🔭 astrophysics

A New PSF Deconvolution Algorithm: Simultaneous Spatial Resolution Enhancement and Point Source Removal for Morphological Analysis of AGN Host Galaxies

The authors propose a novel PSF deconvolution algorithm that simultaneously enhances the spatial resolution of AGN host galaxies and removes their bright central point sources by decomposing images into smooth extended and sparse point-source components under specific constraints, demonstrating its effectiveness on Subaru Telescope data to enable future statistical morphological studies with upcoming wide-field surveys.

Original authors: Ren Kawase, Takatoshi Shibuya, Kazunori Matsuda

Published 2026-05-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ren Kawase, Takatoshi Shibuya, Kazunori Matsuda

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: A Blurry Photo with a Blinding Flash

Imagine you are trying to take a picture of a beautiful, intricate castle (a galaxy) at night. However, right in the center of the castle, someone has turned on an incredibly bright, blinding flashlight (the Active Galactic Nucleus, or AGN).

When you take the photo with a standard camera (a ground-based telescope like Subaru), two things happen:

The Flash: The light from the flashlight is so bright it washes out the details of the castle walls right next to it.
The Blur: Because the atmosphere acts like a foggy window, the whole picture is a bit soft and blurry. You can't see the fine bricks or the spiral towers clearly.

Astronomers want to study the shape and structure of these "castles" to understand how they grow. But the blinding flash and the blur make it impossible to see the details. Traditional methods try to guess what the castle looks like by mathematically subtracting the flash, but they often end up either leaving too much glare (under-subtraction) or erasing parts of the castle along with the flash (over-subtraction).

The Solution: A New "Digital Magic Trick"

The authors of this paper developed a new computer algorithm (a set of mathematical rules) that acts like a super-smart photo editor. Instead of just trying to fix the whole picture at once, it splits the job into three specific tasks simultaneously:

The "Smooth" Filter (The Castle): It creates a version of the image that looks for big, smooth, flowing shapes. It assumes the galaxy is like a soft, rolling hill or a smooth fabric. This helps it ignore the sharp, tiny points of light.
The "Sparse" Filter (The Flash): It creates a separate version that looks for tiny, sharp, isolated points. It assumes the AGN is like a single, bright star. This forces the algorithm to put all the "flash" light into this specific bucket and nowhere else.
The "Balance" Rule (The Handshake): This is the new, clever part. The algorithm has a rule that says: "The smooth castle and the sharp flash cannot occupy the exact same pixel at the same time." If the "flash" bucket claims a pixel is bright, the "castle" bucket must admit that pixel is empty. This prevents the algorithm from accidentally keeping the glare on the castle or erasing the center of the castle while trying to remove the glare.

How They Tested It

To see if their magic trick worked, they did two things:

The Fake Test: They built a computer-generated galaxy with a fake flashlight in the middle. They knew exactly what the "true" picture should look like. They ran their algorithm on the blurry, fake photo, and it successfully separated the smooth castle from the sharp flash, revealing a clear, high-definition castle that looked just like the original blueprint.
The Real Test: They took real photos of actual galaxies from the Subaru Telescope (which are blurry and have bright centers) and compared them to photos of the same galaxies taken by the Hubble Space Telescope (which are naturally sharper and taken from space).

The Results: Ground-Based Telescopes Meet Space Telescopes

The results were impressive. By using their new algorithm:

The Flash Disappeared: The bright central point source was successfully removed from the galaxy image without leaving a "ghost" or erasing the galaxy's core.
The Blur Vanished: The remaining image of the galaxy became much sharper. In fact, the ground-based images processed by their algorithm became almost as sharp as the Hubble Space Telescope images.
Details Emerged: Structures that were previously hidden, like spiral arms and clumps of stars, became clearly visible.

Why This Matters

The paper claims that this method allows astronomers to study the shapes of distant galaxies using data from wide-field surveys (like those from the Subaru Telescope) without needing to rely on the expensive and limited view of space telescopes for every single object.

Think of it this way: Before this, if you wanted to see the fine details of a castle, you needed a very expensive, high-powered telescope in space. Now, with this new "digital magic trick," you can take a photo with a standard telescope on the ground, run it through this algorithm, and get a result that rivals the space telescope. This opens the door to studying thousands of these galaxies at once, rather than just a handful.

Technical Summary: A New PSF Deconvolution Algorithm for AGN Host Galaxy Analysis

Problem Statement
Active Galactic Nuclei (AGN) host galaxies are often difficult to study morphologically because the bright central point source (the AGN) contaminates the light from the host galaxy. While high-resolution space-based telescopes like Hubble and JWST can resolve these structures, their narrow fields of view limit statistical studies to small samples ( $\lesssim 10^4$ objects). Ground-based surveys (e.g., Subaru/HSC, Rubin, Euclid, Roman) offer large samples but suffer from lower spatial resolution and significant Point Spread Function (PSF) blurring. Traditional methods, such as 2D image decomposition (e.g., GALFIT), are sensitive to PSF over-subtraction and complex substructures, and they struggle to separate the host galaxy from the AGN without introducing uncertainties in structural parameters like the Sérsic index. Existing PSF deconvolution algorithms often fail to simultaneously output a separated extended component, a separated point-source component, and a combined reconstructed image, which are all necessary for robust morphological and photometric analysis of faint, high-redshift AGN hosts.

