Subsets of Null Sets and the Completion of a Measure

In mathematics, how do we formalize the idea of a set being "insignificant" or having zero size? This is the role of null sets in measure theory. A powerful intuition suggests that any part of a truly insignificant set must also be insignificant. However, this fundamental idea is not automatically guaranteed in standard mathematical frameworks, creating a critical knowledge gap. This article addresses this problem by exploring the concept of complete measure spaces. First, in the "Principles and Mechanisms" chapter, we will delve into why some measure spaces are 'incomplete,' what that means, and how the elegant process of 'completion' fixes the issue. Following that, the "Applications and Interdisciplinary Connections" chapter will demonstrate why this completion is not just a theoretical nicety, but a practical necessity for the integrity of modern analysis, probability theory, and the modeling of physical phenomena.

Let’s begin with a simple, almost philosophical, question. What does it mean for something to be "small"? Not just small, but truly, fundamentally insignificant? In the world of geometry and measurement, we have a very precise notion for this: a set of ​​measure zero​​. Think of a single point on a line. It has no length. It’s there, but it takes up no space. Its length, or "measure," is zero. Think of a line drawn on a sheet of paper. It exists, but it has no area. Its area-measure is zero. These are what we call ​​null sets​​—sets that are negligible from the perspective of the measurement we are using.

Now, here's a beautifully intuitive idea: if a set is completely insignificant, shouldn't every single one of its parts also be insignificant? If a collection of dust motes has zero total volume, surely a smaller sub-collection of those same motes must also have zero volume. If a region of the real number line has zero total length, then any piece of that region, no matter how jagged or strangely defined, should also have zero length. This seems as self-evident as saying that a part cannot be greater than the whole. In mathematics, we cherish such intuitions, but we also put them to the test. And as we shall see, this one leads us to a fascinating and crucial feature of modern analysis.

To measure sets, we first need to agree on which sets are "measurable." We can't just measure any arbitrary collection of points; some are too pathological, too "badly behaved" to be assigned a consistent size. The collection of all the "good" sets—the ones we agree to measure—is called a ​​$\sigma$-algebra​​. You can think of it as a club. To be in the club, a set must play by certain rules: if a set is in, its complement must also be in; and if you take a countable number of sets from the club, their union must also be in the club. This ensures we have a robust system for building new measurable sets from old ones.

Now, let's return to our intuition. We have a null set $N$—a member of our $\sigma$-algebra with measure zero. And we have a subset $S$ that is contained within $N$. Our intuition screams that $S$ should also be measurable and have measure zero. But what if $S$ is simply not in our $\sigma$-algebra? What if it's not a member of the club? If that's the case, our system of measurement can't even assign it a measure! It's not that its measure is non-zero; it's that the question "What is the measure of $S$?" is meaningless in our framework.

When this happens—when there exists a subset of a null set that is not itself measurable—we say the measure space is ​​incomplete​​. It has a blind spot. It fails to live up to our most basic intuition about insignificance.

Let's make this concrete with a toy example. Imagine a universe $X = \{a, b, c\}$. Suppose our collection of measurable sets (our $\sigma$-algebra) is $\mathcal{M} = \{\emptyset, \{a\}, \{b, c\}, X\}$. We can only measure these four sets and nothing else. Let's define a measure $\mu$ where $\mu(\{a\}) = 1$ and $\mu(\{b, c\}) = 0$. The set $E = \{b, c\}$ is a null set. Now consider its subset $S = \{b\}$. It is clearly a part of a set with zero measure. But look at our $\sigma$-algebra! The set $\{b\}$ is not a member. We cannot measure it. Our system is blind to the set $\{b\}$, and therefore, our measure space is incomplete.

This isn't just a pathology of toy examples. A fundamentally important measure space, the ​​Borel sets​​ on the real line with the standard length measure, is incomplete. The Borel sets are, in a sense, all the "constructible" sets you can make starting from simple intervals. They form the bedrock of much of analysis. Yet, it's possible to find a Borel set $N$ with zero length ($\lambda(N)=0$) that contains a subset $S$ which is not a Borel set. The classic example involves the Cantor set, a fascinating fractal structure that is a Borel set of measure zero. Despite having zero length, it contains as many points as the entire real line, allowing for the existence of non-Borel subsets within it. This discovery was a profound moment, revealing that our initial "constructible" framework of Borel sets was not enough.

