In the published article we used Minkowski functionals, V0, V1, and V2(Mantz et al. 2008), to analyze the statistical properties of small-scale structures produced by our generative adversarial network (GAN)-based Foreground Scale Extender (FORSE) algorithm6. The code used to produce these functionals contained an error that led to it considering only a portion of the input images in the computation. As a result, Figures 4 and 7 in the published article were impacted, and corrected versions are provided in Figures 4 and 7, respectively. The superposition of the Minkowski functional (V0, V1, and V2) computed from the small-scale features generated by the GAN and those obtained from the real total intensity observations are at the level of (64%, 61%, 60%) for Stokes I, (60%, 61%, 62%) for Q, and (64%, 62%, 63%) for U maps, respectively. Despite the lower values obtained in comparison to the previous version, these results still demonstrate the ability of FORSE to generate highly non-Gaussian features, as previously concluded. This is especially clear in the case of polarization, as shown in Figure 7, where the distribution of the functionals from the generated maps (orange) is much closer to the target distribution (black lines) than to that of a Gaussian field (green). We thank Viraj Manwadkar and Susan Clark for finding the error in our computation of Minkowski functionals and Jian Yao for correcting and accelerating the code.
