I am trying to get the probability distribution function of $Z=X-Y$. Given that $f_X(x)$ and $f_Y(y)$ are known, and both variables are chi-square distributed, $X\in\mathbb{R}$, $X\ge 0$, and similarly $Y\in\mathbb{R}$, $Y\ge 0$. So how can I get $f_Z(z)$.

  • 1
    $\begingroup$ Are they $\chi^2$-distributed with the same parameters, or different ? $\endgroup$
    – Sasha
    Nov 24, 2011 at 16:08
  • 5
    $\begingroup$ Regardless of the distribution, you need to know the joint distribution of $X$ and $Y$ in order to answer the question. For the special case when $X$ and $Y$ are independent, the density of $X-Y$ is the convolution of the densities of $X$ and $-Y$, in particular, $$f_{X-Y}(z) = \int_{-\infty}^{\infty} f_X(x)f_Y(x-z)\mathrm dx.$$ As Sasha hints, this integral is easier to compute when the densities of $X$ and $Y$ have the same scale parameter. $\endgroup$ Nov 24, 2011 at 16:35
  • $\begingroup$ @DilipSarwate : I took Sasha's comment to be about degrees of freedom. The number of degrees of freedom is not a scale parameter. $\endgroup$ Nov 24, 2011 at 18:45
  • $\begingroup$ @MichaelHardy : Sasha wrote parameters and so could have meant both scale and degrees of freedom. As you know, $\Chi^2$ random variables are also Gamma random variables, and the sum of independent Gamma random variables with the same scale parameter is a Gamma random variable with the same scale parameter and order parameter equal to the sum of the orders, but the density is a lot messier when the scale parameters are different. I suspect a similar situation for the density of the difference of two $\Chi^2$ random variables; easier to find when the scale parameters are the same. $\endgroup$ Nov 24, 2011 at 19:49
  • 1
    $\begingroup$ A more general question: stats.stackexchange.com/questions/48378/… $\endgroup$ Jun 18, 2016 at 18:13

1 Answer 1


Let's assume that $X$ and $Y$ are independent, and both follow $\chi^2$-distribution with $\nu$ degrees of freedom.

Then $Z = X-Y$ follows symmetric about the origin variance-gamma distribution with parameters $\lambda = \frac{\nu}{2}$, $\alpha=\frac{1}{2}$, and $\beta=0$ and $\mu = 0$.

The best way to see this is through the moment-generating function: $$ \mathcal{M}_X(t) = \mathcal{M}_Y(t) = \left(1-2 \, t\right)^{-\nu/2} $$ Then $$ \mathcal{M}_Z(t) = \mathcal{M}_X(t) \mathcal{M}_Y(-t) = \left( 1-4 t^2 \right)^{-\nu/2} = \left( \frac{1/4}{1/4-t^2}\right)^{\nu/2} $$ We now this that it match the moment generating function of the variance gamma distribution: $$ \mathcal{M}_{\rm{V.G.}(\lambda,\alpha,\beta,\mu)}(t) = \mathrm{e}^{\mu t} \left( \frac{\alpha^2 -\beta^2}{\alpha^2 - (\beta+t)^2 } \right)^\lambda $$ For the said parameters, $\mu=\beta=0$, $\alpha=\frac{1}{2}$ and $\lambda=\frac{\nu}{2}$, the density has the following form: $$ f_Z(z) = \frac{1}{2^{\nu/2}\sqrt{\pi}} \frac{1}{\Gamma\left(\frac{\nu}{2}\right)} \vert z \vert^{\tfrac{\nu-1}{2}} K_\tfrac{\nu-1}{2}\left(\vert z \vert \right) $$ The function $f_Z(z)$ is continuous at $z=0$ for $\nu > 1$, with $$ \lim_{z \to 0} f_Z(z) =\frac{1}{4 \sqrt{\pi }} \frac{\Gamma \left(\frac{\nu }{2}-\frac{1}{2}\right)}{ \Gamma \left(\frac{\nu }{2}\right)} $$

  • $\begingroup$ Thanks, now I understand it better. I have never thought that I can use the MGF with the variable inverted in sign, i.e., $\mathcal{M}(-t)$. $\endgroup$
    – Sophie
    Nov 25, 2011 at 15:03
  • $\begingroup$ @Remmy This simply follows because $\mathcal{M}_Z(t) = \mathbb{E}(\exp(t Z)) = \mathbb{E}(\exp(t X) \exp(-t Y)) = \mathcal{M}_X(t) \mathcal{M}_Y(-t)$, because of independence. $\endgroup$
    – Sasha
    Nov 25, 2011 at 15:05
  • 1
    $\begingroup$ See also: Chi-squared distribution#Linear combination. $\endgroup$ Jun 18, 2016 at 17:52
  • 4
    $\begingroup$ I thought it would be appropriate to add a paper that referenced this answer (been reading it just now) as a token of appreciation to Sasha: arxiv.org/pdf/2005.10009.pdf $\endgroup$
    – NULL
    Jul 28, 2020 at 19:57
  • $\begingroup$ I am comparing this answer with en.wikipedia.org/wiki/…. I suspect that there is a mistake in the expression of the pdf... it looks like $\alpha$ should be incorporated into the expression inside the modified Bessel function, shouldn't it? Correspondingly, the normalizing constant should be adjusted. Am I missing something? $\endgroup$
    – Roja
    Dec 28, 2020 at 13:12

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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