Sinkhorn's Algorithm

Sinkhorn's Algorithm is an iterative numerical method used to obtain an optimal transport plan $\pi \in \Gamma (\alpha ,\beta )$ for the Kantorovich problem with entropic regularization in the case of finitely supported positive measures $\alpha ,\beta \in {\mathcal {M}}_{+}(X)$.

Problem Formulation

Entropic regularization modifies the Kantorovich problem by adding a Kullback-Leibler divergence term to the optimization goal. Specifically, the general form of the problem is now to determine

$L_{c}^{\epsilon }(\alpha ,\beta )=\inf _{\pi \in \Gamma (\alpha ,\beta )}\underbrace {\int _{X\times Y}c(x,y)\mathop {} \!\mathrm {d} \pi (x,y)} _{{\text{Kantorovich functional }}\mathbb {K} (\pi )}+\epsilon \underbrace {\operatorname {KL} (\pi \mid \alpha \otimes \beta )} _{\text{entropic term}}$

where $\alpha \otimes \beta$ is the product measure of $\alpha$ and $\beta$, and where

$\operatorname {KL} (\mu \mid \nu )=\int _{X}\log \left({\frac {\mathrm {d} \mu }{\mathrm {d} \nu }}(x)\right)\mathop {} \!\mathrm {d} \mu (x)+\int _{X}(\mathop {} \!\mathrm {d} \mu (x)-\mathop {} \!\mathrm {d} \nu (x))$

whenever the Radon-Nikodym derivative ${\tfrac {\mathrm {d} \mu }{\mathrm {d} \nu }}$ exists (i.e. when $\mu$ is absolutely continuous w.r.t. $\nu$) and $+\infty$ otherwise. This form of the KL divergence is applicable even when $\mu ,\nu \in {\mathcal {M}}_{+}(X)$ differ in total mass and it reduces to the standard definition whenever $\mu$ and $\nu$ have equal total mass. From this definition it immediately follows that for $\epsilon >0$ an optimal coupling $\pi ^{*}$ must be absolutely continuous w.r.t $\alpha \otimes \beta$. As a result, the optimal plan is in some sense less singular and hence "smoothed out."

To apply Sinkhorn's algorithm to approximate $L_{c}^{\epsilon }(\alpha ,\beta )$, it will be necessary to assume finite support so let $\alpha =\textstyle \sum _{i=1}^{n}a_{i}\delta _{x_{i}}$ and $\beta =\textstyle \sum _{j=1}^{m}b_{i}\delta _{y_{j}}$ and denote the corresponding vector of weights by $\mathbf {a} \in \mathbb {R} _{+}^{n}$ and $\mathbf {b} \in \mathbb {R} _{+}^{m}$. Additionally let $C_{ij}=c(x_{i},y_{j})$ and denote the discrete version $\Gamma (\alpha ,\beta )$ by $U(a,b)=\{P\in \mathbb {R} ^{n\times m}\mid \textstyle \sum _{j}P_{ij}=a_{i},\textstyle \sum _{i}P_{ij}=b_{j}\}$. This let's us write the entropic Kantorovich problem as

$L_{c}^{\epsilon }(\mathbf {a} ,\mathbf {b} )=\inf _{P\in U(\mathbf {a} ,\mathbf {b} )}\sum _{i,j}C_{ij}P_{ij}+\epsilon \operatorname {KL} (P\mid \mathbf {a} \mathbf {b} ^{T})$