I describe the problem dimension $2$, but it could be generalized to $n$ dimensions.

So we have $X_{1}, X_{2}, X_{3}$ three $iid$ random variables of continuous law $F(x,y)$ on $\Bbb R^2$. Let's denote by $S[X_{1}, X_{2}, X_{3}]$ the simplex generated by those random variables. Given a point $(x,y)$, is it possible to write $P((x,y)\in S[X_{1}, X_{2}, X_{3}])$ as a function $g$ of $F(x,y)$ ? For example, in dimension $1$, $P(x\in [X_{1}, X_{2}]) = 2F(x)(1-F(x))$.

Thank you

  • $\begingroup$ This post is somehow related. $\endgroup$ – EditPiAf Mar 21 '17 at 23:43
  • $\begingroup$ Unless I misunderstood the problem, I think it would be a littler neater to speak of four iid random variables, so we want to compute $P(X_4 \in S[X_1,X_2,X_3]$. no? $\endgroup$ – leonbloy Apr 19 '17 at 16:40
  • $\begingroup$ Looking at your second paragraph, it seems that my previous comment is not right, you are thinking on the fouth point as fixed (given). $\endgroup$ – leonbloy Apr 19 '17 at 16:48
  • $\begingroup$ A very particular case ( point is origin, distribution is uniform on a circle) here math.stackexchange.com/questions/268635 $\endgroup$ – leonbloy Apr 19 '17 at 16:49
  • $\begingroup$ I think the answer is in the paper arxiv.org/pdf/1612.08619v5.pdf $\endgroup$ – Eugen Ionascu Apr 5 '18 at 17:34

Given any three distinct points $A, B, C$, let

  • $\overrightarrow{AB}$ be the ray starting from $A$ pointing towards $B$.
  • $\overleftarrow{AC}$ be the ray starting from $A$ pointing away from $C$.
  • $\mathcal{C}(A,B,C)$ be the cone obtained by rotating a ray start from $\overrightarrow{AB}$ to $\overleftarrow{AC}$ counterclockwisely.
  • $\mathcal{H}(A,B) = \mathcal{C}(A,B,B)$ be the half-space on the "left hand" side of the ray $\overrightarrow{AB}$.

Given any fixed point $P = (x_P,y_P)$ and three random points $X_1, X_2, X_3$ sampled from a continuous distribution. Aside from events of probability zero, the four points $P, X_1, X_2, X_3$ will be in general position. i.e. no three of them will be colinear. Let $Y_1, Y_2, Y_3$ be a reordering of the $X_1, X_2, X_3$ so that $Y_1, Y_2, Y_3$ surrounds the triangle $\triangle_{X_1X_2X_3}$ in counterclockwise manner.

There are four mutually exclusive possibilities for the position of $P$ relative to $Y_1, Y_2, Y_3$. $P$ can lie in the interior of triangle $\triangle_{Y_1Y_2Y_3}$ or interior of one the cones $\color{red}{\mathcal{C}(Y_1,Y_3,Y_2)}, \color{green}{\mathcal{C}(Y_2,Y_1,Y_3)}$ and $\color{blue}{\mathcal{C}(Y_3,Y_2,Y_1)}$.

$\hspace{1in}$ an triangle and three cones

There is a simple criterion to test whether $P$ belongs to one of the cones.

As one can see from above diagram, $P \in \mathcal{C}(Y_1,Y_3,Y_2)$ is more or less equivalent to both $Y_2,Y_3$ lie at "left hand" side of ray $\overrightarrow{PY_1}$ (the hatched area in above diagram ). This happens to the other two cones. Aside from events of probability zero, we have $$ \begin{align} P \in \mathcal{C}(Y_1,Y_3,Y_2)\quad\iff\quad Y_2,Y_3 \in \mathcal{H}(P,Y_1)\\ P \in \mathcal{C}(Y_2,Y_1,Y_3)\quad\iff\quad Y_3,Y_1 \in \mathcal{H}(P,Y_2)\\ P \in \mathcal{C}(Y_3,Y_2,Y_1)\quad\iff\quad Y_1,Y_2 \in \mathcal{H}(P,Y_3) \end{align} $$

