Inferential Targets Under Multivariate Contamination

Why the Inferential Target Comes First

A multivariate test answers a question about a particular population quantity. That quantity might be the mean vector, the covariance matrix, or a robust location or scatter functional. These targets can agree under a reference model and differ under contamination. Consequently, choosing a robust procedure does not by itself specify the scientific hypothesis: the target must be stated explicitly.

We use the contamination model $F_{\varepsilon}=(1-\varepsilon)F_0+\varepsilon G$, where $F_0$ is a reference distribution, $G$ is a contaminating distribution, and $0\leq\varepsilon<1$. The mixture describes the distribution of an observation: with probability $1-\varepsilon$ it follows $F_0$, and with probability $\varepsilon$ it follows $G$. It is a model for departures from the reference distribution, not a claim that every unusual observation is erroneous.

Definition: If $X$ is a $p$-dimensional random vector with distribution $F$, its population mean is $\mu(F)=E_F[X]$, when this expectation exists, and its covariance matrix is $\Sigma(F)=E_F[(X-\mu(F))(X-\mu(F))^{\mathsf T}]$, when the second moments exist. A location functional $T(F)$ assigns a $p$-vector to a distribution; a scatter functional $S(F)$ assigns a symmetric positive semidefinite $p\times p$ matrix. These functionals are distinct targets unless they are explicitly defined to coincide.

The mean and covariance are classical moment parameters. Robust location and scatter functionals instead use a specified distributional summary designed to limit sensitivity to some deviations from a model. Examples include location and scatter functionals associated with the minimum covariance determinant (MCD) and S-estimators. Their exact definitions and behavior depend on the chosen functional and its calibration. In particular, a functional that is Fisher-consistent for the mean or covariance under a reference model need not equal those quantities for every distribution.

How Contamination Changes Classical Targets

When the component distributions have finite second moments, the mixture mean and covariance have exact forms. The covariance identity makes clear that contamination can change scatter in two ways: through within-component variability and through separation between component means.

Theorem: Suppose $F_0$ and $G$ have finite second moments. Write their means as $\mu_0,\mu_G$ and their covariance matrices as $\Sigma_0,\Sigma_G$. For $F_{\varepsilon}=(1-\varepsilon)F_0+\varepsilon G$, $$ \mu(F_{\varepsilon})=(1-\varepsilon)\mu_0+\varepsilon\mu_G $$ and $$ \Sigma(F_{\varepsilon})=(1-\varepsilon)\Sigma_0+\varepsilon\Sigma_G +\varepsilon(1-\varepsilon)(\mu_G-\mu_0)(\mu_G-\mu_0)^{\mathsf T}. $$

Proof. Linearity of expectation gives $$ \mu(F_{\varepsilon})=E_{F_{\varepsilon}}[X] =(1-\varepsilon)E_{F_0}[X]+\varepsilon E_G[X] =(1-\varepsilon)\mu_0+\varepsilon\mu_G. $$ For the covariance, condition on which mixture component generated $X$. The expected conditional covariance is $(1-\varepsilon)\Sigma_0+\varepsilon\Sigma_G$. The conditional means are $\mu_0$ and $\mu_G$, and their mixture mean is $\mu(F_{\varepsilon})$. Their covariance is $$ (1-\varepsilon)(\mu_0-\mu(F_{\varepsilon}))(\mu_0-\mu(F_{\varepsilon}))^{\mathsf T} +\varepsilon(\mu_G-\mu(F_{\varepsilon}))(\mu_G-\mu(F_{\varepsilon}))^{\mathsf T}. $$ Since $\mu_0-\mu(F_{\varepsilon})=-\varepsilon(\mu_G-\mu_0)$ and $\mu_G-\mu(F_{\varepsilon})=(1-\varepsilon)(\mu_G-\mu_0)$, this expression reduces to $$ \varepsilon(1-\varepsilon)(\mu_G-\mu_0)(\mu_G-\mu_0)^{\mathsf T}. $$ Adding the expected conditional covariance gives the stated formula. The finite-second-moment assumption ensures that all these covariance matrices exist. $\square$

Worked Example: A Point Contaminant Changes Both Moments

Let $F_0=N_2((0,0)^{\mathsf T},I_2)$, and let $G$ put all its probability at $a=(4,-2)^{\mathsf T}$. Take $\varepsilon=0.05$. Find the mean and covariance of the contaminated distribution.

