1: \begin{abstract}

2: Motivated by the manifold hypothesis,

3: which states that data with a high extrinsic dimension

4: may yet have a low intrinsic dimension,

5: we develop refined statistical

6: bounds for entropic optimal transport

7: that are sensitive

8: to the intrinsic dimension of the data.

9: Our bounds involve a robust notion

10: of intrinsic dimension, measured at only

11: a single distance scale

12: depending on the regularization parameter,

13: and show that

14: it is only the minimum

15: of these single-scale intrinsic dimensions which 

16: governs the rate of convergence. We call this the Minimum

17: Intrinsic Dimension scaling (MID scaling) phenomenon, and establish MID scaling

18: %

19: %

20: with no assumptions on the data distributions

21: so long as the cost is bounded and Lipschitz,

22: and

23: for various entropic optimal transport quantities beyond just

24: values, %

25: %

26: with stronger analogs when one distribution

27: is supported on a manifold.

28: Our results significantly

29: advance

30: the theoretical state of the art

31: by showing that MID scaling is a generic

32: phenomenon, and 

33:  provide the first

34: rigorous interpretation

35: of the statistical

36: effect of entropic regularization as a distance

37: scale.

38: %

39: %

40: %

41: %

42: \end{abstract}

43: