It can be seen that the following random variates have the same distribution:

  1. $\frac{X_1 + X_3}{X_2 + X_3}$, where $(X_1, X_2, X_3) \sim \text{Dirichlet} (\alpha_1, \alpha_2, \alpha_3)$
  2. $\frac{Y_1 + Y_3}{Y_2 + Y_3}$, where $(Y_0, Y_1, Y_2, Y_3) \sim \text{Dirichlet} (\alpha_0, \alpha_1, \alpha_2, \alpha_3)$
  3. $\frac{Z_1 + Z_3}{Z_2 + Z_3}$, where the $(Z_i)_i$ are independent and $Z_i \sim \text{Gamma}(k = \alpha_i, \theta = 1)$

Question: does this distribution have a name? Has it been studied somewhere in the literature? Were it not for $X_1$ in the numerator, it seems that this would be a Beta-Prime distribution.

  • $\begingroup$ There is no $Y_0$ in your second formula. $\endgroup$ Commented Mar 11, 2023 at 15:31
  • $\begingroup$ @StephanKolassa that's intentional. $\endgroup$ Commented Mar 11, 2023 at 22:35
  • $\begingroup$ Once you know that Dirichlet distributions arise as ratios of Gamma variables, everything fits together. Such combinations of variables like $(X_1+X_3,X_2+X_3)$ have been used to formulate multivariate families of correlated Gamma distributions, suggesting one direction to search. $\endgroup$
    – whuber
    Commented Mar 11, 2023 at 23:45
  • $\begingroup$ @whuber thanks, unfortunately I'm not seeing how this gets me farther than the equivalence between 1., 2. and 3. $\endgroup$ Commented Mar 15, 2023 at 13:43
  • 1
    $\begingroup$ @whuber thanks, I appreciate that. $\endgroup$ Commented Mar 17, 2023 at 20:46

1 Answer 1


It turns out that yes, this distribution has indeed been studied:

Lee, Ru-Ying, Burt S. Holland, and John A. Flueck. "Distribution of a ratio of correlated gamma random variables." SIAM Journal on Applied Mathematics 36.2 (1979): 304-320.

The authors have not given a name to this 3-parameters family of distribution. They frame this in the context of the Cherian-David-Fix bivarate structure.

I would personally suggest to call this a Dirichlet Process Mass Ratio distribution, denoted $\text{DPMR}(\alpha_1, \alpha_2, \alpha_3)$, reflecting the fact that such a random variable is the ratio of (random) probability masses given to two sets $S_1$ and $S_2$ by a Dirichlet Process. $\alpha_1$ corresponds to $S_1-S_2$, $\alpha_2$ to $S_2-S_1$, and $\alpha_3$ to $S_1 \cap S_2$, consistently with writing these indices in binary (1 -> 01, 2 -> 10, 3 -> 11).


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.