Supplementary Material A Theory and Algorithm

A.1 Geodesic Autoencoder Algorithm Using the definitions from Sec. 3, we present the algorithm to train the geodesic autoencoder.Algorithm 2 Geodesic Autoencoder (GAE) Theorem 3. We consider a time-varying vector field It also admits a static formulation, i.e. minimizing with respect to The encoder network is three layers with ReLU activation functions in between layers. We used the authors' implementation of DSB with For the EB data, we set the scale to 1500. In Figure 1, we present an ablation study for the density loss and the geodesic embedding. To complement this figure, in Tab. In Tab. 3, we show the W Table 5 shows the 1-NN distance for several methods and datasets.