Mean Testing Under Heavy Tails

Why Heavy Tails Change Mean-Test Calibration

High-dimensional mean tests often turn coordinatewise sample means into a maximum or a quadratic statistic. Their familiar Gaussian or extreme-value calibrations require more than a mean of zero and a finite covariance matrix: as the number of coordinates grows, a few unusually large observations can dominate coordinate sums or make quadratic fluctuations much larger than the calibration anticipates. The issue is not that every Gaussian approximation fails under every heavy-tailed model. It is that finite variance alone does not provide a general justification for those approximations.

We work with independent and identically distributed vectors \(X_1,\ldots,X_n\in\mathbb{R}^p\), where \(X_i=(X_{i1},\ldots,X_{ip})^\top\), and test \(H_0:E[X_i]=0\). Coordinates within one vector may be dependent unless an example says otherwise. Write \(\bar X_j=n^{-1}\sum_{i=1}^n X_{ij}\). The maximum statistic examines \(\max_j|\sqrt n\,\bar X_j|\); a basic quadratic statistic is \(Q=n\|\bar X\|_2^2=\sum_{j=1}^p(\sqrt n\,\bar X_j)^2\). Scale estimates and centering refinements used in particular tests do not remove the need to justify their reference distributions under the assumed tails.

A Pareto Example: The Maximum Can Be Too Large

The following example shows explicitly why a Gaussian extreme-value threshold need not control the null rejection probability. Let \(p=n\). Take all \(n^2\) entries \(X_{ij}\) to be independent, symmetric random variables with \(P(|X_{ij}|>t)=t^{-3}\) for \(t\ge1\). Symmetry gives mean zero. The variance is finite and equals 3, since \(E[X_{ij}^2]=1+\int_1^\infty 2t\,t^{-3}\,dt=3\).

Quadratic statistics are vulnerable for a related reason: they square coordinate means, so an unusually large coordinate can have disproportionate influence. In the example above, each finite-sample coordinate sum has infinite fourth moment. Indeed, condition on the other \(n-1\) entries lying in a fixed bounded interval; this event has positive probability, and the remaining Pareto entry still has infinite fourth moment after addition of a bounded quantity. Hence \(E[(\sqrt n\,\bar X_j)^4]=\infty\), and \(E[Q^2]=\infty\). A Gaussian quadratic calibration justified by finite fourth moments is therefore not justified here. More generally, the appropriate conditions depend on the precise quadratic statistic and its growth regime; finite coordinate variances alone do not establish its usual Gaussian limit.

Truncation: Control the Fluctuation and the Bias

Truncation gives a direct route to valid calibration. Suppose \(E[X_{ij}]=0\) and, for some \(0<\delta\le1\), the uniform moment bound \(E[|X_{ij}|^{2+\delta}]\le M\) holds for every coordinate. The moment condition controls the truncation bias, while the bounded summands permit Bernstein’s inequality. No independence among coordinates is needed for the simultaneous bound below; independence across observations is needed.

To prove the bias bound, use \(E[X]=0\) to write \(E[\psi_\tau(X)]=E[\psi_\tau(X)-X]\). The difference is zero when \(|X|\le\tau\), and its magnitude is at most \(|X|\) otherwise. On \(\{|X|>\tau\}\), \(|X|\le |X|^{2+\delta}/\tau^{1+\delta}\), so

For the fluctuation bound, set \(W_{ij}=\psi_\tau(X_{ij})\), \(m_j=E[W_{ij}]\), and \(v_j=\operatorname{Var}(W_{ij})\). Then \(|W_{ij}-m_j|\le2\tau\). If \(v_j\le v\) for every coordinate, Bernstein’s inequality and a union bound yield the following result.

For each coordinate, apply Bernstein’s inequality to the centered sum of \(W_{ij}-m_j\). The variance of that sum is \(nv_j\le nv\), and each summand is bounded in magnitude by \(2\tau\). In the usual Bernstein tail bound, the threshold \(\sqrt{2nv t}+(2/3)(2\tau)t\) gives a one-sided tail at most \(e^{-t}\). Applying it to both signs gives

Since \(2pe^{-t}=\alpha\), a union bound controls the largest centered coordinate at that threshold with probability at least \(1-\alpha\). The bias lemma gives \(\sqrt n|m_j|\le\sqrt n M/\tau^{1+\delta}\). Adding this maximum bias to the centered threshold proves the theorem. \(\square\)

Worked Calibration Examples

A quadratic statistic can also be calibrated from the same coordinate bounds, although this construction can be conservative. Under the theorem’s assumptions, reject when \(Q_W=n\|\bar W\|_2^2>pR^2\). If \(Q_W>pR^2\), then at least one coordinate satisfies \(|\sqrt n\,\bar W_j|>R\); otherwise summing the \(p\) squared coordinate bounds would give \(Q_W\le pR^2\). The theorem therefore guarantees null rejection probability at most \(\alpha\). This valid quadratic rule is not a claim that its threshold is optimal or that it has good power against every dense alternative.

Choosing the Truncation Level and Avoiding Pitfalls

The truncation level controls a real trade-off. A larger \(\tau\) preserves more of the original observations but enlarges the Bernstein fluctuation term \(4\tau t/(3\sqrt n)\). A smaller \(\tau\) reduces that term but increases the bias bound \(\sqrt n M/\tau^{1+\delta}\). For a growing \(p\), a useful asymptotic choice should make both contributions appropriately small relative to the desired threshold. For example, in the Pareto example, \(\tau=n^{2/5}\) makes the bias and bounded-summand contributions vanish for \(p=n\), while the principal threshold grows on the order of \(\sqrt{\log n}\).

The quadratic rule just derived is intentionally transparent: a simultaneous coordinate guarantee yields a valid threshold, but can discard the benefit of combining evidence across coordinates. More powerful robust quadratic procedures require additional structure or sharper distributional analysis. Their calibration must match their assumptions; truncation alone does not automatically produce a particular Gaussian or chi-squared limit.

• Assuming finite variance is enough. The Pareto example has finite variance, yet its maximum coordinate mean exceeds every fixed Gaussian-scale multiple with probability tending to one.
• Ignoring truncation bias. The clipped variables generally have nonzero means even under \(E[X_{ij}]=0\). Include a bias bound or establish a symmetry condition that makes the clipped mean zero.
• Using a coordinatewise result without simultaneous calibration. A bound for one coordinate does not control the maximum over \(p\) coordinates. The union bound’s \(\log(2p/\alpha)\) term accounts for this multiplicity.
• Transferring a maximum threshold to a quadratic statistic unchanged. A bound on the largest coordinate does not by itself justify a quadratic reference distribution. State and verify the implication or distributional result used for the quadratic test.