Non-Borel Sets: Discovering the Unconstructible in Mathematics

In mathematics, we often seek to classify objects, starting with simple building blocks and applying rules to generate more complex structures. When describing subsets of the real line, this process yields the vast family of Borel sets—seemingly encompassing every set we can imagine. But does this "constructible" universe contain everything? This article addresses the profound question of what lies beyond the Borel sets, exploring the existence of objects that defy this construction yet are essential to a complete understanding of measure and analysis. In "Principles and Mechanisms," we will define Borel sets, introduce the Lebesgue measure, and uncover the stunning proof that non-Borel sets must exist. Following that, in "Applications and Interdisciplinary Connections," we will investigate the surprising behavior of these sets under continuous functions and topological operations, revealing their crucial role in advancing modern measure theory and probability.

Imagine you want to describe every possible shape you can make. A natural way to start is with some basic building blocks—say, simple, solid bricks. You then establish a set of rules: you can stick bricks together, you can take a shape and consider the space it doesn't occupy, and so on. In the world of mathematics, when we try to describe subsets of the real number line, we play a very similar game. The sets we can build this way are wonderfully well-behaved and are at the heart of much of modern analysis. But as we shall see, the universe of sets is far stranger and more subtle than this simple construction game might suggest, holding "unbuildable" objects that are nonetheless perfectly real.

Our building blocks for the real line are the simplest sets imaginable: ​​open intervals​​, like $(0, 1)$ or $(-\pi, \sqrt{2})$. From these, we generate a vast and powerful family of sets called the ​​Borel sets​​. The construction rules are deceptively simple. If you have some sets, you can create new ones by:

1. Taking the ​​complement​​: If you have a set $A$, you can form its complement $A^c$, which is everything not in $A$.
2. Taking ​​countable unions​​: If you have a list of sets $A_1, A_2, A_3, \dots$, you can glue them all together to form their union $\bigcup_{n=1}^{\infty} A_n$.
3. Taking ​​countable intersections​​: From that same list, you can find the region they all share, their intersection $\bigcap_{n=1}^{\infty} A_n$.

The collection of all sets you can possibly create starting from open intervals using these rules, applied over and over again, is the ​​Borel $\sigma$-algebra​​, and its members are the Borel sets.

This process is purely about the structure, or ​​topology​​, of the real line. It has nothing to do with a set's "size" or "length" just yet; it's all about how sets are pieced together from open intervals. And you can build a staggering variety of things.

For instance, any ​​closed set​​, like $[0, 1]$, is a Borel set because it's the complement of an open set $(-\infty, 0) \cup (1, \infty)$. Even a bizarre object like the ​​Cantor set​​ $C$, formed by repeatedly removing the middle third of intervals, is Borel because it can be expressed as an infinite intersection of closed sets.

What about a set like the ​​rational numbers​​, $\mathbb{Q}$? This set is like a fine dust scattered across the real line. It's not open, nor is it closed. But we can write $\mathbb{Q}$ as a countable union of its individual points, like $\mathbb{Q} = \bigcup_{q \in \mathbb{Q}} \{q\}$. Each individual point $\{q\}$ is a closed set, so $\mathbb{Q}$ is a countable union of closed sets, which we can build with our rules. Therefore, $\mathbb{Q}$ is a Borel set. Even more intricate sets, such as the set of all numbers in $[0, 1]$ whose decimal expansion contains infinitely many 7s, can be shown to be Borel by expressing them as a clever sequence of countable unions and intersections of basic intervals.

The Borel sets seem to encompass every subset of the real line we could ever hope to describe or construct. This leads to a natural question: is there anything else? Is it possible for a set to be "measurable" in a meaningful way without being one of these "constructible" Borel sets?

Enter Henri Lebesgue, who revolutionized our notion of "length" or "measure". The ​​Lebesgue measure​​ is a way of assigning a size to a vast range of sets, far beyond simple intervals. A key feature of this measure is how it treats sets of "zero size." The Cantor set $C$, for example, is a remarkable paradox: it contains as many points as the entire real line, yet its Lebesgue measure is zero, $m(C) = 0$. It is an infinitely fine dust of points.

Now, here is a point of physical and logical intuition. If a region of space has zero volume, what is the volume of any part of it? It must also be zero. Lebesgue insisted that his theory of measure respect this intuition. If a set $B$ has measure zero, then any subset $N \subseteq B$ should also be considered measurable and have measure zero. This principle is called ​​completion​​.

The collection of ​​Lebesgue measurable sets​​, denoted $\mathcal{L}(\mathbb{R})$, is defined by taking all the Borel sets and adding to them all subsets of any Borel set of measure zero. This seems like a minor, logical bit of housekeeping. We are simply ensuring our system is tidy and complete. An immediate consequence is that since the Cantor set $C$ is a Borel set with $m(C)=0$, every single subset of the Cantor set is Lebesgue measurable.

