I'm trying to show that a scale mixture of normals yields a Laplace distribution. I've gotten to the point where I have $\int N(0,\tau)\times Ga(\tau\:;\:1,\frac{\lambda^{2}}{2}) \:d\tau$ should equal a Laplace. It's unclear to me how to solve the integral.

Alternatively, I've seen a derivation of the scale of mixture normals as a power exponential here, with certain parameter values yielding the Laplace. However, to be honest it's hard to follow. Can someone please help decipher this derivation or provide a clue as to how to solve the integral?

  • 1
    I found the integral above (with Gamma, not Inverse Gamma) in this presentation: newton.ac.uk/files/seminar/20061214161517004-150400.pdf (page 6)Are they formulating the problem incorrectly? – ilanman Oct 5 '15 at 16:45
  • Ok great. However, the calculation of this integral still eludes me (it's been a while since I integrated anything...) – ilanman Oct 5 '15 at 18:38
up vote 4 down vote accepted

The marginal distribution of $\beta$ associated with $$\beta|\tau\sim\mathcal{N}(0,\tau)\quad\tau\sim\mathcal{E}(\lambda^2/2)$$ [with the convention that $\tau$ is the variance] has density \begin{align*}\mathfrak{t}(\beta) &=\int_0^\infty \tau^{-1/2}\varphi(\beta/\sqrt{\tau})\frac{\lambda^2}{2}\exp\{-\lambda^2\tau/2\}\text{d}\tau\\ &=\frac{\lambda^2 e^{-\lambda|\beta|}}{2\sqrt{2\pi}}\int_0^\infty \tau^{-1/2}\exp\left\{-\frac{1}{2}\left(\frac{|\beta|}{\sqrt{\tau}}-\lambda\sqrt{\tau}\right)^2\right\}\text{d}\tau\\ \end{align*} [where the $\lambda|\beta|$ appears by creating a perfect square in the exponential]. This suggests the change of variable $\nu=\sqrt{\tau}$ and leads to $$\mathfrak{t}(\beta) = \frac{\lambda^2 e^{-\lambda|\beta|}}{2\sqrt{2\pi}}\int_0^\infty \exp\left\{-\frac{1}{2}\left(\frac{|\beta|}{\nu}-\lambda\nu\right)^2\right\}\text{d}\nu$$ [since $\tau^{-1/2}\text{d}\tau=2\text{d}\nu$]. This further suggests the change of variable $$\zeta=\frac{|\beta|}{\nu}-\lambda\nu$$ with its inverse $$\nu=\left\{-\zeta+\sqrt{\zeta^2+4\lambda|\beta|} \right\}\big/2\lambda$$ [obtained by solving a second degree polynomial equation] and the Jacobian $$\frac{\text{d}\nu}{\text{d}\zeta}=\left\{-1+\frac{\zeta}{\sqrt{\zeta^2+4\lambda|\beta|}} \right\}\big/2\lambda$$ which is always negative. Hence \begin{align*}\mathfrak{t}(\beta)&=\frac{\lambda^2 e^{-\lambda|\beta|}}{4\lambda\sqrt{2\pi}}\int_{-\infty}^\infty \exp\left\{-\frac{\zeta^2}{2}\right\}\left\{1-\frac{\zeta}{\sqrt{\zeta^2+4\lambda|\beta|}} \right\}\text{d}\zeta\\ &=\frac{\lambda e^{-\lambda|\beta|}}{2}\int_{-\infty}^\infty \left\{1-\frac{\zeta}{\sqrt{\zeta^2+4\lambda|\beta|}} \right\}\varphi(\zeta)\text{d}\zeta\\ &=\frac{\lambda e^{-\lambda|\beta|}}{2}\left\{1-\int_{-\infty}^\infty\frac{\zeta\varphi(\zeta)}{\sqrt{\zeta^2+4\lambda|\beta|}}\text{d}\zeta\right\}=\frac{\lambda e^{-\lambda|\beta|}}{2}\end{align*} [since the integrand is an odd function of $\zeta$ in the last integral]. This establishes [without complex calculus] that the marginal distribution of $\beta$ is indeed a Laplace or double-exponential distribution.

I've always found the direct integration in this case to be a complicated integral. The Moment Generating Function (MGF) approach works too.

The MGF: $\beta | \tau \sim N(0, \tau)$ and then

$M_{\beta|\tau}(t)=e^{\frac{\tau t^2}{2}}.$ Now to get the MGF of $\beta$ marginally, take the expectation with respect to $\tau$.

$$\mathbb{E}(M_{\beta|\tau}(t)) = \int_0^\infty e^{\frac{\tau t^2}{2}} \frac{\lambda^2}{2}e^{-\tau \frac{\lambda^2}{2}}d\tau =\int_0^\infty \frac{\lambda^2}{2} e^{-\tau \left(-\frac{t^2}{2} +\frac{\lambda^2}{2}\right)} d\tau =\frac{\lambda^2/2}{\lambda^2/2 - t^2/2} =\frac{1}{1 - \frac{t^2}{\lambda^2}}, $$ Now you can recognize this last function as the MGF of a Laplace (double exponential) distribution.

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

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