Simulating Image Coaddition with the Nancy Grace Roman Space Telescope. IV. Hyperparameter Optimization and Experimental Features

TL;DR

The paper tackles image coaddition for the Nancy Grace Roman Space Telescope's weak-lensing program, using Imcom to optimize coadded PSFs by balancing PSF leakage $U_α$ and noise amplification $Σ_α$. It systematically explores hyperparameters (the Lagrange multiplier $κ_α$, the iterative solver tolerance $rtol$, and the acceptance radius $INPAD$) across five bands and 16 blocks drawn from OpenUniverse2024 simulations, evaluating 12 criteria including internal Imcom diagnostics, coadded-noise properties, and injected-star measurements. The authors find that the Cholesky kernel yields the best PSF fidelity and measurement accuracy, Gaussian target PSFs outperform Airy-based targets for photometric reliability, and many experimental features offer little or negative benefit. The results inform practical coaddition choices for Roman, with guidance for future Imcom enhancements and plans to apply these methods to shear and clustering measurements.

Abstract

For weak gravitational lensing cosmology with the forthcoming Nancy Grace Roman Space Telescope, image coaddition, or construction of oversampled images from undersampled ones, is a critical step in the image processing pipeline. In the previous papers in this series, we have re-implemented the {\sc Imcom} algorithm, which offers control over point spread functions in coadded images, and applied it to state-of-the-art image simulations for Roman. In this work, we systematically investigate the impact of {\sc Imcom} hyperparameters on the quality of measurement results. We re-coadd the same $16$ blocks ($1.75 \times 1.75 \,{\rm arcmin}^2$, $2688 \times 2688$ pixels each) from OpenUniverse2024 simulations with $26$ different configurations in each of $5$ bands. We then compare the results in terms of $12$ objective evaluation criteria, including internal diagnostics of {\sc Imcom}, properties of coadded noise frames, measurements of injected point sources, and time consumption. We demonstrate that: i) the Cholesky kernel is the best known linear algebra strategy for {\sc Imcom}, ii) for our measurements, a wide Gaussian target output PSF outperforms a smoothed Airy disk or a narrow Gaussian, iii) kernel-specific settings are worth considering for future coaddition, and iv) {\sc Imcom} experimental features studied in this work are either inconsequential or detrimental. We end this paper by discussing current and next steps of {\sc Imcom}-related studies in the context of Roman shear and clustering measurements.

Figures (10)

• Figure 1: Diagram showing the suite of $16$ test blocks ($1.75 \times 1.75 \,{\rm arcmin}^2$ each, including padding) selected from $36 \times 36$ blocks in the mosaic ($1.0 \times 1.0 \,{\rm deg}^2$). All cases studied in this work, benchmark or variant, are tested on these $16$ blocks.
• Figure 2: Target output PSFs in the H158 band studied in this work. The normalization of PSFs is arbitrary but unified. The upper row shows three PSFs of the same FWHM but different forms; from left to right: Gaussian PSF in the benchmark case, obscured Airy disk convolved with a Gaussian, and unobscured Airy disk convolved with a Gaussian. Parameters of these three PSFs are tabulated in Table \ref{['tab:extrasmooth']}. The left two panels of the lower row show Gaussian PSFs that are narrower and wider (in terms of FWHM) than the benchmark version by $20\%$, respectively. The lower right panel compares the radial profiles of these PSFs, shown as solid curves with colors corresponding to preceding panels. The half widths at half maximum (HWHMs) are shown as white dashed circles in the first five panels and colored dashed vertical lines in the last one.
• Figure 3: Four layers in a field of $17.5 \,{\rm arcsec}$ ($448$ output pixels) on a side, coadded by the Cholesky kernel (upper row) and the iterative kernel (lower row). Each panel is a Y106 ( #001AA6) + J129 ( #006659) + H158 ( #596600) + F184 ( #A61A00) composite; note that these four colors have similar lightnesses and add up to white (#FFFFFF). From left column to right column, the four layers are: simulated science images ( ' SCI'), injected stars drawn by GalSim ( ' gsstar14'), simulated white noise frames ( ' whitenoise10'), and simulated $1/f$ noise frames ( ' 1fnoise9'). The scaling is set following Paper I Figure 8 for ' SCI' and following Paper II Figure 1 for the other three layers.
• Figure 4: Output maps in the H158 band produced by the Cholesky kernel (upper row) and the iterative kernel (lower row). From left to right: fidelity (negative logarithmic PSF leakage in decibels, i.e., $-10 \log_{10} (U_\alpha/C)$), noise amplification (Equation (\ref{['eq:U_Sigma']})), effective coverage (Equation (\ref{['eq:Neff']})), and total weight (Equation (\ref{['eq:Tsum']})). Note that we deliberately choose different color bar ranges to better display spatial structures.
• Figure 5: Histograms of $957$ sets of four Imcom diagnostics yielded by two linear algebra kernels in five bands. (From left to right:) Mean fidelity is defined as $-10\log_{10} \langle U_\alpha/C\rangle$, where $U_\alpha$ is the PSF leakage metric defined in Equation (\ref{['eq:U_Sigma']}), and $\langle\cdot\rangle$ denotes an average over $15 \times 15$ pixels centered at a HEALPix node with ${\tt NSIDE} = 14$. Logarithmic mean noise amplification is defined as $\log_{10} \langle \Sigma_\alpha\rangle$, where $\Sigma_\alpha$ is the noise amplification metric defined in Equation (\ref{['eq:U_Sigma']}). Logarithmic mean effective coverage is defined as $\log_{10} \langle \bar{n}_{{\rm eff}, \alpha}\rangle$, where $\bar{n}_{{\rm eff}, \alpha}$ is the effective coverage metric defined in Equation (\ref{['eq:Neff']}). Logarithmic standard deviation of total weight is defined as $\log_{10} \sigma[T_{{\rm tot}, \alpha}]$, where $T_{{\rm tot}, \alpha}$ is the total input weight metric defined in Equation (\ref{['eq:Tsum']}), and $\sigma[\cdot]$ denotes a standard deviation within $15 \times 15$ pixels centered at a HEALPix node with ${\tt NSIDE} = 14$. Following Paper III, we invert $x$-axes of the second and fourth columns so that "better" values are shown on the right; but unlike in Paper III, we do not explicitly introduce minus signs here. From top to bottom, the five rows present histograms in Y106, J129, H158, F184, and K213 bands; results given by the Cholesky and iterative kernels are shown in blue and orange, respectively.