Lebesgue-Stieltjes measure

How do we measure things? While a ruler measures length and a scale measures weight, mathematics and science often require a more abstract and versatile concept of "measure." We need a single, unified framework that can handle the counting of discrete objects (like point charges), the integration of continuous quantities (like temperature across a surface), and even bizarre, fractal-like distributions that fit neither category. The lack of such a tool creates a disconnect between the worlds of discrete sums and continuous integrals, forcing us to use different methods for problems that feel conceptually similar.

This article introduces the powerful solution to this problem: the Lebesgue-Stieltjes measure. It is a profound generalization of length that provides a universal language for measurement. Over the next two chapters, you will embark on a journey to understand this remarkable concept. First, in "Principles and Mechanisms," we will lift the hood to see how the measure is constructed from a "generating function" and how it naturally decomposes into discrete, continuous, and singular parts. Subsequently, in "Applications and Interdisciplinary Connections," we will explore its immense utility, discovering how it unifies sums and integrals and provides the very foundation for modern probability theory. Let's begin by exploring the core machinery that makes this all possible.

Now that we have a taste for what the Lebesgue-Stieltjes measure is, let’s roll up our sleeves and look under the hood. How does this machine actually work? Like all great ideas in physics and mathematics, it starts with a simple, almost playful, modification of something we already know. We’ll begin by reimagining our concept of "length" and, in doing so, uncover a rich structure of different kinds of measures—some familiar, some surprisingly strange.

Think about how you measure length. You take a ruler, a straight edge with uniform markings, and you read off the numbers. The length of the interval from point $a$ to point $b$ is simply $b-a$. This is the essence of the ordinary Lebesgue measure. It's built on the function $F(x) = x$. The length of $(a, b]$ is $F(b) - F(a) = b-a$. Simple enough.

But what if our ruler wasn't uniform? What if it was made of a strange material that was stretched in some places and compressed in others? This is the core idea of the Lebesgue-Stieltjes measure. We replace the simple function $F(x) = x$ with a more general function, which we'll call the ​​distribution function​​, or ​​generating function​​, $F(x)$. The only rules are that this function can't decrease (our ruler can't have negative length) and it must be right-continuous (a technical detail to keep things tidy).

With this new "ruler" $F(x)$, the measure of an interval $(a, b]$ is defined as:

This little change has enormous consequences. It allows us to define "length" in incredibly flexible ways. But what part of the function $F$ really matters? Suppose you have two such rulers, described by functions $F_1(x)$ and $F_2(x)$, and they are identical except that one is shifted up; that is, $F_2(x) = F_1(x) + c$ for some constant $c$. What happens to the measures they generate? Let’s compute the measure of an interval $(a, b]$ with the second ruler:

They are exactly the same! This simple calculation reveals a deep truth: the absolute value of the generating function is irrelevant. The measure is encoded entirely in the differences—the way the function changes from point to point. Shifting the entire ruler up or down doesn't change any of the lengths you measure with it. It is the slope, the rate of change, the jumps in $F$ that hold all the information. This is our first clue to the rich world we're about to enter.

So, the changes in $F$ are what matter. Let's explore this with a peculiar kind of ruler—one that doesn't stretch smoothly at all. Imagine a function that stays flat, then suddenly jumps up, stays flat, and jumps again. A perfect example is the floor function, $F(x) = \lfloor x \rfloor$ (for $x \ge 0$). This function is constant on intervals like $[0, 1)$, $[1, 2)$, etc., and it jumps by exactly 1 at every positive integer.

What kind of measure does this staircase-like function generate? Let's try to measure a single point, say the point $\{x_0\}$. How can we measure a single point? We can think of it as an infinitesimally small interval. Let's take the interval $(x_0 - \epsilon, x_0]$ and see what happens as $\epsilon$ shrinks to zero:

This limit is precisely the definition of the jump size in the function $F$ at the point $x_0$. We call it $F(x_0) - F(x_0^-)$, where $F(x_0^-)$ is the value $F$ is approaching from the left.

For our function $F(x) = \lfloor x \rfloor$:

• If $x_0$ is not an integer (say, $x_0=3.5$), then for small $\epsilon$, both $F(3.5)$ and $F(3.5 - \epsilon)$ are equal to 3. The jump is $3 - 3 = 0$. The measure is zero.
• If $x_0$ is a positive integer (say, $x_0=4$), then $F(4) = 4$, but as we approach from the left, $F(4-\epsilon) = 3$. The jump is $4 - 3 = 1$. The measure is one!

This is fantastic! Our measure $\mu_F$ is zero everywhere except at the positive integers, where it has a concentrated "lump" of measure equal to 1. A point with a positive measure is called an ​​atom​​ of the measure. For a measure generated by a step function like this, the atoms are precisely the points of discontinuity.

So, for $F(x) = \lfloor x \rfloor$, the "length" of any set $S$ is simply the number of positive integers inside $S$. We've created a "counting measure" out of thin air, just by choosing the right generating function. This type of measure, made up entirely of atoms, is called a ​​discrete measure​​. It's the first fundamental type of measure we've discovered.

What if our function $F$ has no jumps? That is, what if $F$ is continuous? You might think that this guarantees the measure behaves "nicely," perhaps like a stretched version of the ordinary Lebesgue measure. And sometimes, you'd be right.

Consider the case where $F(x)$ is not just continuous, but also smoothly differentiable. For example, let's look at the function in problem, whose derivative is $F'(x) = \arctan(x) + \frac{\pi}{2}$. For such a function, the measure of a small interval $(x, x+dx]$ is approximately $F(x+dx) - F(x) \approx F'(x)dx$. This suggests that the "density" of our new measure at a point $x$ is just $F'(x)$. The total measure of a set $A$ would then be:

where $d\lambda(x)$ is just the standard length element $dx$. This is a beautiful result. It connects our new framework directly back to standard calculus. This kind of measure, which has a density function with respect to the Lebesgue measure, is called ​​absolutely continuous​​. The name comes from a deeper property of the function $F$ itself: if $F$ is what mathematicians call "absolutely continuous" (a stronger condition than mere continuity), it generates an absolutely continuous measure.

This seems like the whole story for continuous functions. But measure theory has a beautiful, ghostly surprise in store for us. Is it possible for a generating function $F$ to be continuous (so there are no jumps, no atoms), yet the measure it generates is not absolutely continuous?

The answer is a resounding yes, and the canonical example is the famous ​​Cantor function​​, or "devil's staircase." Let's sketch its construction. You start with $F(0)=0$ and $F(1)=1$. You remove the middle third of the interval $[0, 1]$, which is $(\frac{1}{3}, \frac{2}{3})$, and you define $F(x)$ to be constant on this interval, with the value $\frac{1}{2}$. Then you take the two remaining intervals, $[0, \frac{1}{3}]$ and $[\frac{2}{3}, 1]$, and repeat the process. You remove their middle thirds and define $F$ to be constant there as well. You continue this forever.

The resulting function $F(x)$ is a marvel. It is continuous everywhere—no jumps! But it only increases on a strange, dust-like set called the Cantor set, which has a total Lebesgue length of zero. Everywhere else, the function is flat, meaning its derivative $F'(x)$ is 0 almost everywhere. If its measure were absolutely continuous, its density would be $F'(x)=0$, and its total measure $\int_0^1 0\ dx$ would be 0. But we know its total measure is $F(1)-F(0)=1$. This is a paradox!

The resolution is that the measure $\mu_F$ lives entirely on the Cantor set, a set that the Lebesgue measure considers to have zero size. The two measures are "mutually singular"—they live in different worlds. This third type of measure is called ​​singular continuous​​. It has no atoms, but it has no density either. It is a ghost in the machine, a kind of measure that standard calculus could never dream of.

So far our journey has been one of divergence. We've discovered three fundamentally different species of measure:

1. ​​Discrete (or atomic) measure​​, generated by the jumps in $F$.
2. ​​Absolutely continuous measure​​, generated by the "smoothly rising" parts of $F$, with a density $F'$.
3. ​​Singular continuous measure​​, the ghostly measure generated by continuous but "non-smooth" parts of $F$, like the Cantor function.

This seems like a zoo of disparate creatures. But the final, beautiful conclusion of this story is one of profound unity. The ​​Lebesgue Decomposition Theorem​​ tells us that any Lebesgue-Stieltjes measure $\mu$ can be uniquely written as a sum of these three pure types:

Every measure is a chord composed of these three fundamental notes. We can see this decomposition beautifully if we construct the right generating function. Consider a function like $F(x) = x^2 + \lfloor 2x \rfloor$ on the interval $[0, 1]$. We can literally see the decomposition in the function itself.

• The $x^2$ part is a smooth, differentiable function. It will generate the ​​absolutely continuous part​​, $\mu_{ac}$, with a density of $(x^2)' = 2x$.
• The $\lfloor 2x \rfloor$ part is a step function. It has jumps at $x=1/2$ and $x=1$. This part generates the ​​discrete part​​, $\mu_d$, consisting of two atoms.

There is no singular continuous part in this particular example, so the measure is a mix of just the first two types. More complex functions can contain all three.

This decomposition isn't just an abstract curiosity; it's an incredibly powerful tool for computation. Suppose we want to calculate an integral with respect to a mixed measure, like $\int_{\mathbb{R}} x \, d\mu(x)$ from problem. Thanks to the decomposition, we can break the problem down:

1. We integrate over the absolutely continuous part using its density: $\int x \cdot F'(x) \, dx$.
2. We sum over the discrete part by taking the value of the function ($x$) at each atom and multiplying by the mass of that atom (the jump size of $F$).

The total integral is simply the sum of these parts. What was once an intractable problem becomes a straightforward exercise. The Lebesgue decomposition provides a universal recipe for handling any measure, no matter how complicated its generating function may be. It transforms a seeming chaos of different behaviors into a simple, elegant, and unified structure. This is the inherent beauty of the Lebesgue-Stieltjes framework: it provides a single, powerful language to describe a vast universe of measuring, from counting discrete objects to analyzing the smoothest of continua, and even to navigating the strange, fractal landscapes in between.

Now that we have grappled with the machinery of the Lebesgue-Stieltjes measure, you might be asking a fair question: "What is all this for?" We have built a rather elaborate new kind of ruler. A standard ruler measures length. But what does our ruler measure, and why go to all the trouble of defining it in such a peculiar way, using this "generating function" $F(x)$?

The answer is that we have invented a tool of astonishing versatility. It is a universal ruler that can measure not just length, but also mass, charge, and, most importantly, probability, even when they are distributed in the most bizarre and counter-intuitive ways imaginable. This chapter is a journey through the landscapes—some familiar, some strange—that our new tool allows us to explore. We will see how this single, elegant idea unifies concepts that seemed disparate, solves problems that were once awkward, and reveals a hidden structure to the world of mathematics and physics.

Let's start with a simple, almost childlike question. What is the difference between adding up a list of numbers and finding the area under a curve? One is a sum, $\sum$, the other an integral, $\int$. They seem like fundamentally different operations. But with our new perspective, we see they are just two faces of the same coin.

Imagine a physicist wants to describe a series of point charges placed on a line. Let's say there's a charge of 1 unit at $x=1$, another at $x=2$, and so on. How could we describe this "distribution" of charge? The Lebesgue-Stieltjes framework gives us a beautiful way to do this. We can choose a generating function $F(x)$ that "jumps" at each of these locations. A perfect candidate is the floor function, $g(x) = \lfloor x \rfloor$, which increases by exactly 1 at every integer.

