The Bregman distance Bξx(y,x)\Breg{\xi_x}(y,x), ξxJ(y),\xi_x \in \partial J(y), associated to a convex sub-differentiable functional JJ is known to be in general non-symmetric in its arguments xx, yy. In this note we address the question when Bregman distances can be bounded against each other when the arguments are switched, i.e., if some constant C>0C>0 exists such that for all x,yx,y on a convex set MM it holds that 1CBξx(y,x)Bξy(x,y)CBξx(y,x).\frac{1}{C} \Breg{\xi_x}(y,x) \leq \Breg{\xi_y}(x,y) \leq C \Breg{\xi_x}(y,x). We state sufficient conditions for such an inequality and prove in particular that it holds for the pp-powers of the p\ell_p and LpL^p-norms when 1<p<1 < p <\infty.

Contact details are reproduced from the original publication and may be historical.

S. Kindermann. “A Note on the Approximate Symmetry of Bregman Distances.” Journal of Convex Analysis 26 (2019), No. 3, 991–999.