13
$\begingroup$

It is apparently the case that if $X_i \sim N(0,1)$, then

$X_1 X_2 + X_3 X_4 \sim \mathrm{Laplace(0,1)}$

I've seen papers on arbitrary quadratic forms, which always results in horrible non-central chi-squared expressions.

The above simple relationship does not seem at all obvious to me, so (if it is true!) does anyone have a simple proof of the above?

$\endgroup$
17
$\begingroup$

An elementary sequence of steps using well-known relationships among distributions and a simple algebraic polarization identity provide an elementary and intuitive demonstration.

I have found this polarization identity generally useful for reasoning about, and computing with, products of random variables, because it reduces them to linear combinations of squares. It is a bit like working with matrices by diagonalizing them first. (There's more than a superficial connection here.)


A Laplace distribution is a difference of two Exponentials (which intuitively makes some sense, because an Exponential is a "half-Laplace" distribution). (The link demonstrates this by manipulating characteristic functions, but the relation can be proven using an elementary integration following from the definition of a difference as a convolution.)

An Exponential distribution (which itself is a $\Gamma(1)$ distribution) is also a (scaled version of a) $\chi^2(2)$ distribution. The scale factor is $1/2$. This can easily be seen by comparing the PDFs of the two distributions.

$\chi^2$ distributions are obtained naturally as sums of squares of iid Normal distributions (having zero means). The degrees of freedom, $2$, count the number of Normal distributions in the sum.

The algebraic relation

$$X_1X_2 + X_3X_4 = \left[\left(\frac{X_1+X_2}{2}\right)^2 + \left(\frac{X_3+X_4}{2}\right)^2\right] - \left[\left(\frac{X_1-X_2}{2}\right)^2 + \left(\frac{X_3-X_4}{2}\right)^2\right]$$

exhibits $X_1X_2 + X_3X_4$ in terms of squares of four distributions, each of which is a linear combination of standard Normals. It is easy to check that all four linear combinations are linearly independent (and each follows a Normal$(0,\sqrt{1/2})$ distribution). Thus the first two terms, which sum the squares of two identically distributed Normal distributions of mean zero, form a scaled $\chi^2(2)$ distribution (and its scale factor of $\sqrt{1/2}\ ^2=1/2$ is exactly what is needed to make it an Exponential distribution) and the second two terms independently have an Exponential distribution, too, for the same reason.

Therefore $X_1X_2+X_3X_4$, being the difference of two independent Exponential distributions, has a (standard) Laplace distribution.

$\endgroup$
  • 4
    $\begingroup$ That is absolutely delightful! $\endgroup$ – Corone Sep 26 '13 at 15:35
  • 2
    $\begingroup$ I just noticed that another answer, based on moment generating functions, appears at stats.stackexchange.com/a/51717/919: see the paragraph in the middle beginning "incidentally" (another name for the Laplace distribution is "bi-exponential"). That thread concerns the MGF of a generalization of the present question. $\endgroup$ – whuber Sep 26 '13 at 17:25
  • $\begingroup$ Nice derivation, but how do you know that the difference of two independent exponential distributed variables has a Laplacian distribution? $\endgroup$ – HelloGoodbye Sep 24 '16 at 14:14
  • $\begingroup$ @Hello Please follow the link: it goes to a Wikipedia article that includes a brief demonstration. $\endgroup$ – whuber Sep 24 '16 at 14:50
13
$\begingroup$

$X\sim \mathrm{Laplace}(0,1)$ has characteristic function $$ \phi_X(t) = \frac{1}{1+t^2} $$ which is the square of the characteristic function of a product standard normal (see https://math.stackexchange.com/questions/74013/characteristic-function-of-product-of-normal-random-variables). The claim follows by the fact that sums of independent random variables relate to products of characteristic functions.

$\endgroup$

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.