Now, suppose we want to calculate some quantity that depends on the position of these charges, say, the total potential energy, which might involve integrating a function like $f(x)=x^3$ against this charge distribution. We would write the integral $\int f(x) dg(x)$. What happens when our machinery gets to work? The integral, this seemingly complex continuous object, sees that the "ruler" $g(x)$ is constant everywhere except at the integers. It recognizes that the only places that can contribute to the total are the points where $g(x)$ jumps. The result? The integral magically transforms into a simple sum over the values of $f(x)$ at the integer points, weighted by the size of the jump (which is just 1 in this case). The continuous integral becomes a discrete sum: $\int_{[0,4]} x^3 d\lfloor x \rfloor = 1^3 \cdot (1) + 2^3 \cdot (1) + 3^3 \cdot (1) + 4^3 \cdot (1)$ This is the essence of the calculation in. The Lebesgue-Stieltjes integral doesn't just calculate an area; it "probes" the structure of the measure. If the measure is a collection of discrete points, the integral naturally becomes a sum.

This idea is not limited to a finite number of points. We could construct a measure from an infinite series, placing smaller and smaller masses at an infinite sequence of points converging to zero. Our integral is still perfectly well-behaved; it simply becomes an infinite series. This is the first glimpse of the unifying power of our new tool: it sees no fundamental difference between the discrete and the continuous.

The most profound and widespread application of the Lebesgue-Stieltjes measure is in the field of probability. In fact, it is the very language of modern probability theory.

Every student of statistics learns about the Cumulative Distribution Function, or CDF. For a random variable $X$, its CDF, $F(x)$, gives the probability that $X$ will take on a value less than or equal to $x$. That is, $F(x) = P(X \le x)$. This function $F(x)$ is non-decreasing, right-continuous, and it runs from 0 to 1. Sound familiar? It's a perfect candidate for a generating function!

The Lebesgue-Stieltjes measure $\mu_F$ generated by a CDF is, in fact, the ​​probability distribution​​ itself. The measure of an interval $(a, b]$ is $\mu_F((a, b]) = F(b) - F(a)$, which is precisely $P(a \lt X \le b)$. Calculating the total measure of the entire real line, as in the case of the Cauchy distribution, confirms that the total probability is 1, just as it must be.

The real beauty here is how this framework effortlessly handles all types of random variables.

• For a ​​discrete random variable​​ (like the roll of a die), its CDF is a step function. The jumps correspond to the points with non-zero probability, and the size of the jump is the probability. The integral $\int x \, dF(x)$ becomes a sum, $\sum_k x_k P(X=x_k)$, which is the definition of the expected value!
• For a ​​continuous random variable​​ (like the height of a person), its CDF is a smooth, continuous function. Its derivative, $F'(x)$, is the familiar Probability Density Function (PDF), $f(x)$. The integral $\int x \, dF(x)$ becomes the standard integral $\int x f(x) \,dx$, which is the definition of expected value for a continuous variable.

But what about something in between? What if a random variable has a chance of taking a specific value, but can also fall within a continuous range? For example, the amount of rainfall on a given day might be exactly 0 with some probability, but if it's not 0, it could be any positive value described by a density. Such a "mixed" distribution would be awkward to handle with separate tools for discrete and continuous cases.

For the Lebesgue-Stieltjes integral, this is no problem at all. As we saw in a hypothetical scenario, if the generating function $F(x)$ has both smooth parts and jumps, the integral automatically and correctly decomposes. It becomes the sum of two pieces: a standard integral over the parts where a density exists, and a sum over the jump points. This is the power of a unified theory. It doesn't care if a distribution is discrete, continuous, or a mix of both; the definition of the integral $\int g(x) \, dF(x)$ gives the correct expectation $E[g(X)]$ in all cases.

So, we have discrete measures (sums) and absolutely continuous measures (integrals with a density). Is that all there is? Is every distribution either a collection of points or a smooth smearing, or a mixture of the two? For a long time, mathematicians thought so. But nature, and mathematics, is more imaginative than that.

