Support errors and metric masking in constrained EDS composition regression

Abstract True-support error can decrease while an elemental composition assigns more mass incor rectly. We demonstrate this discrepancy in constrained regression of 5 keV energy-dispersive X-ray spectra, using 76 audited material–platform records representing 69 reference com positions. With identical compositions held out together, Ridge regression gives a pooled R2 of 0.236, yet its ternary mean absolute error exceeds that of a uniform-on-support base line. After one true element is removed, adding the highest-scored absent element lowers ternary true-support MAE by 3.04 percentage points while raising full-vector L1 error by 33.14 points. An exact mass decomposition attributes the discrepancy to false-positive mass excluded from the support score. For fixed raw scores, Euclidean simplex projection makes full L1 nondecreasing when the allowed support gains only absent elements. With an inter nally trained support detector, lowering the threshold from 0.5 to 0.3 recovers true elements in 37 records and improves both mean errors. Nine of the 76 records have lower true-support MAE and higher full L1