Title: On the accuracy of common moment-based radiative transfer methods for simulating reionization

Abstract: Cosmological simulations of reionization often treat radiative transfer by solving for the monopole and dipoles of the intensity field and by making ansatz for the quadrupole moments to close the system of equations. We investigate the accuracy of the most common closure methods, i.e. Eddington tensor choices. We argue that these algorithms are most likely to err after reionization and study quasi-analytic test problems mimicking these situations: large-scale post-reionization ionizing background fluctuations and radiative transfer in a predominantly ionized medium with discrete absorbers. We show that OTVET and M1 over-ionize self-shielding absorbers when fixing the background photoionization rate, leading to 30-40% higher emissivity to balance the increased recombination. This over-ionization results in a simulation run with these algorithms having a factor of ~2 lower average metagalactic photoionization rate relative to truth given an ionizing emissivity. Furthermore, these algorithms are unlikely to reproduce ionizing background fluctuations on scales below the photon mean path: OTVET tends to overpredict the fluctuations there when the simulation box is smaller than twice the mean free path and underpredict otherwise, while M1 drastically underpredicts these fluctuations. As a result, these numerical methods are likely not sufficiently accurate to interpret the Ly$\alpha$ forest opacity fluctuations observed after reionzation. We show that a high number of angular directions need to be followed to capture the post-reionization ionizing background fluctuations accurately with ray-tracing codes. Lastly, we argue that the strong dependence of the post-reionization ionizing background on the value of the reduced speed of light found in many simulations signals that the ionizing photon mean free path is several times larger in such simulations than the observationally measured value.