In this paper we derive and test the flat sky approximation for galaxy number counts. We show that, while for the lensing term it reduces to the Limber approximation, for the standard density and redshift space distortion it is different and very accurate already at low $\ell$ while the corresponding Limber approximation completely fails. At equal redshift the accuracy of the standard terms is around 0.2% at low redshifts and 0.5% for redshift $z=5$, even to low $\ell$. At unequal redshifts the precision is less impressive and can only be trusted for very small redshift differences, $\Delta z<\Delta z_0 \simeq 3.6\times10^{-4}(1+z)^{2.14}$, but the lensing terms dominate for $\Delta z>\Delta z_1 \simeq 0.33(r(z)H(z))/(z+1)$. The Limber approximation achieves an accuracy of 0.5% above $\ell\simeq 40$ for the pure lensing term and above $\ell\simeq 80$ for the lensing-density cross-correlation. Besides being very accurate, the flat sky approximation is also very fast and can therefore be useful for data analysis and forecasts with MCMC methods.

Next-generation galaxy surveys will be able to measure perturbations on scales beyond the equality scale. On these ultra-large scales, primordial non-Gaussianity leaves signatures that can shed light on the mechanism by which perturbations in the early Universe are generated. We perform a forecast analysis for constraining local type non-Gaussianity and its two-parameter extension with a simple scale-dependence. We combine different clustering measurements from future galaxy surveys -- a 21cm intensity mapping survey and two photometric galaxy surveys -- via the multi-tracer approach. Furthermore we then include CMB lensing from a CMB Stage 4 experiment in the multi-tracer, which can improve the constraints on bias parameters. We forecast $\sigma(f_{\rm NL}) \simeq 0.9$ (1.4) by combining SKA1, a Euclid-like (LSST-like) survey, and CMB-S4 lensing. With CMB lensing, the precision on $f_{\rm NL}$ improves by up to a factor of 2, showing that a joint analysis is important. In the case with running of $f_{\rm NL}$, our results show that the combination of upcoming cosmological surveys could achieve $\sigma(n_{\rm NL}) \simeq 0.12$ (0.22) on the running index.