So what do we do when our system is incomplete? We fix it! If the problem is that some subsets of null sets are missing from our $\sigma$-algebra, the solution is beautifully simple: we add them. This process is called ​​completion​​.

The idea is to create a new, larger $\sigma$-algebra, called the ​​completed $\sigma$-algebra​​ $\overline{\mathcal{M}}$, by throwing in all the "missing" pieces. We start with our original $\sigma$-algebra $\mathcal{M}$ and augment it with every subset of every $\mu$-null set in $\mathcal{M}$. The result is the smallest new $\sigma$-algebra that contains our original one and satisfies our intuition about null sets. We then extend the measure $\mu$ to a new measure $\overline{\mu}$ on this larger collection of sets. How? We simply declare that all these new tiny pieces we've added have a measure of zero.

A measure space that has this property—that every subset of a null set is itself measurable (and thus has measure zero)—is called a ​​complete measure space​​. For a space to be complete, a set having an ​​outer measure​​ of zero is a sufficient condition for it to be measurable. The outer measure $\mu^{*}(A)$ is a way to estimate the size of any set $A$, whether it's in our $\sigma$-algebra or not, by seeing how efficiently we can cover it with measurable sets. Completeness means that if this best-effort estimate is zero, the set $A$ is not just negligible in size, it's officially a well-behaved measurable set.

The structure of the sets in this new, completed $\sigma$-algebra $\overline{\mathcal{M}}$ is wonderfully elegant. Any set $S$ in $\overline{\mathcal{M}}$ can be written in the form:

$S = E \cup N$

Here, $E$ is a set from our original "good" $\sigma$-algebra $\mathcal{M}$, and $N$ is a subset of some original null set $Z \in \mathcal{M}$ with $\mu(Z)=0$. You can think of it like this: every set in the completed world is just an original, well-behaved set ($E$) potentially "dusted" with some negligible garbage ($N$).

Let's see this in action. Consider a space $X = \{1, 2, 3, 4\}$ with the simple $\sigma$-algebra $\mathcal{M} = \{\emptyset, \{1, 2\}, \{3, 4\}, X\}$. Let's say $\mu(\{1, 2\}) = 0$ and $\mu(\{3, 4\}) = 5$. The set $\{1, 2\}$ is our null set. Its subsets $\{1\}$ and $\{2\}$ are not in $\mathcal{M}$. In the completion, we add them. What other new sets do we get? We get sets like $\{1, 3, 4\}$, which can be seen as the union of the original set $E = \{3, 4\}$ and the null subset $N = \{1\}$. The full completed $\sigma$-algebra becomes $\overline{\mathcal{M}} = \{\emptyset, \{1\}, \{2\}, \{1, 2\}, \{3, 4\}, \{1, 3, 4\}, \{2, 3, 4\}, X\}$. Each new set is just one of the old ones combined with a piece of the original null set $\{1, 2\}$.

Interestingly, this decomposition $S = E \cup N$ is not always unique. For the same completed set $S$, there might be different choices for the "original part" $E$. This doesn't cause any problems, but it's a subtle feature showing that the boundary between the original set and the null "dust" can be a little fuzzy.

The most powerful part of the completion process is how we define the measure of these new sets. For a set $S = E \cup N$, its completed measure is defined simply as:

$\overline{\mu}(S) = \mu(E)$

The null part $N$ contributes exactly zero to the measure, just as our intuition demanded! The "dust" is invisible to the measuring tape.

Let's consider a beautiful example. Suppose we're working on the interval $[0, 2]$. Let our measurable sets be just $\emptyset, [0, 1], (1, 2], [0, 2]$, and define a measure where $\mu([0, 1]) = 0$ and $\mu((1, 2]) = 7$. Now consider the strange set $S = (\mathbb{Q} \cap [0, 1]) \cup (1, 2]$, which is the set of all rational numbers between 0 and 1, combined with the interval from 1 to 2. The set of rationals $\mathbb{Q} \cap [0, 1]$ is a messy, porous set. But it is a subset of $[0, 1]$, which is a null set in our space. So, $\mathbb{Q} \cap [0, 1]$ is our "dust" $N$. The set $(1, 2]$ is our "original part" $E$. To find the measure of $S$ in the completed space, we simply ignore the dust: $\overline{\mu}(S) = \mu((1, 2]) = 7$. This principle allows us to handle incredibly complex sets with ease, as long as their complexity is confined to a region of measure zero. This happens all the time in probability and analysis, where we often need to know that a property holds "almost everywhere"—that is, everywhere except on a set of measure zero. Completeness guarantees we don't have to worry about the pathological structure of that exceptional set.

