Under Construction

Visualized examples from Dupui's and Mao's and Birrell's papers.

• Dirac Masses: Example 3.1, in page 13
• Gaussian: Theorem 6.8, in page 114
• Unifrom: Example 1, section (4.1) in page 42.

Here we consider a simple example involving Dirac masses where the (f, Γ)-divergence can be explicitly computed using Theorem 3.3. This example further illustrates the two-stage mass-redistribution/mass-transport interpretation of the infimal convolution formula and demonstrates how the location and distribution of probability mass impacts the result.

Case 1	Case 2

1) P ~ N(1, 0.5), Q ~ N(3, 0.5)	2) P ~ N(1, 0.5), Q ~ N(3, 1)	3) P ~ N(1, 1), Q ~ N(3, 0.5)

For given c, b solves: exp(1 − b) − (1 − b) = 1 + c
When c = 0, b = 1 solves the equation. When c > 0 is small, 1 − b is also close to 0. Moreover, 1 − b can be written as an analytic function of √c around 0 as 1-b = √2√c - c/3 + O(c^(3/2)).

To reproduce the achieved results run the file of your choice with the corresponding arguments. The output is a gif visualizing the Mass Redistribution and Transport respectivly to the chosen case (Dirac, Gaussian or Uniform).

Note: Dirac case 1 has no arguments.

Note: Default values corresponds to Case 2.

Note: c value must be positive and less than e-2, thus 1-b ≥ 0.

@inproceedings{Dupuis2020,
  author    = {"Paul Dupuis, Yixiang Mao"},
  title     = {"Formulation and properties of a divergence used to compare probability measures without absolute continuity and its application to uncertainty quantification"},
  journal   = {arXiv:2011.08441},
  year      = {2020},
  publisher = {"-?-"}
}

@inproceedings{Birrell2022,
  author    = {"Jeremiah Birrell, Paul Dupuis, Markos A. Katsoulakis, Yannis Pantazis, Luc Rey-Bellet"},
  title     = {"(f, Γ)-DIVERGENCES: INTERPOLATING BETWEEN, f -DIVERGENCES AND INTEGRAL PROBABILITY METRICS"},
  journal   = {arXiv preprint arXiv:2011.05953},
  year      = {2022},
  publisher = {"Journal of Machine Learning Research"}
}