Andrew Landgraf
November 18, 2014
Goal minimize \( f(x) \)
Use surrogate function \( g(x|x_m) \), such that
\[ g(x_m | x_m) = f(x_m) \]
and
\[ g(x | x_m) \geq f(x) \]
Let
\[ M(x_m) = x_{m+1} = \text{argmin} ~ g(x |x_m) \]
The local and global convergence properties of [MM] exactly parallel the corresponding properties of the EM and EM gradient algorithms. This is hardly surprising because the relevant theory relies entirely on optimization transfer and never mentions missing data.
– “Optimization Transfer Using Surrogate Objective Functions” by Lange, Hunter, and Yang (2000)
From “On the convergence properties of the EM algorithm” by Wu (1983)
Under certain conditions: