🔭 astrophysics

A useful representation of TESS light curves

This paper presents a computationally efficient, unsupervised representation of TESS light curves using quantile graphs and self-organizing maps to organize ~1.5 million sources by physical properties like amplitude and timescale, providing a stable, interpretable tool for large-scale astronomical exploration and anomaly detection.

Original authors: Dovi Poznanski

Published 2026-05-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dovi Poznanski

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 have a library containing 1.5 million books, but instead of words, every page is a squiggly line representing how a star's brightness changes over time. These lines are called "light curves." Trying to read, compare, or find specific patterns in 1.5 million of these squiggly lines is like trying to find a specific needle in a haystack the size of a city.

This paper presents a new, simple way to organize that massive library so humans can actually explore it.

The Problem: Too Much Data, Too Little Time

Astronomers have a telescope called TESS that watches the sky and records how stars flicker. It has collected data on millions of stars. Traditionally, scientists try to sort these stars by guessing what makes them special (like how fast they pulse or how bright they get). But this is slow, requires human guesswork, and often misses interesting patterns that don't fit into neat boxes.

The Solution: A "Star Map"

The author, Dovi Poznanski, built a system that acts like a geographic map for star behavior. Instead of looking at the raw squiggly lines, the system translates them into a simple, 13-by-13 grid (169 squares total).

Here is how the system works, step-by-step:

1. Taking a Snapshot (The Input)
The system grabs a 4096-point chunk of a star's light curve. It smooths out the long-term drifts (like a slowly rising tide) so it can focus on the interesting wiggles.

2. Turning Lines into Graphs (The Translation)
The system uses a clever trick called a Quantile Graph. Imagine you have a ruler and you divide the star's brightness into 15 "zones" (from very dim to very bright).

If the star stays in the "dim" zone, the graph draws a line from that zone to itself.
If the star jumps from "dim" to "bright," the graph draws a line connecting those two zones.
By counting how often the star jumps between these zones, the system creates a unique "fingerprint" for that star's shape.

3. Compressing the Fingerprint (The Reduction)
These fingerprints are still too big to put on a map. The system uses a mathematical tool (PCA) to squish the information down, keeping only the most important details, much like summarizing a long novel into a one-page plot summary.

4. Placing it on the Map (The Sorting)
Finally, the system uses a Self-Organizing Map (SOM). Think of this as a smart sorting machine. It takes the summarized fingerprints and places them onto the 169-square grid.

Stars that look similar (similar shapes, similar brightness changes) land in the same square or neighboring squares.
Stars that look different land far away.

What Did the Map Reveal?

When the author ran all 1.5 million stars through this system, the map organized itself in a very logical way without being told what to look for:

The Corners: The top-right and bottom-left corners of the map became the "quiet zones." These held stars that were either very noisy (hard to see clearly) or just didn't change much at all.
The Center: The middle of the map held the most interesting, active stars.
The Pattern: The map naturally sorted stars by how much they varied, how clear the signal was, and how fast they pulsed.

The "Repeat Observation" Test

To make sure the map wasn't just random noise, the author checked stars that were observed multiple times (like taking a photo of the same person on different days).

The Result: Most stars landed in the same square (or a very nearby one) every time they were observed. This proved the map is stable and captures the true nature of the star, not just random glitches.
The Exception: If a star suddenly changed its behavior (like a flare or a temporary event), it would jump to a different square. This makes the map a great tool for spotting "anomalies" or weird events.

The Interactive Tool

The author didn't just stop at the math; they built a web interface.

You can click on any square of the map.
It shows you the "average" star for that square (the "medoid").
You can see the actual light curves of the stars in that square.
You can search for a specific star by its ID and see exactly where it lives on the map and how it compares to its neighbors.

The Bottom Line

This paper doesn't claim to have solved the mystery of every star. Instead, it offers a practical, fast, and interpretable way to navigate a massive dataset. It turns a chaotic pile of 1.5 million squiggly lines into a structured map where "nearby" means "similar," allowing astronomers to explore, find weird outliers, and understand the data without needing to know the answer before they start looking.

