Skip to contents

Documentation of the asymptotic robustness property of the Hotelling T² statistic under non-normal distributions, based on Gneri and Barbosa (2006), and its relationship to the IQCC package implementation.

Theorem 3 (Gneri & Barbosa, 2006, Sección 4.2)

Let \(X\) be a \(p\)-dimensional random vector with continuous distribution, mean vector \(\mu\), positive-definite covariance matrix \(\Sigma\), and finite fourth moments. Let \(X_1, X_2, \dots, X_n\) be an i.i.d. sample, and denote by \(\bar{X}(n)\) and \(S(n)\) the sample mean vector and sample covariance matrix. The Hotelling T² statistic is defined as:

$$T^2 = n (\bar{X}(n) - \mu)' S(n)^{-1} (\bar{X}(n) - \mu).$$

Then, as \(n \to \infty\), \(T^2\) converges in distribution to a chi-squared random variable with \(p\) degrees of freedom.

Proof sketch

The proof proceeds in three steps:

  1. The sample covariance converges in probability: \(S(n)_{ij} \to \Sigma_{ij}\) elementwise (Khinchin's LLN), and by continuity of matrix inversion, \(S(n)^{-1} \to \Sigma^{-1}\) in probability.

  2. By the Central Limit Theorem, \(\sqrt{n}(\bar{X}(n) - \mu) \to N_p(0, \Sigma)\) in distribution.

  3. By Proposition 2 (a Slutsky-type result), \(T^2 \to W' \Sigma^{-1} W\) where \(W \sim N_p(0, \Sigma)\), i.e., \(T^2 \to \chi^2_p\).

Discrepancy in moment conditions

The abstract and introduction of Gneri & Barbosa (2006) state the theorem under the assumption of finite second moments only ("momentos de orden 2 finitos"). However, the formal statement of Theorem 3 (Sección 4.2) requires finite fourth moments. The proof uses:

  • Khinchin's Law of Large Numbers for the sample covariance \(S(n)\), which requires \(E[|X_i X_j|] < \infty\). By Cauchy-Schwarz, this holds when second moments are finite, so order 2 suffices for this step.

  • The Central Limit Theorem for \(\sqrt{n}(\bar{X} - \mu)\), which requires finite second moments.

The fourth-moment condition in the theorem statement is therefore conservative. The theorem holds under finite second moments, provided the covariance matrix is finite and positive-definite.

Scope and limitations

The theorem is asymptotic and does not imply:

  • The exact finite-sample distribution of \(T^2\) under non-normality.

  • The joint distribution of a sequence of charted points in a control chart.

  • Valid Average Run Length (ARL), false-alarm risk, or nominal coverage for small samples.

  • Resolution of parameter estimation uncertainty in Phase I.

  • Justification of the finite-sample F or beta control limits derived under normality for arbitrary continuous distributions.

  • Coverage of discrete distributions, singular covariance, increasing dimension with \(n\), temporal dependence, or infinite moments.

Relationship to IQCC functions

Theorem 3 concerns a single-sample T² statistic: $$T^2 = n (\bar{X} - \mu)' S^{-1} (\bar{X} - \mu)$$ where \(\bar{X}\) and \(S\) are the mean and covariance of the same sample, and \(\mu\) is the known null mean.

The IQCC package implements two related but distinct statistics:

ConceptIQCC functionNotes
Theorem 3 statisticsim_t2_asymptoticSingle sample, known \(\mu = 0\), own \(S\)
Phase I T²T2.1Uses grand mean \(\bar{\bar{x}}\) and pooled \(\bar{S}\) from stats
Phase II T²T2.2Tests new obs. against Phase I estimates
Phase I chartcchart.T2.1Beta (n=1) or F (n>1) limits
Phase II chartcchart.T2.2F limits
Auxiliary statsstatsGrand mean, pooled covariance, subgroup means

Key differences between Theorem 3 and T2.1():

  • Theorem 3 uses the true null mean \(\mu\); T2.1() uses the estimated grand mean \(\bar{\bar{x}}\) from Phase I data.

  • Theorem 3 uses the covariance of the same sample; T2.1() uses the pooled covariance from multiple Phase I subgroups.

  • Theorem 3 is asymptotic as \(n \to \infty\) for a single sample; T2.1() has \(m\) correlated statistics (one per subgroup) with finite-sample F/beta distributions under normality.

The control limits in cchart.T2.1() and cchart.T2.2() are based on exact finite-sample distributions under multivariate normality (F and beta distributions). These limits are not justified by Theorem 3 for non-normal data in finite samples. The theorem only guarantees that as the subgroup size \(n \to \infty\), a single T² statistic approaches a \(\chi^2_p\) distribution regardless of the underlying continuous distribution (provided the moment conditions hold).

References

Gneri, M. A. and Barbosa, E. P. (2006). "Robustez Asintótica de la Estadística de Hotelling". Sección 4.2, Teorema 3, pp. 34-36. IMECC-UNICAMP.

Montgomery, D. C. (2009). "Introduction to Statistical Quality Control". Chapter 11. Wiley.