Using symmetry and the fact $X_1, X_2, X_3$ are $iid$ from same continuous distribution, we have

$$ \mathbb{P}\left[P \in \triangle_{X_1X_2X_3}\right] = 1 - 3\mathbb{P}\left[ Y_2,Y_3 \in \mathcal{H}(P,Y_1) \right] = 1 - 3\mathbb{E}\left[ \mathbb{P}\left[ X_2 \in \mathcal{H}(P,X_1) | X_1 \right]^2 \right]$$

This leads to following integral representation of the desired probability $$\mathbb{P}\left[P \in \triangle_{X_1X_2X_3}\right] = 1 - 3\int_{0}^{2\pi} \rho(P,\theta) \left[ \int_{\theta}^{\theta+\pi}\rho(P,\phi)d\phi \right]^2 d\theta $$ where $$\rho(P,\theta)\stackrel{def}{=} \int_{0}^{\infty}F(x_P+r\cos\theta,y_P+r\sin\theta) rdr$$ is the probability density function for ray $\overrightarrow{PX_1}$ making an angle $\theta$ with respect to the $x$-axis.

There is an interesting special case. When $F(x,y)$ is centrally symmetric with respect to $P$, $$\mathbb{E}\left[X_2 \in \mathcal{H}(P,X_1) | X_1\right] = \int_{\theta}^{\theta+\pi} \rho(P,\phi)d\phi = \frac12$$ independent of $\theta$. In that case, we have $$\mathbb{P}\left[P \in \triangle_{X_1X_2X_3}\right] = 1 - 3\left(\frac12\right)^2 = \frac14 $$


This is based on the barycentric coordinates. Applying Cramer's rule or using suitable software, one obtains \begin{aligned} a_1 &= \frac{x_2 y_3 - y_2 x_3 + x_3 y - x y_3 - x_2 y + y_2 x}{-x_2 y_1 + x_2 y_3 - x_1 y_3 + y_1 x_3 + y_2 x_1 - y_2 x_3} \, , \\ \\ a_2 &= \frac{x_1 y - x_1 y_3 - x_3 y + y_1 x_3 + x y_3 - y_1 x}{-x_2 y_1 + x_2 y_3 - x_1 y_3 + y_1 x_3 + y_2 x_1 - y_2 x_3} \, , \\ \\ a_3 &= \frac{-x_2 y_1 - x_1 y + y_1 x + x_2 y + y_2 x_1 - y_2 x}{-x_2 y_1 + x_2 y_3 -x_1 y_3 + y_1 x_3 + y_2 x_1 - y_2 x_3} \, , \end{aligned} where $X_i=(x_i,y_i)$ and $X=(x,y)$. It does not look like there is a simple rule to find the law of $a_i$, but the denominators in $\lbrace a_1, a_2, a_3\rbrace$ are all the same, and the numerators are very similar. One may view the problem with a different point of view. Indeed, one can compute the edge lines of the triangle: \begin{aligned} E_{12}: & \;\left(y - y_1\right) \left(x_2 - x_1\right) = \left(y_2 - y_1\right) \left(x - x_1\right) \, , \\ E_{13}: & \;\left(y - y_1\right) \left(x_3 - x_1\right) = \left(y_3 - y_1\right) \left(x - x_1\right) \, , \\ E_{23}: & \;\left(y- y_2\right) \left(x_3 - x_2\right) = \left(y_3 - y_2\right) \left(x - x_2\right) \, . \\ \end{aligned} The point $X=(x,y)$ belongs to the triangle if $X$ belongs to the same half-plane as $X_3$ with respect to $E_{12}$: \begin{aligned} \text{sign}\left(\left(y - y_1\right) \left(x_2 - x_1\right) - \left(y_2 - y_1\right) \left(x - x_1\right)\right) = \text{sign}\left(\left(y_3 - y_1\right) \left(x_2 - x_1\right) - \left(y_2 - y_1\right) \left(x_3 - x_1\right)\right) , \end{aligned} etc. The latter equation is not very different from $a_3\geq 0$.


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.