This brings us to the triumphant conclusion of our story: the ​​Lebesgue measure​​. When you hear mathematicians talk about "measure theory," they are most often implicitly talking about the Lebesgue measure on the real line. The genius of Henri Lebesgue was to formalize this very process. The Lebesgue $\sigma$-algebra $\mathcal{L}$ is precisely the ​​completion​​ of the Borel $\sigma$-algebra $\mathcal{B}(\mathbb{R})$.

By constructing it this way, the Lebesgue measure is complete by definition. Any subset of a Lebesgue null set is a Lebesgue measurable set with measure zero. This resolves the "wrinkle" we found in the Borel sets and fully restores our initial, powerful intuition. It gives us a system of measurement on the real line that is both incredibly powerful, able to measure a vast collection of sets, and robust, behaving exactly as we would hope when it comes to negligible quantities. It is this property of completeness that makes the Lebesgue integral so much more powerful and flexible than its predecessor, the Riemann integral, and makes it the indispensable language of modern probability theory, functional analysis, andpartial differential equations. What starts as a simple, intuitive demand—that parts of nothing should be nothing—blossoms into one of the most profound and useful constructs in all of mathematics.

It is a curious and profoundly important fact that the elegant world of Borel sets—the mathematical structure we build from simple open intervals—is, in a crucial sense, incomplete. If you think of the Borel sets as a fantastically detailed map of the real number line, it contains every city, town, and highway you could imagine. But it is missing something. It is missing the infinitely many tiny, dusty footpaths that lie within properties so small they have zero area. For a casual traveler, this omission is harmless. But for the detective, the physicist, or the financier—anyone whose work depends on tracking every possibility, no matter how remote—this incomplete map is a liability. The journey to “complete” this map, by adding all these dusty footpaths, takes us from the Borel sets to the Lebesgue measurable sets. This is not just a pedantic exercise for mathematicians. As we shall see, completing the picture is essential for the foundations of analysis, probability theory, and the modeling of the physical world.

Let's start with the most direct consequence. Why aren't the Borel sets enough? The reason is subtle but beautiful. Consider the famous Cantor set, constructed by repeatedly removing the middle third of intervals starting with $[0,1]$. What remains is a "dust" of points. This "dust," the Cantor set $C$, is a closed set, so it is certainly a Borel set. And remarkably, its total length—its Lebesgue measure—is zero. It’s a null set.

Now, how many points are in this dust? Uncountably many! In fact, there are just as many points in the Cantor set as there are on the entire real line. So, what about the subsets of the Cantor set? The collection of all its subsets is enormous, with a cardinality of $2^{\mathfrak{c}}$. In contrast, the collection of all Borel sets has a much smaller cardinality, $\mathfrak{c}$. This simple fact of different-sized infinities leads to an inescapable conclusion: there must be an immense number of subsets of the Cantor set that are not Borel sets.

Here is the problem: we have a perfectly good Borel set $C$ with measure zero, yet it contains subsets whose "measurability" is undefined in the Borel world. The completion of the measure solves this elegantly. The new, larger collection of Lebesgue measurable sets, let's call it $\mathcal{L}$, is defined to include the Borel sets $\mathcal{B}$ plus all subsets of any Borel null set. In this completed world, every subset of the Cantor set now becomes a member of $\mathcal{L}$, and each is assigned the only sensible measure it could have: zero. In fact, we can characterize any Lebesgue measurable set as being the union of a Borel set and a subset of some null set—it’s a major highway from our old map plus some of that fine dust we just added.

This richer collection of sets naturally gives us a richer collection of functions. We can now construct functions that are Lebesgue measurable but not Borel measurable. A stunning example involves composing the strange Cantor-Lebesgue function—which maps the measure-zero Cantor set onto the entire interval $[0,1]$—with the indicator function of a non-measurable set. The resulting function has the curious property that its preimages are Lebesgue measurable, but not all of them are Borel sets, making the function itself a creature that can only exist in our completed world.

This new class of functions is not just a cabinet of curiosities. It is, in fact, the most natural class of functions for modern analysis. Lusin's theorem tells us that a function is Lebesgue measurable if and only if it is "nearly" continuous; that is, you can find a continuous function that agrees with it everywhere except on a set of arbitrarily small measure. The ability to quantify that "small set" depends crucially on the completeness of our measure.