At first glance, this "completion" step doesn't seem to have added anything fundamentally new. The new sets are just slivers of something that was already negligible. It’s natural to assume that these new, measurable pieces of dust must themselves be Borel sets that we just hadn't noticed before. For a long time, it was an open question whether the family of Borel sets and the family of Lebesgue measurable sets were, in fact, one and the same.

The answer to that question is a resounding no, and the proof is one of the most stunning arguments in mathematics, a true "ghost in the machine" discovery. The trick is not to try and build a non-Borel set, but to prove it must exist by comparing two different kinds of infinity.

First, let's count the Borel sets. The collection of open sets has a cardinality of $\mathfrak{c}$, the "cardinality of the continuum" (the number of points on the real line). By applying our construction rules (countable unions and complements) a countable number of times, and even extending this process through transfinite induction, one can show that the total number of sets you can possibly create is still just $\mathfrak{c}$. So, the cardinality of the Borel $\sigma$-algebra is $|\mathcal{B}(\mathbb{R})| = \mathfrak{c}$,. This is a colossal number, but it is the same order of infinity as the real line itself.

Now, let's return to our friend, the Cantor set $C$. It has measure zero, but it contains $\mathfrak{c}$ points. Now for the killer question: how many subsets does the Cantor set have? The set of all subsets of $C$ is its power set, $\mathcal{P}(C)$. By a fundamental theorem of Georg Cantor, the cardinality of a power set is always strictly greater than the cardinality of the original set. Therefore:

And we know that $2^{\mathfrak{c}} > \mathfrak{c}$. The number of subsets of the Cantor set belongs to a higher order of infinity than the number of points on the real line.

The final syllogism is as beautiful as it is devastating:

1. There are $2^{\mathfrak{c}}$ distinct subsets of the Cantor set $C$.
2. Since $m(C)=0$, every one of these $2^{\mathfrak{c}}$ subsets is Lebesgue measurable.
3. There are only $\mathfrak{c}$ Borel sets in the entire universe of real numbers.

Since $2^{\mathfrak{c}} > \mathfrak{c}$, there are vastly more subsets of the Cantor set than there are Borel sets in total. The inescapable conclusion is that there must exist subsets of the Cantor set that are Lebesgue measurable but are ​​not Borel sets​​,.

This is a profound revelation. Our "housekeeping" step of completion was no small thing; it tore open a portal to a new universe of sets. These sets are perfectly well-defined in terms of measure—they have size zero—but they are "unbuildable" from the basic blocks of open intervals. They are ghosts: we can prove they exist and even know where to find them (hiding inside any Borel set of measure zero, like the Cantor set), but we cannot construct them with the standard Borel toolkit. This implies that there must exist a Borel set (for example, the Cantor set itself, or even the whole real line $\mathbb{R}$) which contains a subset that is not Borel.

The cardinality argument is a magnificent proof of existence, but it feels a bit like a cosmic census report telling you ghosts exist without showing you one. Can we actually construct an example of a Lebesgue measurable, non-Borel set? The answer is yes, through a piece of mathematical artistry.

Consider a special function $f(x) = x + c(x)$, where $c(x)$ is the strange and beautiful Cantor-Lebesgue function, also known as the "devil's staircase." This function $f$ is a ​​homeomorphism​​, a kind of perfect, continuous transformation that stretches and bends the interval $[0,1]$ into $[0,2]$ without any tearing or gluing. A key property is that homeomorphisms are topologically faithful: they map Borel sets to Borel sets.

The magic of this particular function is how it treats the Cantor set. It takes the measure-zero dust of points that is $C$ and "stretches" it out, so that its image, $f(C)$, becomes a set with positive measure, $m(f(C)) = 1$.

Now for the clever trap:

1. Because the set $f(C)$ has positive measure, it is large enough to contain truly pathological sets. It is a known fact that we can construct a subset $N \subset f(C)$ that is ​​not Lebesgue measurable​​. (Such sets, like the Vitali set, are the canonical examples of "unmeasurable" sets).
2. Let's look at the origin of this troublemaker. Define the set $A = f^{-1}(N)$. This is the set of points in the original $[0,1]$ interval that get mapped into $N$.
3. Where does $A$ live? Since $N$ is a subset of $f(C)$, its preimage $A$ must be a subset of $C$.
4. Is $A$ Lebesgue measurable? Yes! It is a subset of the Cantor set $C$, and since $m(C)=0$, all of its subsets are Lebesgue measurable with measure zero.
5. Is $A$ a Borel set? Here's the contradiction. ​​Assume​​ $A$ is a Borel set. Since $f$ is a homeomorphism, it must map the Borel set $A$ to another Borel set, $f(A)$. But $f(A) = N$. This would mean $N$ is a Borel set. And since every Borel set is Lebesgue measurable, this would imply $N$ is Lebesgue measurable. But we chose $N$ specifically because it was not Lebesgue measurable!

