The smallest fraction (or proportion) of contamination in the data that can cause an estimator or statistical procedure to produce arbitrarily large, meaningless, or otherwise completely misleading results; a global robustness metric.

Principle

An estimator's breakdown point quantifies its global resistance to outliers: above that fraction of contaminated observations, the estimator can be driven to pathological values regardless of sample size.

Demonstration

Illustrative scenario: the sample mean has breakdown point 0 under an additive contamination model because a single arbitrarily large outlier can make the mean arbitrarily large; the sample median in a symmetric univariate sample has breakdown point 50%, since over half the sample must be corrupted to force it arbitrarily far.

Misapplication

Confusing high breakdown point with overall superiority—an estimator with a high breakdown point can still be inefficient or biased in clean data, and breakdown point ignores local influence measures.

Consequence

High breakdown point estimators are preferred when gross contamination is plausible; practical response includes choosing robust estimators, combining with efficiency considerations, and using diagnostic tools to detect contamination.

Reversal

Zero or low breakdown point: an estimator is highly sensitive to a small number of extreme observations (e.g., mean), so minor contamination can invalidate inference.

Boundary

Applies within specific contamination models (e.g., Huber's contamination) and to the considered estimator class; it does not by itself quantify local robustness, efficiency, or model misspecification effects.

Semantic Tension

Tension with 'influence function' — breakdown point is a global worst-case measure, while influence functions capture local infinitesimal sensitivity; both are complementary but can suggest different robustness trade-offs.

Synthesis

The breakdown point condenses the concept of global robustness into a single contamination threshold: it signals how much gross corruption an estimator can tolerate before yielding arbitrarily bad outputs, and must be balanced against efficiency and other robustness diagnostics.