When we step into the world of probability, measure becomes probability, and measurable sets become events. An incomplete measure space means there are potential outcomes of an experiment—subsets of a zero-probability event—for which we cannot even ask what their probability is. This is an untenable situation for a theory supposed to handle uncertainty!

The completion ensures that if an event $N$ has probability zero, then any sub-event $A \subseteq N$ is also an event and has probability zero. This is not just a matter of convenience; it is a matter of consistency. There exist pathological sets, like the Vitali set, which cannot be assigned a measure without violating fundamental principles like translation invariance. However, these are different beasts. They are not subsets of null sets. By completing our space, we deal with the vast family of subsets of null sets in the only logical way possible, without running into the paradoxes of sets like Vitali's.

This has profound consequences for more advanced concepts, like conditional expectation. The conditional expectation, $E[X | \mathcal{G}]$, represents the best possible guess for a random variable $X$ given only the information contained in a sub-$\sigma$-algebra $\mathcal{G}$. What happens if we complete $\mathcal{G}$ to $\overline{\mathcal{G}}$? The completion adds all the subsets of null events in $\mathcal{G}$. This means $\overline{\mathcal{G}}$ contains more information—it allows us to distinguish between outcomes that were previously lumped together inside a single null event. A beautiful, concrete example shows that $E[X | \mathcal{G}]$ and $E[X | \overline{\mathcal{G}}]$ can be different. The latter, having access to more information, can provide a more refined, accurate prediction of $X$. Ignoring the completion means we are literally throwing away information.

Perhaps the most dramatic application of completing the measure appears when we try to model phenomena that evolve continuously in time, like the jiggling path of a pollen grain in water—Brownian motion. We want to define a probability measure on the space of all possible paths a particle could take. The monumental Kolmogorov extension theorem provides a way to do this. It builds a probability measure, but on a rather limited $\sigma$-algebra generated by "cylinder sets," which are sets defined by the particle's position at a finite number of time points.

This presents a colossal problem. With this limited structure, we can answer questions like, "What is the probability the particle is at position $x_1$ at time $t_1$ and at position $x_2$ at time $t_2$?" But we cannot answer the most natural question of all: "What is the probability that the particle's path is continuous?" The property of being continuous depends on the particle's position at all uncountably many moments in time. The set of all continuous paths is simply not in the cylinder $\sigma$-algebra provided by Kolmogorov's theorem.

The theory seems to be at a dead end. How can we talk about a continuous random walk if the very event of "being continuous" is not recognized by our probability measure?

The answer, once again, is to complete the picture. While the set of continuous paths is not a "Borel set" of our function space, it turns out to be a "Lebesgue measurable set." It lives inside the completion of the Kolmogorov measure space. Thanks to this completion, we can rigorously prove that the set of continuous paths has probability 1. This legitimizes the entire physical model of Brownian motion and other stochastic processes as phenomena with well-behaved sample paths.

This principle is absolutely fundamental to the modern theory of stochastic differential equations (SDEs), which are used to model everything from stock prices to neuronal firing. The famous Itô integral, the cornerstone of stochastic calculus, is defined for a class of processes whose measurability assumes a complete probability space—it’s one of the "usual conditions" that underpins the entire theory. Without completing the measure, the mathematical machinery for a massive portion of modern science and finance would rest on shaky foundations.

So, we have seen that completing the measure is not a mere technicality. It expands the analyst's world of functions, solidifies the probabilist's logic, and makes the physicist's models of continuous random phenomena rigorous.

But here is the final, beautiful twist in the story. We started with the Borel sets and their functions, formed the $L^p$ spaces of functions used in physics and engineering, and then moved to the much larger completed space of Lebesgue measurable sets. It seems like we have added a universe of new, complicated objects. But what happens to our $L^p$ spaces? Do they become unrecognizably different? The answer is a resounding no. The space $L^p$ built on the Borel sets is perfectly, isometrically isomorphic to the $L^p$ space built on the much larger collection of Lebesgue sets. Functionally, they are identical.

It's as if we took our good, but incomplete, map and painstakingly drew in every last footpath and alleyway. The resulting map is now perfect; every location is accounted for. And yet, the total area of the country, its essential geometry and structure, has not changed one bit. We simply see it now with perfect clarity. That is the subtle power and inherent beauty of completing the measure.