4

I am wondering if there is an algebraic solution (i.e. an equation) to find the intersection point(s) of two Weibull probability densities that have different scale and shape parameters. While I have tried myself and searched for a solution, I have come up blank. I am not even sure if an explicit solution is possible. However, if it is, I'd be keen to know what it is.

A couple of clarifications: Firstly, I am not looking for a solution in R or some other package. I am seeking an equation. Secondly, to be clear, this question relates to the Weibull distribution. I am more than familiar how to find the intersection of two normal distributions.

Mari153
  • 880

1 Answers1

5

Let's take two Weibull distributions with parameters $(\alpha,\beta)$ and $(k,\lambda)$, respectively.

We know the general form of the Weibull density is:

$$f(x)=\bigg(\frac{k}\lambda\bigg)\bigg(\frac{x}\lambda\bigg)^{k-1}{e^{-(x/\lambda)}}^{k}$$

Obviously, we can see that for any $k>1$, $\lambda>0$ that there is an intersection at

$$x=0$$

In fact, the number of intersection points is dependent on $\alpha,k$. We can summarise the number of intersection points below:

$$\begin{array}{cc|lll} &&&k&&\\ && <1 & =1 & >1 \\ \hline & <1 & 2 & 2 & 2 \\ \alpha & =1 & 2 & 1^{1} & 2 \\ & >1 & 2 & 2 & 3^{2} \end{array}$$

$^1$This is the special case of the intersection of two exponential density functions. In this case, the intersection point is given by $x=\beta\cdot\lambda\cdot\log(\beta/\lambda)/(\beta-\lambda)$.

$^{2}$Where one of the intersection points is always at $(x,y)=(0,0)$.

If we equate the two density functions, we get:

$$(k-\alpha)\log(x)-\bigg(\frac{x}{\lambda}\bigg)^{k}+\bigg(\frac{x}{\beta}\bigg)^{\alpha}-\log\bigg(\frac{\alpha\cdot\lambda^{k}}{k\cdot\beta^{\alpha}}\bigg)=0$$

Unfortunately, there isn't a simple solution to this. We can solve for the remaining roots numerically:

a=1.7
b=2
k=2
l=1

fun = function(x, pars) {
  return((pars[3]-pars[1])*log(x)-(x/pars[4])^pars[3]+(x/pars[2])^pars[1]+log(pars[3])-pars[3]*log(pars[4])- log(pars[1])+pars[1]*log(pars[2]))
}

ss=uniroot(f = fun, pars = c(a,b,k,l), interval = c(3, 10), extendInt = "yes")

> ss$root
[1] 1.414019

You may need to try different starting points to find all of the roots.

The two curves will always converge in the tails: $$\lim\limits_{x\to\infty}f(x;\alpha,\beta)-f(x;k,\lambda)=0$$

We can visualize an example of these intersections below for each of the cases:

WEIBULLS

statsplease
  • 2,808
  • Is there a final intersection point in the tail, or is that only at the limit (and therefore not a real intersection point)? – jbowman Nov 27 '19 at 05:04
  • @jbowman Yes, not a real intersection point. Just both functions go to zero in the limit. – statsplease Nov 27 '19 at 05:06