Figure 3
From: Estimating the intrinsic dimension of datasets by a minimal neighborhood information
Estimated dimension d vs the number of points N in logarithmic scale; for each value of N the dataset is partitioned in a number of independent sets containing exactly N points, d is computed on each subdataset and a measure d(N) is obtained as an average of these values. In Panel A we study the case of a uniform plane of 50000 points in dimension 2 perturbed by a Gaussian noise with variance σ along 20 independent directions; σ takes the three values 0.0, 0.0001 and 0.0002. In Panel B we analyze a dataset composed of a two-dimensional Gaussian of 50000 points wrapped around a Swiss Roll and perturbed by a gaussian noise with variance σ along 20 independent directions. Again σ takes the three values 0.0, 0.0001 and 0.0002.
