1: \begin{abstract}

2:     We study covariate shift in the context of nonparametric

3:     regression.  We introduce a new measure of distribution mismatch

4:     between the source and target distributions that is based on the

5:     integrated ratio of probabilities of balls at a given radius.  We

6:     use the scaling of this measure with respect to the radius to

7:     characterize the minimax rate of estimation over a family of

8:     Hölder continuous functions under covariate shift.  In comparison

9:     to the recently proposed notion of transfer exponent, this measure

10:     leads to a sharper rate of convergence and is more

11:     fine-grained. We accompany our theory with concrete instances of

12:     covariate shift that illustrate this sharp difference.

13:   \end{abstract}

14: