Similar to Fig. 3, showing model parameter posterior variance, conditioned on noisy datavectors, estimated with the SBI (the method and normalizing flow model used in each experiment is listed in each panel). The variances between NLE and NPE for each compression are similar except for the small underestimation of the variance in the NLE experiments, due to the likelihood being additionally informed by the prior. The additional simulations, not labeled on the x-axis, but required for the separate compressions (where the true covariance is unknown) are noted for each method. When the true data covariance is not known, requiring the use of double the number of simulations, the reconstructed posterior errors from SBI are significantly higher than the Dodelson & Schneider (2013) corrected errors for less than 4 × 10³ simulations.