🔭 astrophysics

Co-addition and Subtraction of Undersampled Images

This paper introduces LUTRA, a novel open-source method that mathematically optimizes the co-addition and background subtraction of undersampled astronomical images to achieve super-resolution, reduce false alarms, and improve signal-to-noise ratios by 25% compared to current techniques.

Original authors: Matan Schlanger, Barak Zackay

Published 2026-07-21

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Matan Schlanger, Barak Zackay

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

Imagine you are trying to take a perfect photo of a starry night sky, but your camera has a strange glitch: its pixels are too big. Instead of capturing the tiny, sharp points of light, the camera smears them out so much that a star landing on the left side of a pixel looks exactly the same as a star landing on the right side. In the world of astronomy, this is called "undersampling." It's like trying to read a fine print book through a pair of foggy, thick glasses; you know something is there, but the details are lost, and the image looks blocky and confusing.

Astronomers need these sharp images to spot "transients"—sudden, fleeting events in the universe like exploding stars (supernovae) or the afterglows of colliding black holes. To see these faint flashes, they take many pictures of the same patch of sky and stack them together, a process called "co-addition," to make the image deeper and clearer. They also need to subtract the old, static background to find what's new. But when the images are undersampled, the usual math breaks down. The "blocky" nature of the pixels makes it impossible to line up the images perfectly, leading to blurry results and a lot of false alarms where the computer thinks it sees a new star but is actually just seeing a glitch in the pixel grid.

This is where a new method called LUTRA (Linear Undersampled Transients & Addition) comes in, proposed by Matan Schlanger and Barak Zackay. Think of LUTRA as a super-smart detective that doesn't just look at the blurry photos one by one. Instead, it uses a clever mathematical trick to look at the pattern of all the blurry photos together. The authors show that even though individual images are fuzzy, the collection of them contains hidden clues about the true, sharp shape of the sky. By solving a specific set of equations, LUTRA can reconstruct a "super-resolution" image that is much sharper than the original camera could ever capture on its own.

The paper demonstrates that this method is mathematically proven to be the best possible way to handle these undersampled images. When the team tested LUTRA on real data from the Zwicky Transient Facility (ZTF), they found it produced images with a signal-to-noise ratio that was 1.25 times higher than current methods. This means it can spot fainter objects and, crucially, it significantly reduces the number of false alarms caused by the pixel grid confusion. The result is a cleaner, sharper view of the universe that allows astronomers to find new cosmic events faster and with more confidence, all without needing to build a new, more expensive camera.

Technical Summary: Co-addition and Subtraction of Undersampled Images

Problem Statement
In astronomical imaging surveys, such as those searching for transients (supernovae, gravitational wave counterparts), repeated observations of the same sky patches are required to achieve deep imaging. A significant portion of these images, particularly from space-based observatories (e.g., JWST, Roman, Spitzer) and ground-based surveys under excellent seeing conditions (e.g., ZTF, VRO), are "undersampled." This occurs when the pixel size is too large relative to the Point Spread Function (PSF), violating the Nyquist-Shannon sampling theorem (typically defined as FWHM < 2 pixels).

Undersampling leads to aliasing, where sub-pixel shifts of a source result in identical pixel values, making image registration and sub-pixel shifts ill-defined. Current methods for handling these images, such as the Drizzle algorithm or simple arithmetic subtraction, are sub-optimal. They often fail to account for varying PSFs, introduce noise correlations, require manual intervention to distinguish artifacts from real signals, or suffer from numerical instability and high computational costs. These limitations result in reduced sensitivity and increased rates of false alarms, hindering the efficient detection of transient phenomena.

Methodology
The authors propose a new method named Linear Undersampled Transients & Addition (LUTRA). This approach provides a mathematically proven optimal framework for co-adding undersampled images and performing background subtraction.

The core of the methodology involves modeling the image formation process in the Fourier domain. The authors define a general model where the true sky $T(x)$ is convolved with the image-specific PSF $P_j$ and shifted by a sub-pixel amount $\Delta_j$ , then sampled with additive white Gaussian noise.

Matrix Formulation: The sampling process is expressed as a linear system in the frequency domain. For each frequency index $k$ , the sampled data vector $\vec{M}_k$ is related to the high-resolution sky vector $\vec{T}_k$ via a system matrix $G_k$ (incorporating PSF and phase shifts) and noise.
Sufficient Statistics: By applying the Fisher-Neyman factorization theorem to the log-likelihood function, the authors derive "sufficient statistics" that retain all information present in the original dataset regarding the static background. These statistics are:
1. Co-addition Statistic: A summary of the reference images that allows for optimal likelihood calculation for any hypothesis without re-processing the raw data.
2. Subtraction Statistic: A derived statistic for detecting new point sources by comparing a new image set against the reference set. This statistic utilizes the sufficient statistics of both sets, enabling efficient computation without loading entire datasets into memory.
3. Cosmic Ray Statistic: A specific statistic for identifying sharp, high-amplitude artifacts (cosmic rays) in new images.

The method is designed to be numerically stable (avoiding the inversion of singular matrices common in other reconstruction algorithms) and updatable, allowing new images to be incorporated without recomputing the entire solution.

Key Contributions

Optimal Co-addition and Subtraction: LUTRA provides the first mathematically proven optimal solution for co-adding and subtracting undersampled images, accounting for varying PSFs and noise levels.
Super-Resolution Output: The derived statistics are vectors in the frequency domain that can be "unfolded" to produce images with a resolution improvement factor of $B$ (the undersampling ratio), effectively achieving super-resolution.
Direct Measurement Capabilities: The method facilitates direct and optimal photometry and astrometry measurements. The authors demonstrate that flux estimation and precise position determination can be performed efficiently using the derived statistics.
Robustness to Noise and PSF Variations: Unlike previous methods, LUTRA explicitly weights images based on noise levels and handles images with different PSFs and shifts naturally.
Open Source Implementation: The authors provide an open-source Python implementation of the algorithm.

Results
The authors demonstrate the performance of LUTRA using public data from the Zwicky Transient Facility (ZTF).

Signal-to-Noise Ratio (SNR): The method achieves a 1.25x higher SNR compared to current standard methods (specifically those involving Drizzle and simple subtraction).
False Alarm Reduction: Analysis of ZTF event data indicates that a significant proportion of "bogus" (false) events occur in undersampled images (approx. 42% overall, rising to >54% when filtering near bright stars), whereas real transients are less affected. The authors argue that applying LUTRA to these images will directly improve the false-alarm rate.
Validation: The method is shown to reproduce results for properly sampled images (reducing to known optimal scalar solutions) and correctly handles interlaced sampling scenarios, yielding results equivalent to high-resolution signals.

Significance and Claims
The paper claims that LUTRA solves the fundamental problem of transient detection in undersampled images by providing an optimal, linear, and computationally efficient framework. The significance lies in its ability to:

Improve the sensitivity of existing and future surveys (including the Vera C. Rubin Observatory and Roman Space Telescope) where undersampling is prevalent.
Reduce the reliance on human supervision for transient classification by minimizing false positives caused by undersampling artifacts.
Enable precise photometry and astrometry in regimes where traditional methods fail or lose information.

The authors note that while the current work assumes aligned images without relative rotation, the framework is designed to be extended to handle rotations in future companion papers. The method is presented as a comprehensive solution for synthesizing undersampled image sets into high-resolution objects, directly addressing the limitations of current pipelines in the era of large-scale time-domain astronomy.

More like this