---
title: Matrix Minimum Covariance Determinant (MMCD)
url: https://www.emergentmind.com/topics/matrix-minimum-covariance-determinant-mmcd
type: topic
---

# Matrix Minimum Covariance Determinant (MMCD)

Searching arXiv for MMCD and related papers to ground the article in the cited literature.
I’m checking arXiv for the Matrix Minimum Covariance Determinant literature and closely related extensions/applications.
Matrix Minimum Covariance Determinant (MMCD) is a robust location and covariance estimation procedure for data that are naturally represented as matrices rather than vectors. It extends the determinant-minimization principle of the classical Minimum Covariance Determinant (MCD) to the matrix-variate setting, so that the mean matrix and the rowwise and columnwise covariance matrices are estimated directly under a matrix-variate elliptical model with Kronecker-structured covariance, rather than after vectorization into dimension \(pq\). In this formulation, MMCD is matrix affine equivariant, admits strong consistency results, achieves a finite-sample breakdown point that can exceed the maximal achievable one for affine-equivariant estimators on vectorized data, and furnishes robust matrix Mahalanobis distances and Shapley-value decompositions for outlier detection and explanation [2403.03975]. The same estimator has subsequently been used as a robust covariance engine in multivariate functional outlier detection and in highly robust factored PCA for matrix-valued data [2605.20325], [2509.25957].

## 1. From classical MCD to the matrix-variate extension

The classical MCD is a high-breakdown estimator of multivariate location and scatter defined by the \(h\)-subset whose sample covariance determinant is smallest. Its raw location estimate is the mean of that subset, and its raw scatter estimate is the corresponding covariance matrix multiplied by a consistency factor. The method is affine equivariant, has bounded influence functions, and is a standard building block for robust distances and multivariate outlier detection. Its main structural limitation is that the covariance matrix of an \(h\)-subset must be nonsingular, so \(h>p\) is required and in practice \(n>5p\) is recommended; when \(p\) is large relative to \(n\), classical MCD becomes unstable or undefined [1709.07045].

MRCD addresses that high-dimensional vector-valued limitation by replacing the subset covariance with a convex combination of a target matrix and the subset covariance, thereby ensuring well-conditioning even when \(p>h\) or \(p>n\). However, MRCD remains a vector-valued construction and is not affine equivariant, even though it is location invariant and scale equivariant [1701.07086].

MMCD addresses a different but related problem: observations that are intrinsically matrix-valued. The motivation is that vectorizing a \(p\times q\) observation into a \(pq\)-vector destroys the natural two-way structure, inflates dimensionality, ignores row/column dependence, and makes robust covariance estimation substantially harder. MMCD therefore preserves the MCD principle of searching for a “best” trimmed subset, but applies it to a matrix-variate model with separable covariance. In the formulation used in robust factored PCA, the raw MMCD estimators reduce to the classical MCD when one dimension equals \(1\) [2403.03975], [2509.25957].

A recurrent source of terminological confusion is that some MCD-based procedures for structured data are not MMCD in the strict matrix-variate sense. For example, robust two-way MANOVA based on the MCD estimator replaces classical SSP matrices by reweighted MCD estimates on vector-valued observations within MANOVA cells; that framework explicitly does not introduce a separate matrix-variate MMCD estimator [1806.04106].

## 2. Model, estimands, and determinant criterion

For a sample
\[
\mathfrak{X} = (X_1,\dots,X_n), \qquad X_i\in\mathbb{R}^{p\times q},
\]
MMCD is formulated under a matrix-variate elliptical model
\[
X \sim ME(M,\Sigma^{\row},\Sigma^{\col},g),
\]
with density
\[
f(X) = \det(\Sigma^{\row})^{-q/2}\det(\Sigma^{\col})^{-p/2} g\!\left( \operatorname{tr}\!\left( (\Sigma^{\col})^{-1}(X-M)^\top(\Sigma^{\row})^{-1}(X-M) \right) \right).
\]
Equivalently, \(\mathrm{vec}(X)\) is multivariate elliptical with covariance
\[
\Sigma = \Sigma^{\col}\otimes \Sigma^{\row}.
\]
The matrix normal distribution is the Gaussian special case of this model class [2403.03975].