The reference mean is $\mu_0=(0,0)^{\mathsf T}$ and $\Sigma_0=I_2$. The point distribution has $\mu_G=a$ and $\Sigma_G=0$. Thus $$ \mu(F_{\varepsilon})=0.95(0,0)^{\mathsf T}+0.05(4,-2)^{\mathsf T} =(0.2,-0.1)^{\mathsf T}. $$ For the covariance, the between-component coefficient is $0.05(0.95)=0.0475$, and $$ aa^{\mathsf T}= \begin{pmatrix}16&-8\\-8&4\end{pmatrix}. $$ Therefore $$ \Sigma(F_{\varepsilon})=0.95I_2+0.0475 \begin{pmatrix}16&-8\\-8&4\end{pmatrix} = \begin{pmatrix}0.95&0\\0&0.95\end{pmatrix} + \begin{pmatrix}0.76&-0.38\\-0.38&0.19\end{pmatrix} = \begin{pmatrix}1.71&-0.38\\-0.38&1.14\end{pmatrix}. $$ The contaminant shifts the mean and adds variability in the direction of $a$, including a negative off-diagonal covariance. A null hypothesis about the reference covariance $I_2$ is therefore not the same as a null hypothesis about the covariance of the contaminated population.

Robust Functionals Are Different Population Targets

A robust location functional $T(F)$ and robust scatter functional $S(F)$ are summaries of the distribution $F$, just as the mean and covariance are. Their definitions determine what a hypothesis about them means. For example, $T(F)=\mu_0$ is a hypothesis about the selected robust location functional. It is not automatically the hypothesis $\mu(F)=\mu_0$. Likewise, $S(F)=\Sigma_0$ and $\Sigma(F)=\Sigma_0$ are generally different assertions.

A functional may agree with the classical parameter for every distribution in a model family yet disagree outside that family. This property is often called Fisher consistency for the parameter under that model. It does not make the functional identical to the parameter on arbitrary contaminated distributions. Robustness concerns the effect of departures on the functional or estimator; it does not erase the distinction between the quantity being estimated and another population quantity.

Result: For a continuous, strictly increasing univariate mixture distribution $H_{\varepsilon}(x)=(1-\varepsilon)H_0(x)+\varepsilon H_G(x)$, its unique median $m_{\varepsilon}$ satisfies $H_{\varepsilon}(m_{\varepsilon})=1/2$. If $H_{\varepsilon}(0)<1/2$, then $m_{\varepsilon}>0$; if $H_{\varepsilon}(0)>1/2$, then $m_{\varepsilon}<0$.

Justification. Strict increase and continuity imply that $H_{\varepsilon}$ crosses the value $1/2$ at exactly one point, which is the median. If $H_{\varepsilon}(0)<1/2$, the crossing must occur to the right of zero, so $m_{\varepsilon}>0$. If $H_{\varepsilon}(0)>1/2$, it must occur to the left. This establishes the result without assuming that the median is unchanged by contamination. $\square$

Worked Example: Contamination Moves a Robust Location Target

Consider a bivariate distribution with independent coordinates, each having the mixture distribution $0.8N(0,1)+0.2N(3,1)$. Find the direction and a useful bound for each coordinate’s population median.

The marginal mixture distribution function is $$ H(x)=0.8\Phi(x)+0.2\Phi(x-3), $$ where $\Phi$ is the standard normal distribution function. At zero, $$ H(0)=0.8(0.5)+0.2\Phi(-3)\approx 0.4+0.2(0.00135)=0.40027<0.5. $$ At $x=0.5$, $$ H(0.5)=0.8\Phi(0.5)+0.2\Phi(-2.5) \approx 0.8(0.69146)+0.2(0.00621) \approx 0.55317+0.00124=0.55441>0.5. $$ The mixture CDF is continuous and strictly increasing. Its unique median is consequently between $0$ and $0.5$. The coordinatewise median vector is therefore $(m,m)^{\mathsf T}$, with $0<m<0.5$, whereas the mixture mean vector is $(0.2(3),0.2(3))^{\mathsf T}=(0.6,0.6)^{\mathsf T}$. Both targets move from zero, but they move by different amounts because they summarize the distribution differently.

