Is there any probability distribution supported on a compact or a half-open interval (of $\mathbb{R}$) such that if a vector $\vec{x} \in \mathbb{R}^n$ is sampled by sampling its coordinates like that then there is a closed form expression for the distribution of $\langle \vec{a} , \vec{x}\rangle$ (as a function of $\vec{a}$)?

The closest example I know of is the Gaussian distribution which is supported on the whole of $\mathbb{R}$.


(1) supported on half planes of $\mathbb{R}^n$, you may want to look at folded Gaussian distributions.

(2) supported on a compact surface like $\mathbb{S}^n$, you may want to look at projected Gaussian distributions. If $\vec{a}$ is fixed you will get a degenerated version of projected Gaussian.

  • $\begingroup$ Thanks! Could you share your intuition about why you think the PDF of $\langle \vec{a}, \vec{x} \rangle$ has a closed form expression with coordinates of $\vec{x}$ coming from say the folded normal? $\endgroup$ Aug 1 '17 at 22:51
  • $\begingroup$ @gradstudent You are basically projecting the folded normal $vec{x}$ onto the direction defined by $\vec{a}$ $\endgroup$
    – Henry.L
    Aug 1 '17 at 22:55

If each coordinate is distributed exponentially with parameter $\lambda$, then there are explicit expressions for the distribution of the inner product. E.g.:

In two dimensions where $\vec{a}=(a,b)$: $$P[\langle \vec{a},\vec{x} \rangle < d] = \frac{f(a)-f(b)}{a-b}\ \text{ with }\ f(u) = \max(u,0)-|u|e^{\min(0,-\lambda d/u)} $$

In three dimensions where $\vec{a}=(a,b,c)$, with $a,b,c,d>0$:

$$P[\langle \vec{a},\vec{x} \rangle < d] = \frac{g(a)}{(c-a)(a-b)}+\frac{g(b)}{(a-b)(b-c)}+\frac{g(c)}{(b-c)(c-a)} $$ $$\text{ with } g(u)=u^2 (e^{-\lambda d/u}-1)$$

  • $\begingroup$ Thanks! Are these obtained by some brute force computation like just doing iterated convolutions or are there tricks to quickly generalize this to $n-$dimensions? $\endgroup$ Aug 2 '17 at 15:58
  • $\begingroup$ Brute force will get you formulas, but not nice formulas. It took substantial tinkering with Mathematica results to get something readable. Maybe tinkering with the n=4 case will make a generalization obvious. Try Probability[ a x + b y < d, {x [Distributed] ExponentialDistribution[lambda], y [Distributed] ExponentialDistribution[lambda]}] and the like, but beware that this command for two dimensions gives a result with 13 cases before tinkering! $\endgroup$
    – Matt F.
    Aug 2 '17 at 16:10

In general, you are just asking about a weighted sum of i.i.d. variables from distribution $D$, with weights $\alpha_1,\dots,\alpha_n$. The Gaussian distribution is the only one that is rotationally invariant when coordinates are sampled i.i.d., so I'd expect it to be the only one that depends only on $\|\alpha\|$ rather than on all of the $\{\alpha_i\}$.

So one direction to look is for distributions where weighted sums belong to some other known distribution, stable distributions being a special case. (I guess the Levy is supported only on half the line, so it should fit what you're looking for...but it may not be a very "nice" distribution.)

Another approach is via the moment generating function or characteristic function. Suppose $S = \sum_{i=1}^n X_i$, where $X_i = \alpha_iY_i$ and all $Y_i$ are i.i.d. Then

\begin{align*} \mathbb{E} e^{tS} &= \prod_{i=1}^n \mathbb{E} e^{tX_i} \\ &= \prod_{i=1}^n \mathbb{E} e^{t \alpha_i Y_i} \\ &= \prod_{i=1}^n f(t \alpha_i) \end{align*} where $f$ is the MGF of the distribution. For example, with exponential$(\lambda)$ variables, the MGF is $\frac{\lambda}{\lambda-t}$, and if $Y_i$ is exponential(1) then $\alpha_i Y_i$ is exponential$(1/\alpha_i)$, so its MGF is $\frac{1}{1-\alpha_i t}$. So for exponentials, the MGF of $\langle \vec{\alpha}, \vec{X} \rangle$ is $$ \prod_{i=1}^n \frac{1}{1-\alpha_i t} . $$ Now whether you can recover a distribution from the MGF is another question, I don't have expertise in this but can point to https://www.quora.com/Given-an-MGF-of-a-random-variable-how-does-one-derive-the-PDF-or-PMF and https://math.stackexchange.com/questions/655302/gamma-distribution-out-of-sum-of-exponential-random-variables


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.