Our initial assumption must be false. The set $A$ cannot be a Borel set.

And there we have it. The set $A$ is our concrete apparition: a Lebesgue measurable set that is provably not a Borel set. The study of these sets marks the transition from the relatively tame world of classical analysis to the wild and fascinating landscape of modern measure theory and descriptive set theory, where mathematicians continue to explore the intricate hierarchy of the definable and the measurable.

After our journey through the intricate construction of sets that are not Borel, a question naturally arises: "So what?" Are these sets merely esoteric oddities, mathematical monsters conjured by set theorists to haunt the dreams of students? Are they simply pathologies that demonstrate the limits of our intuition? It is a fair question. To a physicist or an engineer, a set so bizarre that it cannot be constructed in a finite number of steps might seem like a philosophical indulgence, far removed from the "real world."

The beautiful answer is that these sets are much more than that. Their existence is not a flaw in our mathematical landscape; it is a crucial feature. Non-Borel sets are not just there to break our tools. They are there to teach us the precise limits of those tools, forcing us to build sharper, more powerful ones. By studying them, we don't just learn about what can go wrong; we gain a profoundly deeper understanding of the structures that hold mathematics together, from the properties of continuous functions to the very foundations of probability theory. They mark the boundary between the "tame" and the "wild," and in doing so, reveal the true nature of both.

One of the first things we discover is that some of our most basic mathematical operations have a "civilizing" effect. They can take a wildly complicated non-Borel set and produce from it something perfectly simple and well-behaved.

Imagine you have a set $A$ that is not a Borel set. It's a truly intricate object, beyond any construction using countable unions, intersections, and complements of open sets. Now, let's perform a simple topological operation: find the closure of $A$, denoted $\mathrm{cl}(A)$. This is like taking our "monster" and filling in all the little gaps to make it solid, including all of its limit points. The result, $\mathrm{cl}(A)$, is by definition a closed set. And here is the magic: every closed set in $\mathbb{R}$ is a Borel set! The act of taking the closure has tamed the monster. Its "shadow" or outline is a well-behaved Borel set, no matter how pathological the original set was. The same remarkable thing happens if we consider the derived set, which is the set of all accumulation points of $A$. This set, too, is always closed and therefore always a Borel set. It seems that some fundamental topological operations cannot tolerate the complexity of a non-Borel set and collapse it into the Borel world.

This principle finds a powerful expression in the world of continuous functions. Continuous functions are, in many ways, the gold standard of "niceness" in analysis. A function is continuous if it doesn't have any sudden jumps or breaks. Surely, such a well-behaved function can only interact with well-behaved sets? This intuition is partially correct, in a very specific way. Consider the zero set of a continuous function $f: \mathbb{R} \to \mathbb{R}$, which is the set of all points $x$ where $f(x)=0$. This set is nothing more than the preimage of the singleton set $\{0\}$. Since $\{0\}$ is a closed set and $f$ is continuous, the very definition of continuity guarantees that its preimage must also be a closed set. As we've just seen, all closed sets are Borel sets. This leads to a stunning conclusion: the zero set of a continuous function is always a Borel set. A non-Borel set, in all its complexity, can never be described as the set of points where some continuous function vanishes. Continuity, in this sense, is blind to non-Borel sets; it cannot "pick them out" from its domain.

So, continuity tames non-Borel sets when we look at preimages. But what happens when we look at the forward direction? If we take a perfectly simple Borel set and apply a continuous function to it, must the result also be a simple Borel set? Our intuition, buoyed by the previous example, might scream "Yes!"

And our intuition would be spectacularly wrong.

This is one of the great surprising truths of modern analysis. While the preimage of a Borel set under a continuous function is always a Borel set, it can map a Borel set's image to a non-Borel set. This reveals a fundamental asymmetry in the nature of continuity. How is this possible? The secret lies in the interplay between dimensions.

Imagine you have a continuous function that can map a one-dimensional line segment onto a two-dimensional square, visiting every single point. Such "space-filling curves," like the Peano curve, actually exist. Now, within this 2D square, it is known that there are Borel sets (which are "well-behaved") whose projection, or shadow, onto one of the axes is a non-Borel set. Now, let's put the pieces together. Start with the Peano curve $p$, which maps the interval $[0,1]$ continuously onto the square $[0,1]^2$. Let's take that nasty 2D Borel set whose projection is non-Borel, and call it $B_{NB}$. The preimage of $B_{NB}$ under our Peano curve, let's call it $E = p^{-1}(B_{NB})$, is a subset of the original 1D interval. Since $p$ is continuous and $B_{NB}$ is Borel, the set $E$ must be a Borel set. So we have a perfectly "nice" Borel set $E$ on the line.