Enter the Cantor set. You construct it by taking the interval $[0,1]$ and repeatedly removing the open middle third of every segment. What's left is a strange, disconnected "dust" of points. This set has a total length of zero, yet it contains more points than all the integers and rational numbers combined—it is uncountably infinite.

Now, one can define a function, the Cantor function $c(x)$, that is continuous and non-decreasing, goes from 0 to 1, yet is flat everywhere except on this dusty Cantor set. This function generates a Lebesgue-Stieltjes measure, $\mu_c$. What kind of measure is this?

• It is ​​not discrete​​, because the Cantor function is continuous, so there are no jumps. The measure of any single point is zero.
• It is ​​not absolutely continuous​​, because its derivative is zero almost everywhere. It has no density function you can write down and integrate.

This is a new beast, a third fundamental type of measure: a ​​singular continuous measure​​. It assigns its entire mass of 1 to the Cantor set, a set of Lebesgue measure zero! It is as if you have a pound of dust, but the dust is so fine that it occupies no volume.

This might seem like a pathological "monster" of interest only to mathematicians. But these ideas have found their way into physics, describing phenomena like chaotic dynamics and the energy spectra of quasicrystals. And our Lebesgue-Stieltjes framework can handle it perfectly. We can compute integrals against this strange measure, finding moments and expected values. We can even perform elegant calculations, like integrating the Cantor function against its own measure, revealing the beautifully simple result of $1/2$.

This leads us to a grand, unifying statement: the ​​Lebesgue Decomposition Theorem​​. It tells us that any probability distribution can be uniquely written as a sum of three parts: a discrete part, an absolutely continuous part, and a singular continuous part. An example that explicitly combines an absolutely continuous part with the singular Cantor measure shows how the integral naturally splits to handle this decomposition. Our framework provides a complete classification of the ways probability can be distributed.

Just to see how subtle and powerful this thinking is, consider the set of rational numbers $\mathbb{Q}$, which are famously dense—between any two real numbers, there's a rational one. What if we try to integrate a function which is 1 on the rationals and 0 elsewhere (the Dirichlet function) with respect to the Cantor measure? You might think that since the rationals are "everywhere," the integral should pick up something. But the Cantor measure is continuous, meaning the measure of any single point is zero. Since the rationals are a countable set of points, their total measure under the Cantor measure is a sum of infinitely many zeros, which is zero. The integral is zero. The Cantor measure manages to lay all of its "mass" down on the interval $[0,1]$ while completely avoiding every single rational number!

The ideas we've discussed extend even further. We started by defining our measure with respect to the standard notion of length on the real line. But what if we want to compare two arbitrary measures, neither of which is the standard one?

Suppose we have two different distributions of mass, $\mu_F$ and $\mu_G$, generated by functions $F$ and $G$. The ​​Radon-Nikodym Theorem​​ gives us a way to define the "density" of one with respect to the other. This "density," or Radon-Nikodym derivative $\frac{d\mu_F}{d\mu_G}$, acts like a conversion factor between the two measures. This concept is the mathematical engine behind many advanced topics in science and finance. It allows statisticians to compare different hypothetical models for data and physicists to relate the behavior of a system under different external conditions.

Finally, a word on why this framework has superseded older ones. The Riemann-Stieltjes integral, an earlier attempt to do something similar, exists only under much stricter conditions. For a function to be integrable with respect to the Cantor measure in the Riemann sense, for instance, it has to be continuous at most points of the Cantor set itself. The Lebesgue-Stieltjes integral, however, is far more robust; it happily exists for a much wider class of functions. In a world where the functions that model reality are often "rough" and not perfectly smooth, the Lebesgue-Stieltjes integral is the powerful, reliable tool that a working scientist needs.

From unifying sums and integrals to providing the very foundation of probability and revealing the existence of strange singular measures, the Lebesgue-Stieltjes integral is far more than a technical curiosity. It is a profound enlargement of our ability to measure and to reason, a language that brings clarity and unity to a vast range of human inquiry.