Data Compression
Papers
p-MOPED data compression under noisy covariance
In Sugiyama and Park 2025, we developed powered MOPED, or \(p\)-MOPED, a data-compression method designed for parameter inference when the covariance matrix is estimated from a limited number of simulations. Standard MOPED preserves Fisher information in idealized settings, but it can still propagate noise from a noisy sample covariance into parameter constraints. The \(p\)-MOPED method suppresses this covariance-noise propagation by applying a tunable power-law transformation to the sample correlation matrix, balancing information retention against noise reduction. We tested the method on toy models and on Subaru Hyper Suprime-Cam Year 3 weak-lensing data, showing that \(p\)-MOPED gives more robust compressed likelihood analyses, especially when the number of simulations is limited.