Robust Covariance Estimation (MCD)
Minimum Covariance Determinant Estimation · Also known as: minimum covariance determinant, MCD estimator, robust covariance estimation, Robust Kovaryans Tahmini (MCD)
Robust Covariance via the Minimum Covariance Determinant (MCD) estimates a multivariate mean vector and covariance matrix that are not distorted by outliers. It was made practical by the Fast-MCD algorithm of Rousseeuw and Van Driessen (1999), building on Rousseeuw's earlier work on robust estimation.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use robust covariance when you work with continuous, multivariate data that may contain outliers and you need a mean and covariance you can trust for downstream tasks such as outlier detection or robust distances. It tolerates up to about 50% contamination and needs a reasonable sample (at least about 50 observations, and more observations than variables). It is not appropriate for very small samples (n < 20) or when the outlier share is so high that even the clean subset is unreliable.
Strengths & limitations
- Resistant to outliers: tolerates up to roughly 50% contamination while still recovering the true location and scatter.
- Does not assume normality — it only requires continuous multivariate data.
- Made computationally practical for real datasets by the Fast-MCD concentration-step algorithm.
- Needs more observations than variables and a moderate sample (at least about 50); it is unreliable for n < 20.
- When the outlier ratio is very high (above about 25%), the quality of the selected clean subset degrades.
- It summarises location and scatter only; it is a building block for diagnostics rather than a full inferential model.
Frequently asked
What does the 'minimum determinant' actually minimise?
MCD searches over subsets of the data for the one whose covariance matrix has the smallest determinant. A small determinant means the smallest-volume scatter ellipsoid, i.e. the most tightly concentrated subset, which is taken to be the clean, outlier-free core of the data.
How many outliers can MCD tolerate?
Up to roughly 50% of the observations can be contaminated before the estimator breaks down. In practice, when the outlier ratio rises above about 25% the quality of the selected subset begins to degrade, so very heavy contamination should be handled with care.
Why is the Fast-MCD algorithm important?
The exact MCD requires searching every possible subset, which is computationally impossible for realistic data. The Fast-MCD algorithm of Rousseeuw and Van Driessen (1999) uses iterative concentration steps to converge to the minimum-determinant subset quickly, which is what makes MCD usable in practice.
What should I use when my sample is too small for MCD?
MCD needs more observations than variables and becomes unreliable below about 20 observations. For very small samples, a coordinate-wise robust scale such as the median absolute deviation (MAD) is a safer fallback.
Sources
- Rousseeuw, P. J. & Van Driessen, K. (1999). A Fast Algorithm for the Minimum Covariance Determinant Estimator. Technometrics, 41(3), 212-223. DOI: 10.1080/00401706.1999.10485670 ↗
- Rousseeuw, P. J. & Leroy, A. M. (1987). Robust Regression and Outlier Detection. Wiley. ISBN: 978-0471488552
How to cite this page
ScholarGate. (2026, June 1). Minimum Covariance Determinant Estimation. ScholarGate. https://scholargate.app/en/statistics/robust-covariance
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Least Trimmed SquaresStatistics↔ compare
- MAD EstimationStatistics↔ compare
- Robust ANOVAStatistics↔ compare
- Theil-Sen EstimatorStatistics↔ compare