A prior from shared motor input did not reduce mean error in a fixed linear model of wing motion from calcium on 31 fly recording days

We tested whether shared input to wing motor neurons improves a linear estimate of wing motion from muscle calcium. We used 72,219 wingbeats from 31 recording days and all 80 released left-wing angle coefficients. Each outer test excluded a complete day. Inner day tests chose the ridge penalty. A fixed prior used input cosine from 11 matched muscles, with unit prior diagonal. The primary MANC prior increased mean error by 0.0170% relative to generic ridge. A 95% range from day resamples with fits fixed was -0.0802% to +0.1180% for this error increase. Male CNS and female BANC priors increased error by 0.0114% and 0.0365%. The MANC two-edge prior increased it by 0.4777%. Complete parameter searches for 199 full and 199 attachment-group label changes gave tails of 0.275 and 0.095 for the recorded labels. The generic model reduced error by 4.94% relative to the training response mean, so this instantaneous linear task leaves most error unexplained. We independently scanned all 170,782,903 source connection rows and matched every direct and required upstream record. The result limits this fixed use of shared input. It does not reject motor modules, dynamic models, or precise spike-phase control. Day groups do not establish independent flies. The three maps share the same functional data. The package preserves exact day statistics, complete input extracts, source hashes, code, all fits and label draws, figures, and an offline repeat.

connectome · Drosophila · motor control · wing motion · muscle calcium · shared input · generalized ridge regression · nested validation · sufficient statistics · reproducibility