Noise placement changes edge deletion effects in linear models of selected Allen slice graphs

We tested single-edge deletion in equal-weight linear models of 247 mouse and 17 human Allen cortical slice graphs. With independent noise at all cells, deletion of an excitatory-to-inhibitory edge increased mean stationary variance for 42 of 113 mouse edges and four of 15 human edges. All such edges had a reverse inhibitory connection. The counts held at three gain budgets. At budget 0.8, noise applied only to excitatory cells reduced the counts to 26 and zero. Seven mouse edges and two human edges changed from an increase to a decrease. All models were stable before and after deletion by an absolute-weight bound. An exact two-cell formula shows how noise placement and weight balance control the deletion sign. Independent covariance solvers agreed within 1.8e-14. These are model results, not measured neural reliability or a species effect. Equal weights, small selected slice graphs, common decay, assumed independent noise, and no linked activity limit inference. Code, aggregate evidence, source hashes and external retrieval steps support reproduction. External record-level data are excluded.

Allen Institute · edge deletion · noise covariance · negative feedback · linear circuit model · structure and function