Relative shifts up to 2 ms permit model source ambiguity at 5 percent error in 29 of 133 selected Allen input pairs

We tested whether fitted voltage waveforms can identify which of two connected source cells was active. We used 133 pairs of excitatory inputs to shared targets in Allen cortical slice data: 89 mouse VisP pairs and 44 human TCx pairs. With exact fitted kernels and free positive amplitudes, the least relative L2 waveform error that permits a common response is sqrt((1-rho)/2), where rho is the normalized waveform overlap. At a 5 percent error bound, 5 mouse pairs and 1 human pair permit source ambiguity with known spike time. A grid of relative shifts from -2 to +2 ms gives ambiguity witnesses in 20 mouse pairs and 9 human pairs. The grid counts are lower bounds for continuous timing freedom. An independent high precision calculation agrees within 1.3e-11 in the error threshold. These are deterministic model results, not measured decoding errors. Uncertain fits, separate recordings, selected slice circuits and assumed event timing limit biological inference. Original code, aggregate results, a figure, source hashes and an external data recipe support reproduction. External source records are excluded from the bundle.

Allen Institute · source identification · synaptic kinetics · waveform overlap · timing uncertainty · convergent inputs