Methodology
The authors propose a new PSF deconvolution algorithm designed to simultaneously enhance spatial resolution and remove the central point source. The method models the intrinsic image $x$ as the sum of two distinct components: an extended component $I_{sm}$ (the host galaxy) and a point-source component $I_{sp}$ (the AGN). The observation model assumes the observed flux follows a Gaussian distribution based on the convolution of these components with the PSF.

To reconstruct the intrinsic image, the algorithm minimizes a negative log-posterior objective function subject to three specific constraints:

Smooth Constraint ( $I_{sm}$ ): A Tikhonov-like regularization applied to the extended component to enforce spatial smoothness, modeling the host galaxy structure.
Sparse Constraint ( $I_{sp}$ ): An $L_1$ norm (LASSO) regularization applied to the point-source component to localize the AGN flux to a small number of pixels.
Point-Source Balance Constraint: A novel constraint based on the pixel-wise product $I_{sm} \times I_{sp}$ . This term penalizes the simultaneous presence of high flux values in both the extended and point-source components at the same pixel location. This prevents the "leakage" of point-source flux into the host galaxy (over-subtraction) or the suppression of the host galaxy's central flux (under-subtraction).

The optimization is solved using the Iterative Shrinkage-Thresholding Algorithm (ISTA) with a backtracking scheme for the Lipschitz constant to ensure stable, monotonic convergence. The algorithm outputs three images: the separated host galaxy ( $I_{sm}$ ), the separated AGN ( $I_{sp}$ ), and the combined reconstruction ( $I_{sm} + I_{sp}$ ).

Key Contributions

Novel Algorithm: Development of a three-output PSF deconvolution method specifically tailored for AGN images, explicitly separating the host galaxy and the AGN while reconstructing the total image.
Point-Source Balance Constraint: Introduction of a new regularization term ( $I_{sm} \times I_{sp}$ ) that effectively decouples the extended and point-source components, addressing the interdependence of hyperparameters that plagues previous methods.
Ground-to-Space Resolution Enhancement: Demonstration that ground-based images (Subaru/HSC) can be deconvolved to achieve spatial resolution comparable to Hubble Space Telescope (HST) images while removing the AGN contamination.

Results
The algorithm was tested on both artificial AGN images and real AGN data from the Subaru/HSC COSMOS field (objects J0959 at $z=0.34$ and J1000 at $z=1.38$ ), using HST/ACS images as ground truth.

Qualitative Analysis: The reconstructed host galaxy images ( $I_{sm}$ ) show significantly enhanced spatial resolution, revealing fine structures like spiral arms and clumpy substructures with high contrast. The central point source is effectively removed without significant over- or under-subtraction. The combined reconstructed images ( $I_{sm} + I_{sp}$ ) closely resemble the Hubble reference images.
Quantitative Analysis:
- Convergence: The Hellinger distance between iterations decreased rapidly, stabilizing after ~100 iterations.
- Accuracy: The Root-Mean-Square Error (RMSE) between the reconstructed image (convolved with HSC PSF) and the original HSC observation was extremely low ( $\sim 10^{-6}$ to $10^{-5}$ ). The RMSE between the reconstructed image and the Hubble ground truth was also small ( $\sim 10^{-2}$ ).
- Resolution Improvement: The Full Width at Half Maximum (FWHM) of the reconstructed host galaxy structures improved by a factor of $\sim 2-3$ compared to the original HSC images, reaching levels comparable to Hubble (e.g., from $1.06''$ to $0.57''$ for J0959).
- Constraint Efficacy: The inclusion of the point-source balance constraint ( $\alpha_{bl}$ ) was shown to be critical for real AGNs, further reducing FWHM values and improving point-source removal compared to runs without this constraint.

Significance
The paper claims that this algorithm enables statistical morphological studies of distant AGN host galaxies using wide-field survey data from current and future facilities, including the Vera C. Rubin Observatory, the Euclid Space Telescope, and the Roman Space Telescope. By effectively removing AGN contamination and enhancing resolution, the method allows for the application of non-parametric morphological indices (e.g., concentration, asymmetry, Gini coefficient) to large samples ( $>10^4$ objects) without the biases previously caused by AGN flux. This capability is expected to resolve the current ambiguity regarding the relationship between AGN activity and galaxy mergers/interactions. Additionally, the method reduces reliance on visual classification and can be extended to studies of supermassive black hole–bulge relations and the extraction of other structures like giant stellar clumps.