EM is cast as a relaxation flow whose local dynamics are controlled by the spectrum of the missing-information operator G, from which two new accelerators (G-Accelerator and Geo-Adaptive) are derived and shown to outperform standard EM.
Blahut–Arimoto algorithms as mirror descent
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
stat.ML 1years
2026 1verdicts
UNVERDICTED 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Expectation-Maximization as a Spectrally Governed Relaxation Flow
EM is cast as a relaxation flow whose local dynamics are controlled by the spectrum of the missing-information operator G, from which two new accelerators (G-Accelerator and Geo-Adaptive) are derived and shown to outperform standard EM.