Chapter 4 — Finite-moment fat tails

Chapter 4 asks how a distribution can depart sharply from a Gaussian model while retaining finite moments. Its first mechanism is to randomize a Gaussian’s variance. Its second question is statistical: which moments must exist for a chosen estimate to have a finite error variance?

We use SCoFT’s 2025 third edition, arXiv:2001.10488v4, ch. 4, printed pp. 69–89. The chapter’s broad use of “fat tails” includes comparisons with a Gaussian at fixed variance. That usage does not assert regular variation or subexponentiality.

A reading path through the chapter¶

V4 locator, printed pages	Mathematical question	Our treatment
§§4.1–4.1.2, pp. 69–74	Can variance stay fixed while higher moments increase?	Variance mixtures: conditioning, finite-state examples, kurtosis, and Gaussian-relative tails.
§4.2, pp. 74–75	Does mixing variance or its reciprocal give the same tail?	Variance and precision: gamma, lognormal, and Student models with proved moment conditions.
§§4.3–4.3.1, pp. 75–78	Where does a perturbation add or remove density?	Body, shoulders, and tails: local density curvature, survival curvature, and exact finite-mixture crossings.
§§4.4–4.4.3, pp. 79–84	What differs between quadratic and absolute deviations?	Moment estimation and dispersion ratios: estimands, estimator moments, Gaussian calibration, and repeated samples.
§4.4.4, pp. 84–86	Why do higher powers emphasize large observations?	Power means: Jensen proof, equality, and the maximum limit.
§4.4.5, pp. 87–88	Does a dispersion comparison justify a general recommendation?	Different estimands: state the target and loss before comparing methods.
§4.5, pp. 88–89	Why do geometric unit balls expand with norm order?	Norm geometry: distinguish sums of powers from averages of powers.

Start with a conditional model¶

Let $X=\sqrt YZ$ , with $Z\sim N(0,1)$ independent of positive variance $Y$ . The shared mixing symbol is $Y$ ; see the notation canon and Normix foundation. With $\mathbb EY=1$ and finite $\mathbb EY^2$ ,

\mathbb EX=0,\qquad \operatorname{Var}(X)=1,\qquad \mathbb EX^4=3\mathbb EY^2.

(1)

The mixture note proves these identities. Equal probabilities on variances 0.4 and 1.6 give fourth moment 4.08; probabilities 0.99 and 0.01 on variances 0.5 and 50.5 give 77.25. Both mixtures have a finite MGF everywhere. A sharper peak and larger far-tail probabilities relative to a unit normal do not make either mixture a power law.

Specify what is perturbed¶

The source’s Gaussian boundaries in §4.3.1 concern differentiation with respect to scale. Our body/shoulder note also computes variance curvature. For a unit normal the positive density boundaries are

\begin{aligned} \text{variance perturbation:}&\quad \sqrt{3\pm\sqrt6},\\ \text{scale perturbation:}&\quad \sqrt{(5\pm\sqrt{17})/2}. \end{aligned}

(2)

Neither pair is a threshold for fitting a Pareto tail. A small-perturbation expansion is pointwise, and its zero-curvature points need not equal the crossings of a finite mixture. Tail probabilities have their own curvature: convexity in variance holds only above a threshold relative to that variance, not for every tail event and every mixing interval.

Separate a moment from its estimator¶

For iid Pareto observations, the moment note proves

\mathbb E X^q<\infty\iff q<\alpha,\qquad \operatorname{Var}(\widehat m_{q,n})<\infty\iff 2q<\alpha \quad\text{when }q<\alpha.

(3)

At exponent 3 the second moment exists, its sample average is unbiased and consistent, and that average has infinite variance. Its square root has a finite second moment. This last distinction matters when reading the standard-deviation comparison around §4.4.3, p. 83: conclusions about an average of squares do not transfer unchanged to its square root.

Even with every moment finite, a rare high-variance component can be absent from most short samples. The repeated-sample demo shows both mechanisms. Neither mechanism implies that every sample mean is downward biased. The plug-in singularity adds a third issue: fitting a model may create divergent estimator moments near a parameter boundary.

Notation and source qualifications¶

Source usage	Convention used here
Normal scale parameter	$N(\mu,v)$ uses variance as its second argument; write $v=\sigma^2$ explicitly.
Stochastic variance or reciprocal variance	$Y$ is positive variance; local $Q=1/Y$ is precision.
Student tail parameter	$\nu$ is degrees of freedom; the power-tail exponent is $\alpha=\nu$ . Unit Student scale is not unit standard deviation.
Mean deviation / MAD	State the center and whether the deviation is a mean or a median; our dispersion demo applies Gaussian calibration.
Rising norms	$A_p$ averages powers; geometric $\ell_p$ norms sum them. Their monotonicities differ.

Zero mean does not identify an MGF with a characteristic function; the finite-mixture formulas show their opposite signs in the exponent. Missing a power moment also does not by itself establish a power law. The explicit mixture and Pareto calculations provide narrower, checkable statements than either shortcut.

The chapter’s efficiency discussion is conditional on a model and an estimand. A Gaussian calibration does not make mean absolute deviation an unbiased estimator of standard deviation under every other law. We do not adopt a universal preference for one dispersion statistic. The bounded Cauchy experiment is reserved for chapter 7; recursive mixing for chapter 20.

Worked cases¶

Use normal_mixture_absolute_moment, power_mean.

from xmath.incerto.finite_moments import normal_mixture_absolute_moment, power_mean

variances, weights = [0.5, 50.5], [0.99, 0.01]
for order in (2, 4):
    print(f"{order:d} {normal_mixture_absolute_moment(order, variances, weights):.4g}")
means = [power_mean([0, 0, 0, 4], p) for p in (1, 2, 4)]
print("Power means: [" + ", ".join(f"{v:.4g}" for v in means) + "]")

Output

2 1
4 77.25
Power means: [1, 2, 2.828]

The finite mixture and the fixed vector answer different moment and norm questions.

Figure regeneration: uv run python demos/incerto/finite_moments.py for the deterministic mixture, curvature and norm figures and the seeded moment experiment. Figure regeneration: uv run python demos/incerto/dispersion_ratio.py for the existing Gaussian, Student, and Pareto comparison. Model curves and simulations are identified separately in their concept-page captions; no external data are needed. The scalar variance–mean gamma derivation is in the mixture note.

References¶

Taleb (2025), Statistical Consequences of Fat Tails, third edition, 2025, arXiv:2001.10488v4, ch. 4, §§4.1–4.5, printed pp. 69–89. The table gives the section locators used. Our examples, proofs and figures are developed independently; the concept pages identify cited inputs.

References¶

Taleb, N. N. (2025). Statistical Consequences of Fat Tails (3rd ed.). https://arxiv.org/abs/2001.10488v4