Technical Summary: A Useful Representation of TESS Light Curves

Problem Statement
Modern time-domain astronomy faces a bottleneck not in data acquisition, but in the ability to organize, compare, and interpret light curves at scale. While surveys like TESS provide millions of high-cadence light curves, traditional approaches relying on explicit feature engineering (periods, amplitudes, Fourier coefficients) impose strong priors on variability and often miss structure not captured by predefined descriptors. Conversely, deep learning approaches often lack interpretability and require fixed taxonomies. The authors seek a representation that is computationally efficient, interpretable, and capable of mapping raw, unfolded TESS light curves into a low-dimensional space where proximity reflects genuine morphological similarity, without using labels or explicit period information as inputs.

Methodology
The authors propose a pipeline that transforms raw light curves into a discrete, topology-preserving map using a combination of deterministic feature extraction, dimensionality reduction, and unsupervised clustering.

Data Preprocessing: The study utilizes the TESS 2-minute cadence PDCSAP light curves. The authors impose a continuity criterion, selecting 4096-point contiguous segments (with limited gaps) to satisfy the requirements of Scattering Transforms. Light curves are detrended by removing a linear fit through 5-sigma clipped binned means, resulting in a normalized vector with a mean flux of one.
Feature Extraction: Two families of representations are evaluated:
- Quantile Graphs (QGs): The flux range is discretized into $q$ quantiles. Transitions between quantiles over specific time steps ( $k$ ) are counted to form adjacency matrices. This method is deterministic, requires no training, and scales linearly with data length.
- Scattering Transforms (STs): A wavelet-based method (using Morlet wavelets) that generates a hierarchy of coefficients capturing multi-scale temporal structure while suppressing phase sensitivity.
Dimensionality Reduction: Principal Component Analysis (PCA) is applied to the high-dimensional feature vectors (either QG or ST), retaining the top 200 components to reduce noise and accelerate training.
Mapping to Self-Organizing Maps (SOM): The reduced features are projected onto a $13 \times 13$ hexagonal SOM grid (169 nodes). The SOM organizes the data such that nearby nodes represent similar light curve morphologies.
Model Selection: Approximately 1,500 model configurations were evaluated using a combination of standard embedding diagnostics (Quantization Error, Topological Error, Trustworthiness, Continuity) and a novel Shape Cohesion metric. The Shape Cohesion metric measures the median cyclic cross-correlation distance among period-folded light curves assigned to the same node, favoring maps where nodes contain morphologically similar sources.

Key Results

Selected Model: The optimal configuration utilizes Quantile Graphs with 15 quantiles and 8 skip values, followed by PCA reduction and a $13 \times 13$ SOM. This model was selected for its balance of stability, interpretability, and performance across all evaluation metrics.
Dataset Application: The pipeline was applied to $\sim 1.5$ million TESS 2-minute cadence light curves. The resulting map organizes sources primarily by variability amplitude, signal-to-noise ratio (SNR), characteristic timescale (period), and light-curve shape.
Stability and Persistence: Analysis of repeat observations (stars observed in multiple sectors) reveals that most sources occupy stable, contiguous regions of the map. The ratio of distinct nodes visited to the number of sectors observed is close to the ideal limit of 1, indicating the representation captures persistent astrophysical properties rather than noise or systematics.
Anomaly Detection: The distance of a light curve from its Best Matching Unit (BMU) within a node ( $D_{bmu}$ ) serves as a useful indicator of membership quality. High $D_{bmu}$ values correlate with lower visual purity, allowing for the identification of outliers or transient phenomena that cause a source to shift between nodes across sectors.
Web Interface: An interactive tool (tess-l8) was developed to allow users to inspect nodes, view medoid light curves (representative templates), browse nearest neighbors, and track individual stars across sectors.

Significance and Claims
The authors position this work not as a replacement for existing classification catalogs or period-finding pipelines, nor as a universal latent space for all astronomical time series. Instead, the significance lies in providing a simple, transparent, and computationally efficient mapping that preserves meaningful structure in large datasets.

The paper claims that:

Deterministic encodings (specifically Quantile Graphs) can yield useful, interpretable structure in large-scale time-series data without the need for training complex neural networks.
The resulting representation serves as a practical tool for exploratory analysis, quality control, and anomaly detection, enabling human-guided navigation of massive light-curve samples.
The map successfully captures persistent properties (amplitude, timescale, shape) while remaining robust to noise, as evidenced by the stability of repeat observations.
The approach demonstrates that a low-dimensional, discrete organization can effectively summarize the diversity of TESS variability, facilitating the construction of training sets and the discovery of unusual sources without relying on pre-existing labels.