It’s been a while since I made posts last time. For the past few months, I graduated in June this year and just started a new journey as a data scientist recently. While enjoying the work I do, I still find it interesting to think small problems. Here are the two little problems I encountered a while ago.
-
Suppose we break a stick in two at random. What is the expected ratio of the length of the shorter to the longer one?
-
Now we have two sticks one is clearly longer than the other and we make measurements of their length with a random instrumental error. The error distribution is zero mean with a constant variance. If only two measurements are allowed, can we improve the accuracy than just measuring each stick once?
Solution (CLICK ME):
1. The problem formulation is straightforward. WLOG, let's assume the length of the stick is one. Denote the uniform distribution on [0,1] by $U\sim \text{Unif}[0,1]$. Let $X$ be the length of the shorter one. Then, we can derive the distribution of $X$ as follows: $\forall ~0\le x\le \frac{1}{2}$, $$ F(x) = P(X\le x) = P(U\le x \text{ or } U\ge 1-x) = x+(1-(1-x)) = 2x. $$ Therefore, the density of $X$ is given by $f(x) = 2$ if $x \in [0,\frac{1}{2}]$, which is the uniform distribution on $[0,\frac{1}{2}]$. Now we can calculate the expected ratio of the length of the shorter to the longer one: $$ \E\left[\frac{X}{1-X}\right] = \int_0^{0.5} \frac{x}{1-x}\cdot 2~dx = 2\log 2-1 \approx 0.386. $$ Note that the expected ratio of the longer to the shorter one doesn't exist: $$ \E\left[\frac{1-X}{X}\right] = \int_0^{0.5} \frac{1-x}{x}\cdot 2~dx = 2\int_0^{0.5} \frac{1}{x}~dx - 1 = \infty. $$ The problem with the ratio $\frac{1-X}{X}$ is that as the denominator approaches zero the density doesn't decay. Actually, the phenomenon when a r.v. $Y$ has finite expected value while $\frac{1}{Y}$ does not is quite common. One may consider uniform distribution and inverse uniform distribution, exponential distribution and inverse exponential distribution, etc. 2. Let $\varepsilon$ denote the random instrumental error every time we make a measurement. Say two sticks have lengths $x,y$ and $x>y$. If we put two sticks together and measure $x+y$, we get $x+y+\varepsilon_1$. Next, we align two sticks and measure $x-y$ to get $x-y+\varepsilon_2$. By simple algebra, we obtain our measurement of $x$: $x+\frac{\varepsilon_1+\varepsilon_2}{2}$ and measurement of $y$: $y+\frac{\varepsilon_1-\varepsilon_2}{2}$. Note that even though the expectation of the error parts are the same as those when we taking measurements directly ($\E\frac{\varepsilon_1+\varepsilon_2}{2} = \E\varepsilon_1 = \E\varepsilon_2 = 0$), the variances actually decrease: $\Var\left(\frac{\varepsilon_1+\varepsilon_2}{2}\right) = \Var\left(\frac{\varepsilon_1-\varepsilon_2}{2}\right) = \frac{\Var(\varepsilon)}{2} < \Var(\varepsilon)$. By doing what's stated above, we take two measurements of $x$ and $y$ and thus get better estimates.