Estimating the intrinsic dimensionality using Normalizing Flows - Supplementary A Theoretical appendix A.1 Singular value evolution under normal space noise In [ 4

Neural Information Processing Systems 

The complexity of a) depends on the choosen architecture for the NFs. The complexity of b) depends on the operational complexity of the NF i.e. the number of operations We repeat this experiment and report the result in Figure S2. The rows show the datasets we trained on, while the columns represent the datasets on which we estimate the ID. All the intrinsic dimensionalities are correctly retrieved. Though, when the flow is not expressive enough, we failed to estimate the true ID. 4 Figure S2: Samples from the lolipop dataset.