Embedding Reliability for Unsupervised Classification of Gamma Ray Burst progenitors from Prompt Gamma-ray Emission

Abstract

We present a statistical method based on scDEED to assess the reliability of a 2D embedding showing a low-dimensional representation of the distribution of Gamma-Ray Bursts (GRBs) detected by the Fermi Gamma-ray Burst Monitor (GBM). The original dataset consists of 12 waterfall plots for each event, which contain key information about the prompt emission of each GRB. The dataset's dimensionality is first reduced to a 30-dimensional latent space using an autoencoder, and subsequently to 2D using UMAP. While the methodology and results are discussed in a previous work (arXiv:2406.03643), here we introduce a statistical approach to evaluate the reliability of the final 2D distribution based on the scDEED algorithm. Our analysis shows that the 2D embedding demonstrates overall good reliability, with more than 90\% of the events classified as trustworthy.

Figures (1)

• Figure 1: Left panel: Pearson correlation values for each event, computed as in Step 3, are shown for the real dataset (gray histogram) and for the permuted datasets (black histogram), which serve as the null distribution. The 95th and 5th percentile thresholds are indicated by green and red lines, respectively. Right panel: Final 2D embedded distribution of the GRB dataset. Marker colors indicate the reliability measure: green for trustworthy, gray for intermediate, and red for dubious. Circled points indicate the GRB progenitor type: BNS $\rightarrow$ Binary neutron stars merger (blue), MGF $\rightarrow$ Magnetar giant flare (yellow), LBNS $\rightarrow$ Long binary neutron stars merger (purple), CCSN $\rightarrow$ Core-collapse supernova (pink).