Principal Component‐Based Classification of Blazar Candidates of Uncertain Type in the 4FGL / LAT

ABSTRACT

We adopted principal component analysis (PCA), linear discriminant analysis (LDA), and support vector machines (SVM), to study classification for a sample of 2708 Fermi blazars: 1142 BL Lac objects (BLLs), 759 flat spectrum radio quasars (FSRQs), and 807 blazar candidates of uncertain type (BCUs). From nine observables, the first three PCA components explained of the variance, with the main differences among blazars linked to spectral shape, luminosity, and synchrotron dominance. We introduced a dimensionless separability metric to quantify and compare the ability of different observable to distinguish BLLs from FSRQs. Using known BLLs and FSRQs, we trained an LDA model that achieved accuracy () and yielded 339 BLLs and 468 FSRQ candidates for the 807 BCUs, while the SVM model achieved 83% accuracy, yielding 350 BLL and 384 FSRQ candidates, with 73 identified as low‐confidence cases. The two models agreed on 81.5% of the BCUs, with most differences occurring near classification boundaries. Combining LDA and SVM improves classification reliability: LDA reveals the global separation pattern, while SVM supports the linearity of LDA. These results indicate that the BLL‐FSRQ distinction reflects genuine physical differences and provide a robust approach for identifying uncertain sources.