Subsequently for every set of runs done on the same cell pair and the same stimulation protocol, we computed :
Moreover we created <ItoPeak>scaled and <<ItoPeak>>scaled , scaled versions of <ItoPeak> and <<ItoPeak>> so that the values cover the range [-1,1] by scaling positive (negative) values by the maximum (minimum) response in the dataset. 

Distance from control and significance

Based on light level anatomy, we labeled each tested pair as overlapping or non-overlapping. We used the set of non overlapping pairs as a control (the null sample). Considering only two parameters, the scaled normalized integral and the correlation across flies (see Figure \ref{489066}), we calculated the Mahalanobis distance between the null sample and each data point, using a robust estimate of the covariance matrix \cite{Rousseeuw_1999} of the null sample. While single statistics never were sufficient to capture all relevant aspects of the response, we found that these two measurements recapitulated well distance measurements obtained by combining all the statistics. We then computed 99% confidence intervals on the distribution of distances by bootstrapping to determine significance.

Acknowledgments

We are grateful to Tanya Wolff and Arnim Jenett for help selecting the driver lines, to Janelia FlyCore and Karen Hibbard in particular for providing assistance in setting the initial crosses. We'd like to thank Heather Dionne for help with the alternative landing site LexA injections and Yi Sun, Allan Wong, Stephanie Hampel and members of the Jayaraman lab for fruitful discussions and comments on the manuscript.