The squared matrix Mahalanobis distance is
\[
\mmd^2(X;\,M,\Sigma^{\row},\Sigma^{\col}) = \operatorname{tr}\!\left( (\Sigma^{\col})^{-1}(X-M)^\top (\Sigma^{\row})^{-1}(X-M) \right).
\]
Using binary weights \(w_i\in\{0,1\}\) with \(\sum_i w_i=h\), the weighted matrix-normal log-likelihood takes the form
\[
-\frac{1}{2} \sum_{i=1}^n w_i \Bigl( p\ln(\det(\Sigma^{\col})) + q\ln(\det(\Sigma^{\row})) + \mmd^2(X_i) + pq\ln(2\pi) \Bigr).
\]
MMCD chooses the subset \(H\) of size \(h\) that minimizes
\[
p\ln\!\bigl(\det(\hat\Sigma_H^{\col})\bigr) + q\ln\!\bigl(\det(\hat\Sigma_H^{\row})\bigr),
\]
equivalently
\[
\ln\det(\hat\Sigma_H^{\col}\otimes \hat\Sigma_H^{\row}) = p\ln(\det(\hat\Sigma_H^{\col})) + q\ln(\det(\hat\Sigma_H^{\row})).
\]
This is the matrix-valued analogue of selecting the subset with smallest covariance “volume” [2403.03975].

For a fixed subset \(H\subset\{1,\dots,n\}\), \(|H|=h\), the fitted estimators are
\[
\hat M_H = \frac{1}{h}\sum_{i\in H} X_i,
\]
\[
\hat\Sigma_H^{\row} = \frac{1}{qh} \sum_{i\in H} (X_i-\hat M_H)\,\hat\Omega_H^{\col}\,(X_i-\hat M_H)^\top,
\]
\[
\hat\Sigma_H^{\col} = \frac{1}{ph} \sum_{i\in H} (X_i-\hat M_H)^\top\,\hat\Omega_H^{\row}\,(X_i-\hat M_H),
\]
where \(\hat\Omega_H^{\row}=(\hat\Sigma_H^{\row})^{-1}\) and \(\hat\Omega_H^{\col}=(\hat\Sigma_H^{\col})^{-1}\). Because \(\Sigma^{\row}\) and \(\Sigma^{\col}\) are only identifiable up to a multiplicative constant, an identifiability convention is imposed by setting the first diagonal entry of \(\Sigma^{\col}\) to \(1\). For existence and uniqueness of the subset estimators, the required sample-size condition is
\[
h \ge \left\lfloor \frac{p}{q} + \frac{q}{p} \right\rfloor + 2.
\]
In the multivariate functional setting, the same MMCD criterion is applied to basis coefficient matrices, with \(h=\alpha n\) and \(\alpha\in[0.5,1]\) [2403.03975], [2605.20325].

## 3. Equivariance, breakdown, and statistical guarantees

A central property of MMCD is matrix affine equivariance. If
\[
Z_i = A X_i B + C,
\]
with invertible \(A\in\mathbb{R}^{p\times p}\), invertible \(B\in\mathbb{R}^{q\times q}\), and \(C\in\mathbb{R}^{p\times q}\), then
\[
\hat M_{\mathfrak Z}=A\hat M_{\mathfrak X}B+C,\qquad \hat\Sigma^{\row}_{\mathfrak Z}=A\hat\Sigma^{\row}_{\mathfrak X}A^\top,\qquad \hat\Sigma^{\col}_{\mathfrak Z}=B^\top\hat\Sigma^{\col}_{\mathfrak X}B.
\]
The squared matrix Mahalanobis distance is invariant under these transformations:
\[
\mmd^2(Z_i;\hat M_{\mathfrak Z},\hat\Sigma^{\row}_{\mathfrak Z},\hat\Sigma^{\col}_{\mathfrak Z}) = \mmd^2(X_i;\hat M_{\mathfrak X},\hat\Sigma^{\row}_{\mathfrak X},\hat\Sigma^{\col}_{\mathfrak X}).
\]
The matrix-variate paper emphasizes that this is stronger and more natural for matrix-valued data than vectorizing and allowing arbitrary affine transforms only on \(\mathrm{vec}(X)\), because vectorization respects only the Kronecker-structured subgroup \((B^\top\otimes A)\,\mathrm{vec}(X)+\mathrm{vec}(C)\) [2403.03975].

The finite-sample breakdown theory is likewise matrix-specific. Let
\[
d=\left\lfloor \frac{p}{q}+\frac{q}{p}\right\rfloor,\qquad m=\left\lfloor \min(n-h+1,\; h-(d+1))\right\rfloor.
\]
Then the MMCD estimators satisfy
\[
\varepsilon^\ast(\hat M,\mathfrak X) = \varepsilon^\ast(\hat\Sigma^{\row},\hat\Sigma^{\col},\mathfrak X) = \frac{m}{n}.
\]
The maximum breakdown point is attained for
\[
h=\left\lfloor \frac{n+d+2}{2}\right\rfloor,
\]
giving approximately
\[
\varepsilon^\ast_{\max} \approx \frac{1}{n}\left\lfloor \frac{n-d}{2}\right\rfloor.
\]
A major result is that, when \(p\ge 2\) and \(q\ge 2\), this is higher than the classical upper bound for affine-equivariant multivariate estimators on vectorized \(pq\)-dimensional data, which is bounded by about
\[
\frac{1}{n}\left\lfloor \frac{n-pq+1}{2}\right\rfloor.
\]
The robust FPCA paper presents the same phenomenon in its own notation and states that the corresponding maximum breakdown point is essentially close to \(50\%\) and approaches \(50\%\) as \(n\to\infty\) [2403.03975], [2509.25957].

