A useful representation of TESS light curves

TL;DR

This paper introduces an interpretable, computationally efficient representation of TESS light curves using quantile graphs or scattering transforms, enabling large-scale exploratory analysis without labels.

Contribution

The authors develop a novel, label-free mapping of TESS light curves that captures meaningful similarities and organizes sources by key astrophysical properties.

Findings

The map organizes sources by variability amplitude, signal-to-noise ratio, and shape.

Most sources occupy stable, contiguous regions indicating persistent properties.

The representation aids in exploration, anomaly detection, and dataset characterization.

Abstract

We present a simple and interpretable representation of TESS light curves designed for large-scale exploratory analysis. Our goal is not to optimize classification performance, but to construct a computationally efficient mapping in which proximity reflects meaningful similarity, without using labels or explicit period information as inputs. We represent each light curve using either quantile graphs or scattering transforms, reduce dimensionality with principal component analysis, and project the resulting features onto a self-organizing map (SOM). We evaluate ~1500 model configurations using a combination of standard embedding diagnostics and a light-curve-shape-based cohesion metric, and select a compact quantile-graph-based model that balances interpretability, stability, and performance. Applying the model to ~1.5 million TESS 2-minute cadence light curves, we find that the map organizes sources primarily by variability amplitude, signal-to-noise ratio, characteristic timescale, and light-curve shape. Repeat observations of the same stars show that most sources occupy stable and contiguous regions of the map, indicating that the representation captures persistent properties rather than noise and systematics. We provide an interactive web interface at http://tess-l8.space that enables inspection of nodes, nearest neighbors, and individual sources across sectors. The resulting representation serves as a practical tool for exploration, anomaly detection, and dataset characterization, and illustrates how simple, deterministic encodings can yield useful structure in large astronomical time-series datasets.