Three-dimensional scene reconstruction using Roman slitless spectra

TL;DR

This work tackles host-galaxy contamination in Roman slitless supernova spectra by reconstructing a 3D host datacube from observations at multiple roll angles. The forward model maps the datacube to a 2D spectrum via a linear operator H = R C L that encompasses wavelength interpolation, PSF convolution, and drizzle-based geometric transformation, enabling accurate host predictions at new roll angles for subtraction. By constructing a priors from multi-filter imaging and optimizing a regularization parameter $\alpha$ through cross-validation, the method achieves extremely low systematic residuals and reduced noise compared with single-spectrum subtraction. The approach significantly improves the reliability of SN spectra for cosmological inferences, and is adaptable to arbitrary scenes, though it relies on accurate characterizations of the PSF, dispersion, and instrument throughputs; future work will address PSF kernel size, oversampling, and realism of galaxy scenes.

Abstract

The Nancy Grace Roman Space Telescope will carry out a wide-field imaging and slitless spectroscopic survey of Type Ia Supernovae to improve our understanding of dark energy. Crucial to this endeavor is obtaining supernova spectra uncontaminated by light from their host galaxies. However, obtaining such spectra is made more difficult by the inherent problem in wide-field slitless spectroscopic surveys: the blending of spectra of close objects. The spectrum of a supernova will blend with the host galaxy, even from regions distant from the supernova on the sky. If not properly removed, this contamination will introduce systematic bias when the supernova spectra are later used to determine intrinsic supernova parameters and to infer the parameters of dark energy. To address this problem we developed an algorithm that makes use of the spectroscopic observations of the host galaxy at all available observatory roll angles to reconstruct a three-dimensional (3d; 2d spatial, 1d spectral) representation of the underlying host galaxy that accurately matches the 2d slitless spectrum of the host galaxy when projected to an arbitrary rotation angle. We call this ``scene reconstruction''. The projection of the reconstructed scene can be subtracted from an observation of a supernova to remove the contamination from the underlying host. Using simulated Roman data, we show that our method has extremely small systematic errors and significantly less random noise than if we subtracted a single perfectly aligned spectrum of the host obtained before or after the supernova was visible.

Figures (52)

• Figure 1: This illustration shows simulated observations of a supernova and its host galaxy, with images in five Roman filters (top panels) and the corresponding spectral image taken using the slitless prism (middle panel). The location of the supernova relative to its host galaxy is marked by the red arrows. If we extract the 1d spectrum by concentrating on the boxed area marked in red, we will obtain the SN+galaxy 1d signal-to-noise-ratio (SNR) spectrum as shown in red on the bottom-left panels. If there is no supernova, the similarly-extracted galaxy-only 1d SNR spectrum in the same area is overplotted in black on this panel. Suppose there is no host galaxy, the observed supernova-only 1d SNR spectrum is shown in red at the bottom-right panel. For comparison, the noise-free supernova-only spectrum is overplotted as the green curve. Sky background, the dominant source of noise (see Section \ref{['sec:noise']}), has been subtracted from these spectra. Note the different vertical axis ranges between the panels in the bottom row. The exposure times of these observations are 900 s and 1800 s for respectively the filter images and the spectral image. These are single images and not co-added images. To generate these data, for the host galaxy we use a datacube based on spectral energy distribution (SED) fitting of a galaxy from the VELA galaxy simulation sim19, using the CIGALE boq19 code. The galaxy is at redshift $z=1.0$. We use SNCosmosncosmo to generate the injected supernova spectrum, which is an extended SALT2 spectrum guy10bet14hou18 and redshifted to the same $z$ as the host galaxy.
• Figure 2: In this illustration, Roman observed a galaxy at seven different roll angles, whose values are shown on the top left corners in the left column of each row. The images at each roll angle are shown on the left column, while the corresponding spectra are shown on the right column. At roll angle $\phi = 195.4^\circ$ (outlined in red), a supernova occurred (marked with red arrow). We set aside this spectrum and use the other roll angles to predict what the galaxy-only spectrum would look like at roll angle $\phi = 195.4^\circ$, in order to subtract the host-galaxy spectrum and obtain the supernova-only spectrum (In later demonstrations we will use more roll angles). This is the same simulated galaxy and supernova shown in Figure \ref{['fig:example']}.
• Figure 3: An illustration of a datacube and its various products through projections. A datacube is a flux mapping of a scene of the sky in two spatial axes $(\xi,\eta)$ and the wavelength axis $\lambda$. If we slice the datacube at a constant wavelength, each slice is a monochromatic image at wavelength $\lambda$ (blue arrow). If we look through a particular spatial coordinate $(\xi,\eta)$ along the wavelength axis (red arrow), we will obtain a 1d spectrum. In imaging mode, we convolve the datacube with the corresponding PSF and the passband of a particular filter, and integrate it along the wavelength range of the filter (highlighted as the green box), obtaining a filter image as recorded by the detector. In slitless spectroscopy mode, the prism is put in the optical path rather than the filter. The datacube is convolved with the PSF and the prism passband then integrated along the prism dispersion curve (see Section \ref{['sec:datasim']} for details) to obtain a 2d spectral image. Note that the horizontal axis of the 2d spectral image is actually not an exact mapping of the wavelength axis, but rather it is a mapping of both the spatial axis $\xi$ and wavelength axis $\lambda$ (and in some cases also $\eta$) through the dispersion curve. See Section \ref{['sec:prism']}, Section \ref{['sec:Dp']}, and Figure \ref{['fig:disp']} for details. Illustration adapted from pon16, with the same datacube used in Figure \ref{['fig:example']}.
• Figure 4: The effective area of the Wide Field Instrument (WFI) filters, grism, and prism. Not shown are the F213 filter and the very wide F146 filter.
• Figure 5: The spectral sampling (black line) at the center of the field of view (FoV) of the WFI and the two-pixel resolution (red line) of the prism. The axis on the left corresponds to the black line, while the axis on the right (in red) corresponds to the red line. Note that although the spectral sampling changes by a factor of $\sim$4 across the prism bandpass, the resolution changes by only a factor $\sim$2 across the same bandpass range.