Information Graphs of Statistical Summaries
Afrika Statistika · 01.04.2025 · 20 (2) · 4411-4426
DOI: 10.16929/as/2025.4411.352 ↗Kurz gefasst
Fachartikel in der Zeitschrift Afrika Statistika, online erschienen am 01.04.2025. Verfasst von Idriss Olivier Bado.
Abstract (Originalsprache)
A statistic compresses an observation and may therefore preserve some parameter directions more faithfully than others. We study this directional loss of Fisher information in regular parametric models. The key observation is that conditional expectation given the statistic is the orthogonal projection of the score onto the subspace of functions measurable with respect to that statistic. Consequently, the Fisher information of the induced experiment is the Gram matrix of the projected score. After normalizing by the full Fisher information, we obtain a positive semidefinite contraction whose diagonal entries quantify directional retention and whose off-diagonal entries describe mixing between parameter directions. This matrix defines a weighted information graph. We establish the exact retained--lost information decomposition, identify the coordinate-free spectral content of the construction, derive closed-form expressions for linear and noisy linear summaries in Gaussian location models, and prove monotonicity under further summarization. The framework complements scalar measures of information loss by showing not only how much information remains, but also the form in which it remains
Zitieren
APA
Bado, I. O. (2025). Information Graphs of Statistical Summaries. Afrika Statistika, 20(2), 4411-4426. https://doi.org/10.16929/as/2025.4411.352
Vancouver
Bado IO. Information Graphs of Statistical Summaries. Afrika Statistika. 2025;20(2):4411-4426. doi:10.16929/as/2025.4411.352
BibTeX
@article{Bado2025,
author = {Bado, Idriss Olivier},
title = {Information Graphs of Statistical Summaries},
journal = {Afrika Statistika},
year = {2025},
volume = {20},
number = {2},
pages = {4411--4426},
doi = {10.16929/as/2025.4411.352}
}
EUVIMED