Appendix A Further wavelet details

Using this identity, it is easy to check that the highpass filter must have zero-mean, i.e., X Then Eq. 12 and Eq. 13 provides the sufficient and necessary conditions on the highpass filter to build In this section, we show additional results for the experiments with synthetic data in Sec 4.1 . All experiments were run on an A WS instance of p3.16xlarge for a few days. Fig B4 calculates the distance between the learned wavelets and the groundtruth (DB5) wavelet, defined as in Sec 4.1, as the interpretation penalty varies. For a detailed overview of the data, see the original study [ 50 ]. In order to convert the raw fluorescence images to time-series traces, we use tracking code from previous work [ 52 ].