The n'th cumulant is defined to be the n'th derivative of the CGF (cumulant generating function).

$$\kappa_n = \frac{d^n K(t)}{dt^n} |_{t=0} $$

But I'm reading in a book (p.215, chapter5, eq. 5.8) now that for the exponential family / exponential dispersion model, this is actually equal to:

$$ K \underset{exp.}{=} \frac{\kappa(\theta+t\phi) - \kappa(\theta)}{\phi}\\ \kappa_n = \phi^{n-1} \frac{d^n\kappa(\theta)}{d\theta^n} $$


  1. $\theta$ is the canonical/natural parameter in exponential family.

  2. $K$ is cumulant generating function

  3. $\kappa_n$ is the nth cumulant

I'm not really sure how come you get this result. The $\phi^{n-1}$ I understand, but not how the derivative changes from $t$ to $\theta$.


1 Answer 1


As you did'nt define your terms, I will use the definition of exponential dispersion model used in Definition of exponential family with dispersion parameter, which is $$ f(y|\theta,\phi) = \exp\left(\frac{y\theta - b(\theta)}{a(\phi)} + c(y, \phi)\right) $$ which is an exponential family in $\theta$, for each fixed $\phi$. It seems you have in your notation $a(\phi)=\phi$. Then we can calculate the mgf (moment generating function) as $$ M(t)=\exp\left( \frac{b(t a(\phi)+\theta)-b(\theta)}{a(\phi)} \right) $$ so the cumulant generating function $$ K(t)=\log M(t)= \frac{b(t a(\phi)+\theta)-b(\theta)}{a(\phi)} . $$ Then $$ K'(t)=\frac{b'(t a(\phi)+\theta)\cdot a(\phi)}{a(\phi)}=b'(t a(\phi)+\theta) $$ where $'$ means the derivative of a function with respect to its argument. It does not matter if that argument is named $t$ or $\theta$. Now it is easy to see that we can continue differentiating to get $$ K^{(n)}(t)=b^{(n)}(t a(\phi)+\theta)\cdot a^{n-1}(\phi) $$ by repeated application of the chain rule. Now, if you set $t=0$ and revert to the $\frac{d}{d t}$ notation, you get the result. Now, since on the RHS the argument of $b$ is $\theta$, we differentiate there $b$ with respect to its argument, which explains your confusion.


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.