Entropic regularisation of unbalanced optimal transportation problems

Maciej Buze,Manh Hong Duong

2023-05-04

Abstract:We develop a mathematical theory of entropic regularisation of unbalanced optimal transport problems. Focusing on static formulation and relying on the formalism developed for the unregularised case, we show that unbalanced optimal transport problems can be regularised in two qualitatively distinct ways - either on the original space or on the extended space. We derive several reformulations of the two regularised problems and in particular introduce the idea of a regularised induced marginal perspective cost function allowing us to derive an extended space formulation of the original space regularisation. We also prove convergence to the unregularised problem in the case of the extended space regularisation and discuss on-going work on deriving a unified framework based on higher order liftings in which both regularisations can be directly compared. We also briefly touch upon how these concepts translate to the corresponding dynamic formulations and provide evidence why the extended space regularisation should be preferred. This is a preliminary version of the manuscript, to be updated in the near future.

Optimization and Control

What problem does this paper attempt to address?

### What problem does this paper attempt to solve?

This paper aims to study and develop the entropy - regularization theory of the Unbalanced Optimal Transportation (UOT). Specifically, the author explores two different entropy - regularization methods:

1. **Regularization on the original space**: Regularize the UOT problem on the original space \(X\).
2. **Regularization on the extended space**: Regularize the UOT problem on the extended space \(Y\).

#### Main problem description

The standard balanced optimal transportation problem requires that the masses of the two input measures are equal, that is, \(\mu_0(X)=\mu_1(X)\), otherwise the problem has no solution. However, in practical applications, this assumption is often too strict. Therefore, the unbalanced optimal transportation problem is introduced, which allows handling the transportation problem between measures of different masses.

#### The role of entropy regularization

Entropy regularization smoothes the optimization problem by adding an entropy term, making the problem easier to solve and enabling more efficient algorithms to be introduced. For example, in the standard optimal transportation problem, entropy regularization is usually used to introduce convexity, thereby simplifying numerical calculations.

#### Main contributions of the paper

1. **Theoretical development**: Based on the existing un - regularized UOT theory, the author proposes two different entropy - regularization methods and proves their equivalent forms.
2. **Formula derivation**: The expression of the regularized induced marginal - perspective cost function \(H_{\varepsilon}\) is derived in detail:
\[
H_{\varepsilon}(x_0, s_0, x_1, s_1, S)=s_0 + s_1+\varepsilon S-(2 + \varepsilon)\left(\left(s_0s_1\right)^{\frac{1}{2+\varepsilon}}S^{\frac{\varepsilon}{2+\varepsilon}}\exp\left(-\frac{c(x_0, x_1)}{2+\varepsilon}\right)\right)
\]
3. **Convergence analysis**: It is proved that the regularization problem on the extended space converges to the original UOT problem when the regularization parameter \(\varepsilon\rightarrow0\).
4. **Dynamic formula comparison**: Briefly discuss the differences between these two regularization methods in the dynamic formula and provide evidence that the regularization on the extended space has a faster asymptotic convergence rate when \(\varepsilon\rightarrow0\).

#### Conclusion

By introducing entropy regularization, this paper provides a more flexible and practical framework for the unbalanced optimal transportation problem, which helps to solve complex transportation problems in practical applications and lays a theoretical foundation for further research.