Shot noise in multi-tracer constraints on $f_\text{NL}$ and relativistic projections: Power Spectrum

TL;DR

This paper investigates how shot noise impacts multi-tracer power spectrum measurements of primordial non-Gaussianity and relativistic effects, finding that marginalization can mitigate degradation but ignoring cross shot noise causes systematics.

Contribution

It provides a detailed analysis of shot noise effects in multi-tracer surveys, highlighting the importance of modeling or marginalizing non-Poissonian noise for accurate parameter constraints.

Findings

Marginalization over shot noise limits constraint degradation to ~30%.

Ignoring cross shot noise induces large systematics at z<1.

Optimal sample division improves constraints and reduces errors.

Abstract

Multiple tracers of the same surveyed volume can enhance the signal-to-noise on a measurement of local primordial non-Gaussianity and the relativistic projections. Increasing the number of tracers comparably increases the number of shot noise terms required to describe the stochasticity of the data. Although the shot noise is white on large scales, it is desirable to investigate the extent to which it can degrade constraints on the parameters of interest. In a multi-tracer analysis of the power spectrum, a marginalization over shot noise does not degrade the constraints on $f_\text{NL}$ by more than $\sim 30$% so long as halos of mass $M\lesssim 10^{12}M_\odot$ are resolved. However, ignoring cross shot noise terms induces large systematics on a measurement of $f_\text{NL}$ at redshift $z<1$ when small mass halos are resolved. These effects are less severe for the relativistic projections, especially for the dipole term. In the case of a low and high mass tracer, the optimal sample division maximizes the signal-to-noise on $f_\text{NL}$ and the projection effects simultaneously, reducing the errors to the level of $\sim 10$ consecutive mass bins of equal number density. We also emphasize that the non-Poissonian noise corrections that arise from small-scale clustering effects cannot be measured with random dilutions of the data. Therefore, they must either be properly modeled or marginalized over.