Is the cosmic microwave background really non-Gaussian?

Two recent papers have claimed detection of non-Gaussian features in the COBE Differential Microwave Radiometer sky maps of the cosmic microwave background. We confirm these results, but argue that Gaussianity is still not convincingly ruled out. Since a score of non-Gaussianity tests have now been published, one might expect some mildly significant results even by chance. Moreover, in the case of one measure that yields a detection, a bispectrum statistic, we find that if the non-Gaussian feature is real, it may well be due to detector noise rather than a non-Gaussian sky signal, since a signal-to-noise analysis localizes it to angular scales smaller than the beam. We study its spatial origin in case it is nonetheless due to a sky signal (e.g., a cosmic string wake or flat-spectrum foreground contaminant). It appears highly localized in the direction b = 39.degrees 5 l = 257 degrees, since removing a mere five pixels inside a single COBE beam area centered there makes the effect statistically insignificant. We also test Gaussianity with an eigenmode analysis which allows a sky map to be treated as a random number generator. A battery of tests of this generator yields results consistent with Gaussianity.