Day 868

Thursday.


1504.03979
Do we need model-dependent covariances when we test cosmological models with galaxy power spectra?
Kalus, Percival, Samushia

Consider the shape of the posterior distribution to be used when fitting cosmo models to power spectra measured from galaxy surveys.  At very large scales, Gaussian posterior distributions in the power do not approximate the posterior distribution P_R expected for a Gaussian density field delta_k, even if the covariance matrix is varied according to the model to be tested.  Compare alternative posterior distributions with P_R, both mode-by-mode and in terms of expected f_NL-measurements.  Marginalizing over a Gaussian posterior distribution P_f with fixed covariance matrix yields a posterior mean value of f_NL which, for a data set with the characteristics of Euclid, will be underestimated by Delta f_NL=0.4, while for SDSS DR9 BOSS it will be underestimated by Delta f_NL=19.1.  The inverse cubic normal distribution (P_ICN) agrees very well with P_R at all scales and for all data sets, hence providing the same marginalized value.  Adopting this likelihood function means that different covariance matrix for each model is not required to be tested:  this dependence is absorbed into the functional form of the posterior.  Thus, the computational burden of analysis is significantly reduced.