Strong consistency is established for elliptical matrix-variate distributions:
\[
\|\hat M-M\|\xrightarrow{a.s.}0,\qquad
\|c(\alpha)\hat\Sigma^{\col}\otimes\hat\Sigma^{\row}-\Sigma^{\col}\otimes\Sigma^{\row}\|\xrightarrow{a.s.}0.
\]
The same work reports that raw MMCD has low efficiency at the normal model, whereas reweighted MMCD achieves much better efficiency, often above \(90\%\) for larger \(n\), while preserving robustness [2403.03975].

## 4. Optimization via C-steps and flip-flop updates

MMCD adapts the FastMCD concentration-step strategy to matrix-valued observations. Starting from an \(h\)-subset \(H_{\text{old}}\), the algorithm computes subset MLEs using the flip-flop algorithm for matrix-normal covariance estimation, evaluates all matrix Mahalanobis distances
\[
d_i^2(H_{\text{old}})=\mmd^2(X_i;\hat M_{H_{\text{old}}},\hat\Sigma^{\row}_{H_{\text{old}}},\hat\Sigma^{\col}_{H_{\text{old}}}),
\]
selects the \(h\) observations with smallest \(d_i^2\) to form \(H_{\text{new}}\), and refits. The determinant criterion is non-increasing across C-steps:
\[
p\ln(\det(\hat\Sigma^{\col}_{H_{\text{new}}})) + q\ln(\det(\hat\Sigma^{\row}_{H_{\text{new}}}))
\le
p\ln(\det(\hat\Sigma^{\col}_{H_{\text{old}}})) + q\ln(\det(\hat\Sigma^{\row}_{H_{\text{old}}})).
\]
The sequence of objective values is decreasing and bounded below, so it converges after finitely many subset changes; as in FastMCD, convergence to a global optimum is not guaranteed, hence the use of multiple starting subsets [2403.03975].

The implemented MMCD procedure draws \(500\) random elemental subsets of size
\[
d+2=\left\lfloor \frac{p}{q}+\frac{q}{p}\right\rfloor+2,
\]
runs only \(2\) MLE/C-step iterations on each for speed, keeps the \(10\) best candidates by covariance determinant, runs full C-steps to convergence from those \(10\) starts, chooses the best final subset, applies a consistency correction, and then performs a reweighting step. The normal-model consistency factor is
\[
c(\alpha)=\frac{\alpha}{F_{\chi^2_{pq+2}(\chi^2_{\alpha;pq})}},
\]
where \(\alpha=h/n\). The implementation uses the hard-threshold weight
\[
w(\mmd^2(X_i))= \begin{cases}
1,& i\in H \ \text{or}\ \mmd^2(X_i)<\chi^2_{pq;0.975},\\
0,& \text{otherwise}.
\end{cases}
\]
The method is implemented in the R package **robustmatrix** with a parallelized C++ backend, and subsampling is used for large \(n\) [2403.03975].

The computational advantage of respecting matrix structure is substantial. For \((p,q)=(5,20)\), contamination \(\varepsilon=0.2\), and \(99\%\) probability of a clean initial subset, the vectorized MCD would require about \(2.8\times 10^{10}\) elemental subsets, while MMCD needs only \(16\) [2403.03975].

## 5. Robust matrix Mahalanobis distances and Shapley explanations

Once MMCD estimates \((\hat M,\hat\Sigma^{\row},\hat\Sigma^{\col})\) are available, the robust matrix Mahalanobis distance is
\[
\mmd^2(X_i;\hat M,\hat\Sigma^{\row},\hat\Sigma^{\col}) = \operatorname{tr}\!\left( \hat\Omega^{\col}(X_i-\hat M)^\top \hat\Omega^{\row}(X_i-\hat M) \right).
\]
Under a matrix-normal model, the squared distance is compared with
\[
\chi^2_{pq;0.975}
\]
to flag outliers. This is the matrix analogue of MCD-based robust distance screening in ordinary multivariate analysis, but now adapted to row/column covariance structure [2403.03975].

