Can the Dvoretzky–Kiefer–Wolfowitz (a.k.a. Massart) inequality

$$ \Pr\left( \sup |\hat F_n(x) - F(x)| > \lambda \right) \le 2\exp(-2n\lambda^2) $$

where $\hat F_n(x)$ is the empirical distribution function of the sample $X$ from the distribution $F$, be extended to comparing two empirical distributions coming $\hat F_n,\hat G_n$ coming from equally-sized samples $X,Y$ from the same distribution? What is the $ \Pr\left( \sup |\hat F_n(x) - \hat G_n(x)| > \lambda \right) $ ?

  • $\begingroup$ Maybe it's too early for me but can't we deduce from $sup|\hat{F}_n(x) - F(x)| < \lambda$ and $sup|\hat{G}_n(x) - F(x)| < \lambda$ that $$sup|\hat{F}_n(x) - \hat{G}_n(x)| < 2\lambda$$ $\endgroup$ Apr 18, 2017 at 9:34
  • $\begingroup$ Adding the maximum errors together works, but is too conservative. $\endgroup$
    – Bscan
    Apr 16, 2018 at 21:00

1 Answer 1


Yes, the DKW inequality is essentially an inversion of the one sample Kolmogorov Smirnoff test. A two-sample version exists for it too where the primary difference is the scaling of the constant. So yes, you could invert this test too.

$$ D>c(\alpha)\sqrt{\frac{n + m}{n m}} $$

Here's a paper describing the process in detail: https://arxiv.org/pdf/1107.5356.pdf



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.