requires second moments and the appropriate scalar normalization.

Which limit is being taken?¶

For random variables $Y_n,Y$, convergence in probability means $P(|Y_n-Y|>\varepsilon)\to0$ for every $\varepsilon>0$. Almost-sure convergence means $P(\lim_nY_n=Y)=1$. Convergence in distribution means convergence of the CDFs at continuity points of the limiting CDF. These are the conventions behind the arrows in §2.1, p. 8.

For iid $X_i$ with $\mathbb E|X_1|<\infty$, the weak LLN gives $\bar X_n\xrightarrow{P}\mu$, where $\mu=\mathbb E X_1$. With the stronger assumption $0<\sigma^2=\operatorname{Var}(X_1)<\infty$, the classical CLT gives

These two general theorems are cited, not proved here (SCoFT v4 §§2.2.2–3, p. 10). Neither statement gives a universal finite-sample error bound. The exceedance note does prove a pointwise probability bound for its bounded indicators.

By contrast, the subexponential relation compares $P(S_n>x)$ and $P(M_n>x)$ as $x\to\infty$ with $n$ fixed. It does not say that $M_n/S_n$ tends to one as $n$ grows. The Pareto sums section keeps the growing-sample regimes and moment orders separate.

A probability does not specify the tail payoff¶

For Pareto $(\alpha=3/2,x_m=1)$, threshold $u=4$ is the quantile $Q(7/8)$. The exceedance probability is $1/8$, but the conditional mean above that threshold is 12. The Pareto derivation gives mean excess 8 and tail contribution $3/2$. All four numbers describe the same distribution and threshold.

For an integrable nonnegative loss, define expected shortfall at level $0<q<1$ by $(1-q)^{-1}\int_q^1Q(v)\,dv$. When there is no atom at the quantile, this agrees with the conditional tail mean. At an atom, conditioning on $X>Q(q)$ generally selects a different probability mass. For the empirical law of $(1,2,2,4)$ and $q=1/2$, the upper half consists of equal masses at 2 and 4, so the quantile integral gives 3. Strict conditioning on $X>2$ instead gives 4. This calculation explains why the probability level, sign convention, and treatment of ties must accompany the terms VaR and expected shortfall.

Even uniform CDF accuracy does not control a mean without additional tail conditions. To prove this, let $P_j$ put mass $1-1/j$ at zero and $1/j$ at $j^2$, for integers $j\ge2$. Its CDF differs from that of a point mass at zero by at most $1/j$, yet its mean is $j$. This deterministic sequence of laws is a counterexample to such an implication; it is not a simulation of an iid empirical process. It isolates the probability-versus-moment issue behind §§2.2.27–28, pp. 19–20, without importing a uniform convergence theorem.

Reproduce the worked cases¶

Use empirical_survival_function, pareto_tail_quantities.

from xmath.incerto.stats import empirical_survival_function
from xmath.incerto.glossary import pareto_tail_quantities

support, survival = empirical_survival_function([1, 2, 2, 4])
print("Support: [" + ", ".join(f"{v:.4g}" for v in support) + "]")
print("Strict survival: [" + ", ".join(f"{v:.4g}" for v in survival) + "]")
for alpha in (0.8, 1, 1.5, 3):
    result = pareto_tail_quantities(alpha, x_m=1, threshold=4)
    print(f"alpha={alpha:.4g}")
    for name, value in result.items():
        print(f"  {name}: {value:.4g}")

Support: [1, 2, 4]
Strict survival: [0.75, 0.25, 0]
alpha=0.8
  log_survival: -1.109
  mean_excess: inf
  conditional_mean: inf
  tail_first_moment: inf
alpha=1
  log_survival: -1.386
  mean_excess: inf
  conditional_mean: inf
  tail_first_moment: inf
alpha=1.5
  log_survival: -2.079
  mean_excess: 8
  conditional_mean: 12
  tail_first_moment: 1.5
alpha=3
  log_survival: -4.159
  mean_excess: 2
  conditional_mean: 6
  tail_first_moment: 0.09375

Figure regeneration: uv run python demos/incerto/exceedances.py; uv run python demos/incerto/glossary_worked_cases.py.

The second script uses the fixed list $(1,2,2,4)$ and exact population formulas. It prints conditional Pareto quantities at $u=1,4,16$ for $\alpha=0.8,1,1.5,3$, the uniform-CDF counterexample for $j=10,100,1000$, and multiplier ratios for the catalog’s logarithmic correction. It evaluates four Pareto maximum CDFs with $\alpha=1.5$ on 301 points from 0.25 to 4, and a 161-point multiplier grid from 1 to 10⁴. Units are dimensionless. There is no RNG, external data, model fitting, or distribution truncation; the finite grids bound the work to a few seconds and two figure exports.

Unexpected non-finite arithmetic raises an error. CDF values may round to zero or one at floating-point limits; this does not assert finite support. Tests compare finite formulas with independent integration and distribution calculations. The probability bounds, conditional identities, maximum limit, and counterexamples are proved in the notes, not by the tests.

References and next reading¶

• Taleb, Statistical Consequences of Fat Tails, third edition (2025), arXiv:2001.10488v4, ch. 2, §§2.1–2.2.32, pp. 7–21. Section/page locators above identify source context.

• Tail class catalog gives the precise classification sources and separates proved formulas from cited theorems.

• Chapter 3 — Non-technical overview connects the definitions to statistical consequences; its source references retain the edition identified on that page.