®®®® SIIA Público

Título del libro: Ieee International Symposium On Information Theory - Proceedings
Título del capítulo: Tensorization of f-Divergences

Autores UNAM:
MARIO ALBERTO DIAZ TORRES;
Autores externos:

Idioma:

Año de publicación:
2025
Palabras clave:

Differential privacy; Independent component analysis; Condition; Differential privacies; F-divergence; Formal definition; Independent components; KL-divergence; Machine-learning; Random vectors; Simple++; Tensorization; Learning systems


Resumen:

In many applications across statistics, differential privacy, and machine learning, it is necessary to evaluate or bound an f-divergence between distributions of random vectors with independent components. Even for relatively simple distributions, such computations can become intractable unless the chosen f-divergence tensorizes. In this work, we introduce a formal definition of tensorization for f-divergences and present a necessary condition under which such tensorization can occur. Moreover, we demonstrate - under certain assumptions - that the only f-divergences admitting a polynomial tensorization formula of degree at most two are, essentially, the KL divergence, cross-entropy, and Hellinger divergences of order a. Taken together, our findings represent an initial step toward a complete characterization of f-divergences that tensorize. © 2025 IEEE.


Entidades citadas de la UNAM: