At 1000 model time units, inverse covariance ranked more links than short-lag curvature in two passive worm circuit models

We tested whether short-lag covariance curvature can recover direct electrical links from finite stationary records in two partial adult worm circuit models. Each symmetric passive model had 180 cells, unit independent white current noise and 90 observed cells. Exact stationary transitions supplied 12 trajectories and random masks per map. At duration 1000 model time units, mean precision at the known link count was 0.9647 and 0.9398 for inverse covariance, versus 0.1029 and 0.0961 for curvature at lag 0.01. Curvature had precision one with exact population covariance. Inverse covariance won all 24 primary finite-record comparisons. A lag of 0.1 improved finite-record curvature ranks but increased population drift error. Code, source hashes, aggregate scores and independent checks support reproduction. No recorded neural activity was tested. Two partial maps, assumed conductance and noise, known link counts, few replicas and no method tuning limit transfer.

electrical synapses · finite recordings · network inference · covariance curvature · hidden neurons · C. elegans · sampling error