Classifying Gamma-Ray Bursts Using Dimensionality Reduction

The observed bimodality in the log-duration of gamma-ray bursts (GRBs) strongly suggests that they should be classified into two broad types, termed short and long bursts. Although several separations have been proposed, an unambiguous classification of every burst has thusfar not been possible. In this work we propose the use of a dimensionality reduction algorithm, t-distributed stochastic neighborhood embedding (t-SNE), to classify GRBs based upon their light curves. This method produces two distinct and clearly separated groups, so that each burst can be unambiguously classified. The two groups exhibit approximately normal distributions in log duration overlapping at approximately 2 seconds, and other properties previously identified as associated solely with short or long bursts match this proposed classification, including associated supernova. The mapping also produces substructure, suggesting that both long and short GRBs may include bursts with distinct physical origins.