Now, what happens if we apply a new function, built by first applying our Peano curve $p$ and then projecting the result back down to the x-axis? This composite function, $f(x) = \pi_1(p(x))$, is a continuous map from the interval to itself. But what is the image of our Borel set $E$ under this function? It's $f(E) = \pi_1(p(E))$. Since the Peano curve is surjective, $p(E) = p(p^{-1}(B_{NB})) = B_{NB}$. So, $f(E)$ is just the projection of $B_{NB}$, which we know is not a Borel set! We have found a continuous function that maps a perfectly respectable Borel set to a non-Borel monster.

This is a profound result. It tells us that continuity, while preserving structure backwards, can create immense complexity forwards. It is not, however, the only story. Some continuous transformations are more "honest." A function like $f(x) = \sqrt{x}$ on $[0, \infty)$ is not just continuous; it is a homeomorphism. This means it has a continuous inverse, $f^{-1}(y) = y^2$. It's a true two-way street. Such functions preserve the Borel hierarchy perfectly; they map Borel sets to Borel sets and non-Borel sets to non-Borel sets, and vice versa. Similarly, if you take a non-Borel set $A$ on the real line and lift it to the diagonal in the plane, forming the set $S = \{(x,x) : x \in A\}$, this new set $S$ will also be non-Borel in the plane's Borel structure. This is because the map $x \mapsto (x,x)$ is a homeomorphism between the line and the diagonal. The geometric change is simple, so the structural complexity is preserved.

We have seen that non-Borel sets can be tamed or created. But what happens when we combine them? If we take a sequence of non-Borel sets, what can we say about their limit? If we intersect a non-Borel set with a transformed copy of itself, does the complexity compound or cancel? Here, our intuition is once again humbled.

Consider a sequence of non-Borel sets, $A_1, A_2, A_3, \dots$. Let's look at their limit superior, the set of points that belong to infinitely many of the $A_n$. If we take the simple case where all the $A_n$ are the same non-Borel set $A$, the limit superior is just $A$, which is non-Borel. This seems to suggest the pathology persists. But it's not so simple. We can cleverly construct a sequence of distinct non-Borel sets whose limit superior is a perfectly ordinary Borel set—say, the interval $[0,1]$. It is possible to create a parade of monsters that, in the limit, outline a simple, familiar shape.

The same unpredictability arises with even simpler operations. Take a non-Borel set $A$ and its reflection through the origin, $-A = \{-x : x \in A\}$. What is their intersection, $A \cap (-A)$? One can construct a non-Borel set $A$ that lies entirely on the positive real line; in that case, its intersection with its reflection is the empty set, which is trivially Borel. On the other hand, one can construct a symmetric non-Borel set where $A = -A$. In this case, the intersection is $A$ itself, which is not Borel. There is no general rule; the result of this simple, symmetric operation could be tame or could be just as wild as the original set.

At this point, one might be tempted to throw up one's hands and declare these sets to be pure chaos. They defy simple rules and confound our intuition. But this is where the story takes its most important turn. The existence of these sets did not lead mathematicians to despair; it led them to a deeper level of organization.

The analytic sets—those continuous images of Borel sets that we saw can be non-Borel—form a new, larger class. While they may not be Borel, they possess a property of immense practical importance: they are ​​universally measurable​​. This is a powerful idea. A set is universally measurable if it is measurable with respect to the completion of every possible probability measure you can define on the space.

Think about what this means for a field like stochastic processes or mathematical finance. In these fields, we model random phenomena. We often do not know the "true" probability distribution governing the system. We might have a whole family of possible probability models. If the events we care about correspond to sets that are universally measurable, then we are guaranteed to be able to assign a probability to them, no matter which model from our family turns out to be the right one. The analytic sets, including those non-Borel ones, fall into this wonderfully robust category. Even if we restrict ourselves to so-called atomless measures (which assign zero probability to single points, typical for continuous phenomena), the conclusion holds: there are non-Borel sets that are measurable under all such measures.

So, far from being useless pathologies, these non-Borel analytic sets are citizens of a larger, more accommodating kingdom. Their discovery didn't break measure theory; it enriched it, forcing the development of a more robust framework—the framework of universally measurable sets—that is precisely what is needed to handle the complexities of advanced probability theory.

Thus, our journey comes full circle. The strange sets that exist in the gaps of the Borel hierarchy are not just cautionary tales. They are signposts that point toward a richer, more powerful understanding of measurement, continuity, and probability. They show us the boundaries of our simplest ideas and, in doing so, invite us to build the more profound theories that lie beyond.