Abstract
Analysis of pulsar timing data have provided evidence for a stochastic
gravitational wave background in the nHz frequency band. The most plausible
source of such a background is the superposition of signals from millions of
supermassive black hole binaries. The standard statistical techniques used to
search for such a background and assess its significance make several
simplifying assumptions, namely: i) Gaussianity; ii) isotropy; and most often
iii) a power-law spectrum. However, a stochastic background from a finite
collection of binaries does not exactly satisfy any of these assumptions. To
understand the effect of these assumptions, we test standard analysis
techniques on a large collection of realistic simulated datasets. The dataset
length, observing schedule, and noise levels were chosen to emulate the
NANOGrav 15-year dataset. Simulated signals from millions of binaries drawn
from models based on the Illustris cosmological hydrodynamical simulation were
added to the data. We find that the standard statistical methods perform
remarkably well on these simulated datasets, despite their fundamental
assumptions not being strictly met. They are able to achieve a confident
detection of the background. However, even for a fixed set of astrophysical
parameters, different realizations of the universe result in a large variance
in the significance and recovered parameters of the background. We also find
that the presence of loud individual binaries can bias the spectral recovery of
the background if we do not account for them.