mathematics//statistics//robust statistics

Robust statistics is the set of estimators that limit how much weight any single data point can carry, so that a few wrong values cannot drag the answer arbitrarily far, and engineers reach for it whenever real data contain **outliers**: a bit flipped on a bus, a loose connector, a GPS jump from reflections off a building. The ordinary tools are fragile because they square errors. A least-squares fit gives an error of 100 ten thousand times the weight of an error of 1, so one absurd reading pulls the line, the mean or the calibration as far as it likes.

Robust statistics is the set of estimators that limit how much weight any single data point can carry, so that a few wrong values cannot drag the answer arbitrarily far, and engineers reach for it whenever real data contain outliers: a bit flipped on a bus, a loose connector, a GPS jump from reflections off a building. The ordinary tools are fragile because they square errors. A least-squares fit gives an error of 100 ten thousand times the weight of an error of 1, so one absurd reading pulls the line, the mean or the calibration as far as it likes.

The robust replacements each cap that weight. The median replaces the mean: one reading of 1,000 among readings near 20 does not move it. The median absolute deviation (MAD), the median of the distances to the median, replaces the standard deviation; multiplied by 1.4826 it equals σ\sigmaσ for Gaussian data, and a single 1,000 cannot move it either. In a fit, the Huber loss is quadratic near zero and linear far away, so small residuals are treated as in least squares and large ones only pull in proportion to their size; the sum of absolute residuals (least absolute deviations) is the maximum likelihood fit under Laplace noise, whose spikes are more frequent than a Gaussian's, and in the simplest case its minimum is the median. When outliers are a large share of the data, RANSAC fits on random small subsets and keeps the one most of the data agree with.

An outlier may be a measurement error or the first symptom of the fault being looked for.

Deleting outliers until the model fits is taking the batteries out of the smoke detector because it beeps. Mark them, count them and look at when they appear: concentrated on one machine or one shift, they are already a lead (anomaly detection).

The cheapest robust monitor fits in a PLC. Flag any reading farther than 3.5 MADs from the median of a clean week; the median and the MAD ignore the very spikes the rule is meant to catch, where a mean and a σ\sigmaσ computed on the same data would be inflated by them (detection threshold).

Robustness has a measurable size, the breakdown point: the share of arbitrarily bad points an estimator tolerates before its answer can be pushed anywhere. The mean breaks with one point; the median and the MAD hold up to half the data.

The price is efficiency. On clean Gaussian data the median needs about 57 % more samples than the mean for the same precision, and robust fits have no closed form (they iterate, as reweighted least squares). A filter can be robustified the same way (robust Kalman filter).