Figure 2:

Relative improvements of the DR-ITR over the standard ITR as the difference of relative regrets on testing distributions ${\mathcal{N}}_2(\mathit{\mu },{\mathbf{\text{I}}}_2)$ of $\mathit{\mu }\in \{{({\mu }_1,{\mu }_2)}^{\top }\in {ℝ}^2:{\mu }_1^2+{\mu }_2^2⩽4\text{log}5\}$ (lighter the better).