Reading a Null Hypothesis Under Contamination

A test is interpretable only after its null hypothesis is connected to its target. A test built to assess a mean vector concerns $H_0:\mu(F)=\mu_0$, subject to its model and calibration assumptions. A procedure targeting a robust location functional concerns $H_0:T(F)=t_0$. These nulls need not be equivalent. The same distinction applies to a covariance matrix and a robust scatter functional.

Worked Example: One Distribution, Distinct Location and Scatter Statements

Let $F=0.9N_2((0,0)^{\mathsf T},I_2)+0.1N_2((4,0)^{\mathsf T},I_2)$. Compare the classical mean and covariance targets with the coordinatewise median target. This comparison identifies population quantities; it does not prescribe a test.

The component mean difference is $\delta=(4,0)^{\mathsf T}$, and both components have covariance $I_2$. The mixture mean is $$ \mu(F)=0.9(0,0)^{\mathsf T}+0.1(4,0)^{\mathsf T}=(0.4,0)^{\mathsf T}. $$ The mixture covariance is $$ \Sigma(F)=I_2+0.1(0.9)\delta\delta^{\mathsf T} =I_2+0.09\begin{pmatrix}16&0\\0&0\end{pmatrix} =\begin{pmatrix}2.44&0\\0&1\end{pmatrix}. $$ The first coordinate’s mixture CDF at zero is $$ 0.9\Phi(0)+0.1\Phi(-4)\approx 0.45+0.1(0.0000317)=0.45000317<0.5. $$ At $0.2$, it is $$ 0.9\Phi(0.2)+0.1\Phi(-3.8)\approx 0.9(0.57926)+0.1(0.000072)=0.52134>0.5. $$ So its median lies strictly between $0$ and $0.2$; the second coordinate’s median is zero because its distribution is unchanged standard normal. Thus the mean vector, coordinatewise median vector, and covariance matrix give distinct descriptions of the same population. A test of zero mean and a test of zero robust location would ask separate questions, even though both location targets are nonzero here. A robust scatter test would likewise require specifying its scatter functional and null, rather than silently interpreting its result as a test of $\Sigma(F)=I_2$.

Common Pitfall: “Robust” Does Not Mean “Tests the Clean Model”

A common error is to reason that because a robust method downweights unusual observations, its null hypothesis must be about the uncontaminated distribution $F_0$. That conclusion does not follow. The method produces an estimator of a specified functional of the observed population distribution, and the functional’s value under $F_{\varepsilon}$ may differ from its value under $F_0$. Downweighting can reduce sensitivity to some contaminating patterns; it does not reveal which observations came from which component or automatically recover $F_0$.

The correction is to state the target and the model separately. Ask whether the scientific question concerns the observed population mean or covariance, a robust functional of the observed distribution, or a parameter of a stipulated reference model. Then verify that the estimator and inference procedure are calibrated for that target under the assumptions actually being used. This distinction matters even when classical and robust summaries happen to agree at a clean reference distribution.

Key takeaway: Contamination changes population targets, not merely sample calculations. A claim about a mean or covariance is not interchangeable with a claim about a robust location or scatter functional; interpretation begins by naming the functional in the null hypothesis.

Check Your Understanding

For each question, identify the target before deciding what a result would mean.

Under what moment conditions do the mixture mean and covariance formulas apply?
In the point-contamination example, which term in the covariance formula creates the off-diagonal entries?
Why does a Fisher-consistent robust location functional not have to equal the mean under every contaminated distribution?
For a continuous strictly increasing mixture CDF, what does $H_{\varepsilon}(0)<1/2$ imply about its median?
Explain why rejecting a null about a robust scatter functional does not automatically establish that the population covariance differs from a specified matrix.