Notes. For each dataset (left column), the table shows maximal values of the estimated bispectral coherence (), estimates of the skewness from the data () and from the bispectrum (), as well as the maximal value obtainable for low-pass filtered data (γ^{filt, max}). By modeling the data with a Gaussian field plus a weak quadratic nonlinear component, we obtain estimates of the fraction of variance due to this nonlinear component (), the maximal value obtainable when low-pass filtering the data (), and the true value of f_NL, which is known for the synthetic datasets only. See Figs. 6 and 7 for how and depend on low-pass filters applied to the data. The estimate is biased and differs from the true f_NL due to second-order effects, see Eqs. (D.5) and (D.6). For the smoothed TD data, was computed here by taking the apodized region in Fig. 1 into account to show the consistence with .