Yes,werequire2n > dforthereason9 you mention. A viable stopping criterion of Algorithm 2is to check the angle difference of projection directions10 between two consecutive iterations. The algorithm is terminated when the angle is close to zero. Our theory can be readily applied to this case.

But in the field of reinforcement learning (RL),L2-norm is the most common choice due to its5 efficiency and effectiveness. Thus we adoptL2-norm in the paper to ensure consistency between the objective of6 Andersonacceleration(AA)andthelossofQ-valuefunction(critic).7 Minors.

When standardizing as in Figure 1, the10 smallest width produced by [Thm. We feel it is important to also maintain our existing real-data24 experiments, as these best reflect how the competing CIs and tests perform in practice, under the eccentricities of25 real data which are hard to capture with synthetic data. For example, itiscommon in real data to haveone method26 dominate for smaller sample sizes and the other dominate for larger sample sizes; this is precisely what we see in27 the right column of Figure 1.