Residual Source-Participation Penalties in Calibration-Limited sEMG Personalization: Matched LDA and Fixed-Representation Comparisons

Cross-user variability motivates personalization of surface electromyography (sEMG) gesture recognition from limited user-specific calibration. Source weighting and representation reuse are established approaches, but improvements over ordinary pooling need not imply an advantage over target-centered fitting. We studied offline recognition in intact-participant NinaPro DB1 and DB2 tasks using complete-repetition calibration, linear discriminant analysis (LDA), and classifier-head adaptation on frozen convolutional neural network (CNN) representations. Initial evaluation identified a residual deficit after target weighting improved ordinary pooling. Stage 6C then provided a prespecified controlled re-evaluation of that previously observed deficit under matched preprocessing. Its sole primary contrast, H3, compared source-inclusive LDA at 75% target/25% source global mass with target-only LDA using identical target-calibration-derived preprocessing and matched fitting settings. Mean subject-level Anchor macro-F1, the equal mean at the lowest and maximum tested nonzero calibration budgets, was lower by 0.0465 in DB1 and 0.0420 in DB2; both nominal 95% t and bias-corrected and accelerated bootstrap intervals lay below zero. H3 was fixed before execution and targets were held out from mechanism development, but their participants, recordings and calibration/test partitions had already contributed Phase 5 outcomes that informed the question. Stage 6C is therefore not statistically independent evidence from Phase 5, and its uncertainty does not adjust for the outcome-informed research trajectory or hypothesis selection. Secondary H5 showed penalties of 0.0138 in DB1 and 0.0090 in DB2 within the same retained source-trained encoder and normalization. H5 changes both loss coefficients, relative regularization and the finite Adam trajectory. The contribution is comparator control and conditional empirical characterization: the residual LDA deficit persists when preprocessing is identical, and a related CNN penalty occurs for the specified finite head-training procedures. Representation reuse and continued source-example participation warrant separate evaluation; encoder benefit and a unique causal mechanism remain unestablished.