Refined Convergence Rates for Maximum Likelihood Estimation under Finite Mixture Models

Authors: Tudor Manole, Nhat Ho

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our theoretical findings are supported by a simulation study to illustrate these improved convergence rates.
Researcher Affiliation Academia 1Department of Statistics and Data Science, Carnegie Mellon University 2Department of Statistics and Data Sciences, University of Texas, Austin.
Pseudocode Yes Algorithm 1 Modified EM Algorithm.
Open Source Code Yes All code for reproducing our simulation study is publicly available.1 https://github.com/tmanole/Refined-Mixture-Rates
Open Datasets No The paper uses synthetic data generated for the simulation study rather than publicly available datasets. For each model, we generate 20 samples of size n, for 100 different choices of n between 10^2 and 10^5.
Dataset Splits No The paper describes generating synthetic data for a simulation study and does not specify explicit training, validation, or test dataset splits.
Hardware Specification No All simulations hereafter were performed in Python 3.7 on a standard Unix machine, which is not a specific hardware detail.
Software Dependencies Yes All simulations hereafter were performed in Python 3.7
Experiment Setup Yes We chose the convergence criteria ϵ = 10^-8 and T = 2,000. Since our aim is to illustrate theoretical properties of the estimator b Gn, we initialized the EM algorithm favourably. In particular, for any given k and k0, and for each replication, we randomly partitioned the set {1, . . . , k} into k0 index sets I1, . . . , Ik0, each containing at least one point. We then sampled θ(0)j (resp. Σ(0)j) from a Gaussian distribution with vanishing covariance, centered at θ0ℓ(resp. Σ0ℓ), where ℓis the unique index such that j Iℓ.