A distinctive feature of MMCD is its associated Shapley-value decomposition of squared outlyingness. For cellwise contributions,
\[
\Phi(X)= (X-M)\circ \Omega^{\row}(X-M)\Omega^{\col}\in\mathbb{R}^{p\times q},
\]
and in indexed form
\[
\phi_a(x) = (x_a-\mu_a)\sum_{b=1}^{pq}(x_b-\mu_b)\omega_{ab}
= (x_{jk}-m_{jk})\sum_{i=1}^p\sum_{l=1}^q (x_{il}-m_{il})\omega_{ij}^{\row}\omega_{kl}^{\col}.
\]
These contributions sum to the squared Mahalanobis distance:
\[
\sum_{j=1}^p\sum_{k=1}^q \phi_{jk}(X)=\mmd^2(X).
\]
Rowwise and columnwise decompositions are
\[
\phi_{\row}(X)=\operatorname{diag}\!\bigl(\Omega^{\row}(X-M)\Omega^{\col}(X-M)^\top\bigr)\in\mathbb{R}^p,
\]
\[
\phi_{\col}(X)=\operatorname{diag}\!\bigl((X-M)^\top\Omega^{\row}(X-M)\Omega^{\col}\bigr)\in\mathbb{R}^q.
\]
The cellwise Shapley values are shift invariant, scale invariant under diagonal scaling matrices, and permutation equivariant under row/column permutations, but they are not generally matrix affine equivariant [2403.03975].

In multivariate functional data, this Shapley framework is generalized to decompose overall multivariate functional outlyingness into time-coordinate-specific contributions. The same paper states that the otherwise exponential computational complexity relative to the number of components is reduced to linear complexity while retaining the key properties of the Shapley value [2605.20325].

## 6. Functional data, robust FPCA, and empirical uses

In multivariate functional data with separable covariance structure,
\[
\bm{X}(t)=\bigl(X_1(t),\dots,X_p(t)\bigr)', \qquad \K(s,t)=\bm{\Sigma}^{\row}\,\kappa(s,t),
\]
basis expansion produces coefficient matrices
\[
\bm{A}_i=(\bm{a}_1^{(i)},\dots,\bm{a}_p^{(i)})\in\mathbb{R}^{m\times p},
\]
and the key theoretical link is
\[
\Cov(\vect(\bm{A}))=\bm{\Sigma}^{\row}\otimes \bm{\Sigma}^{\col}.
\]
If the process is Gaussian, then the coefficient matrix is matrix normal. The truncated multivariate functional Mahalanobis semi-distance satisfies
\[
\fmmd^2(\boldsymbol{X};mp) = \tr\!\left((\boldsymbol{\Sigma}^{\row})^{-1} (\bm{A}-\bm{M}_{\bm{A}})' (\boldsymbol{\Sigma}^{\col})^{-1} (\bm{A}-\bm{M}_{\bm{A}})\right) = \mmd^2(\bm{A}),
\]
so MMCD on the coefficient matrices directly induces a robust functional distance. In simulations, the resulting robust distances perform best or near-best across outlier types, including shift, shape, isolated, and covariance-induced outliers. In the ENSO application, the multivariate robust distance has the strongest Spearman correlation with an ENSO activity proxy, \(\rho=0.65\), higher than all competing methods; in resistance spot welding data, the robust MMCD-based method yields AUC close to \(1\); and in the fertility appendix, MMCD-based robust distances avoid masking under classical estimation [2605.20325].

In highly robust factored PCA, MMCD replaces maximum likelihood estimation in matrix-normal FPCA. The pipeline first estimates
\[
\widehat{\mathbf{M}}^H,\quad \widehat{\mathbf{\Sigma}_c}^H,\quad \widehat{\mathbf{\Sigma}_r}^H
\]
from a clean \(h\)-subset \(H\), and then performs PCA on the robust row and column covariance matrices. The resulting robust scores support a score–orthogonal distance analysis (SODA) plot that separates regular observations, good leverage points, orthogonal outliers, and bad leverage points. Across multiple contamination levels and outlier types, the reported practical ranking is
\[
\text{HRFPCA} > \text{TPCA} > \text{RFPCA} > \text{FPCA},
\]
and HRFPCA maintains excellent covariance estimation up to around \(49\%\) contamination [2509.25957].

The original MMCD paper also demonstrates direct applied use on matrix-valued data. In glacier weather data, MMCD flags \(23\) years as outliers and Shapley values identify the responsible variables and months. In the DARWIN handwriting data, Alzheimer’s patients have generally larger robust matrix Mahalanobis distances than healthy subjects, and rowwise Shapley values indicate which handwriting features are most distinctive. In surveillance video frames treated as matrices, robust distances spike when a man enters the scene, and cellwise Shapley values localize the outlying pixels on the contour and head [2403.03975].

Source: https://www.emergentmind.com/topics/matrix-minimum-covariance-determinant-mmcd