FrugalMadStat

Constant-memory adaptive median and median absolute deviation estimates.

The first observation seeds the center and zero deviation. Each later update moves both estimates toward their targets by at most half stepSize, clamped to avoid overshooting. The deviation target is the absolute residual against the pre-update center. Constant steps forget earlier centers gradually; these estimates track a changing distribution rather than reconstructing its historical MAD.

Positive weights each count as one observation, as in com.eignex.kumulant.stat.quantile.FrugalQuantileStat. Zero and non-finite weights are inert; negative weights throw because the recurrence has no inverse. Choose stepSize in the input's units: larger steps adapt faster with more jitter.

Use cases: lightweight robust center/scale monitoring where bounded state and adaptation matter more than quantile accuracy. Use MadStat for digest-based historical estimates, observation weights, and digest-backed merging.

Memory: O(1); two estimates, an initialization flag, and a lock.

Update: O(1); two bounded adjustments, with no per-update allocation.

Concurrency: Body locked under any concurrent level. Updates, reads, and resets see a coherent pair; estimates remain sensitive to observation order. Merge adopts a nonempty remote pair while empty and otherwise averages both estimates equally, without reconstructing the combined distribution or its observation weights.