To measure or not to measure... a real problem!


This post will be more theoretical than usual. If you are afraid of mathematics, go away now before it is too late!

I will write about a set, the Vitali set, which arises in measure theory, to show that there is no coherent way of assigning a size to every subset of the real line. Because of that I will try to define size for subsets of the real line and then verify that such task is impossible.
I will also enforce a couple of (seemingly reasonable) restrictions on my definition.

I challenge you to read this post and to calmly follow my reasoning. Whenever something doesn't seem obvious, try to make it clear by yourself with a piece of paper and pen/pencil. If any doubts persist, drop me your question(s) in the comments and I will answer gladly!

Let us call $m$ to the function that, given a subset of the real line, returns its size; that is, let us try to define $m: \mathcal{P}(\mathbb{R}) \to [0, \infty]$.

Let us also suppose that the function $m$ satisfies the following restrictions:
  • If $I$ is an interval with endpoints $a \leq b$, $m(I) = b-a$. That is, the size of an interval is what we intuitively call its length. For example, $m([3,4]) = 4-3 = 1$;
  • $m(E+x) = m(E)$: the size of a set remains unchanged under translations, i.e. even if a set is "pushed over", its size remains the same;
  • If we have a finite number of sets, or a numerable quantity of them, and if they are all disjoint, then the size of their union must equal the sum of their individual sizes: $m (\cup_{n=1}^{\infty}E_n) = \sum_{n=1}^{\infty}m (E_n) $ whenever all the $E_n $ are disjoint.

Under these restrictions, can we create the function $m$?

Giuseppe Vitali, an italian mathematician, found out that the answer was "no". Vitali discovered a set $V \subset \mathbb{R}$ that allows one to conclude that we cannot create a function $m$ satisfying the restrictions above. I will now present the construction of the Vitali set:

Let us take a real number $x$ and define the set $C_x = \{y \in \mathbb{R}: y-x \in \mathbb{Q}\}$, i.e. the set $C_x$ contains all numbers that differ of $x$ by a rational amount: $C_x = \{x + q, \forall q \in \mathbb{Q}\}$ (in fact, the set $C_x$ represents the equivalence class of $x$ under the equivalence relation $x \sim y \iff x-y\in\mathbb{Q}$).

Having defined the sets $C_x$ for all real $x$, we pick out of every unique $C_x$, a value $v_x$ such that $v_x \in ]0, 1[$.
I said "out of every unique $C_x$" because, for example, $C_0 = C_1$; for this construction, we do not want repeated sets $C_x$. We will also say that the number $v_x$ is a representative of the set $C_x$ and all its elements. As an example, $0.3$ could be a representative of the set $C_0$ and $\pi - 3.1$ could be a representative of the set $C_\pi$.
(By the axiom of choice) we have a set $V$ - which we will call the Vitali set - where we have, from each unique set $C_x$, a single representative $v_x \in ]0, 1[$.
We now ask what should be the value $m(V)$. Either $m(V) = 0$ or $m(V) > 0$. We will build an auxiliary set which will allow one to conclude that both cases lead to incoherent results.

To build the auxiliary set, let us consider $\{r_1, r_2, \cdots, r_n, \cdots \}$ an enumeration of all the rationals in $]-1, 1[$ and let us define $V_j = V + r_j$: $V_j$ is the translation of the set $V$ by $r_j$. Then we define $$G = \cup_{j=1}^\infty V_j$$ We notice that the sets $V_j$ are all disjoint! Suppose $y$ is an element lying simultaneously in $V_j$ and in $V_k$ (with $j \neq k$). This means that
  1. $x_j = y-r_j$ is in $V$ because $y \in V_j = V + r_j$;
  2. $x_k = y-r_k$ is in $V$ because $y \in V_k = V + r_k$;
  3. From $1.$ and $2.$ we have that $x_j - x_k = (y - r_j) - (y - r_k) = r_k - r_j$ is a rational number, i.e. $x_j - x_k \in \mathbb{Q}$;
From $3.$ we get that $x_k \in C_{x_j}$ but, by definition, $V$ only has a representative from each $C_x$. Thus we obtain a contradiction and there can be no real number $x$ lying in two different $V_j$: all the $V_j$ are disjoint.

Because of the restrictions we imposed on $m$, we have that $m(V_j) = m(V + r_j) = m(V)$ because $m$ is constant under translations; we also have that $$m(G) = m\left(\cup_{j=1}^\infty V_j\right) = \sum_{j=1}^\infty m(V_j) = \sum_{j=1}^\infty m(V)$$ because the $V_j$ are all disjoint and we are performing their union.

Now we can see that if it is the case that $m(V) = 0$, then $m(G) = 0$. But $m(G) = 0$ is absurd because the interval $]0, 1[$ is inside $G$ and $m(]0, 1[) = 1$ implying that $m(G) \geq 1$. To verify that $]0, 1[ \subset G$, let $x \in ]0,1[$. This element $x$ belonged to some initial $C_\bullet$ and therefore has a representative $y$ in $]0, 1[$. If $y$ is a representative of $x$, then $x$ and $y$ differ by a rational number. But $x,y \in ]0, 1[$ means that their distance is smaller than $1$ and, in particular, $x - y \in ]-1, 1[$. Adding everything together, $x-y \in \mathbb{Q}$ and $x-y \in ]-1, 1[$ gives that $x-y = r_k$ for some $k$, as $\{r_1, \cdots, r_n, \cdots\}$ was an enumeration of all the rationals in $]-1, 1[$.
If it was the case that $m(V) > 0$, then $m(G) = \infty$. But $m(G) = \infty$ is another absurd given that $G \subseteq ]-1, 2[$ implying that $m(G) \leq m(]-1, 2[) = 3$. To see that $G \subseteq ]-1, 2[$ we just have to remember that $G$ is a numerable union of the sets $V_j$ and every one of those is $V_j = V + r_j$. We had that $V \subset ]0, 1[$ and $r_j \in ]-1, 1[$ hence $V_j \subseteq ]-1, 2[$.
(the image above has some points from $V$ in red, one of the rationals $r_j$ and the respective set $V_j$ in blue; because every $r_j$ lies in $]-1, 1[$ it is obvious that $V_j \subset ]-1, 2[$)

Looking back at what we have, neither $m(V) = 0$ nor $m(V) > 0$ produce coherent results! We conclude that our function $m$ cannot exist as I envisioned it in the beginning of the post... even though the restrictions I imposed seemed perfectly reasonable!

I understand if this post isn't the easiest to digest. If you have any doubts regarding the construction I described or anything related, drop a line on the comments section and I will be glad to answer!

  - RGS

Popular posts from this blog

Pocket maths: how to compute averages in your head

Tutorial on programming a memory card game

Markov Decision Processes 02: how the discount factor works