Unbounded Variation

In the world of mathematics, we often begin by studying smooth, predictable functions—curves that can be drawn without lifting a pen and whose behavior at any point can be described by a simple tangent line. But what happens when we venture beyond this orderly landscape? Is it possible for a continuous path, packed into a finite space, to have an infinite length, like a coastline whose complexity increases the closer you look? This paradox lies at the heart of the concept of ​​unbounded variation​​. It addresses the critical knowledge gap between the "tame" functions of classical calculus and the "wild," infinitely jagged reality found in nature and finance.

This article provides a journey into this fascinating topic. First, in the "Principles and Mechanisms" chapter, we will precisely define what it means for a function to have bounded or unbounded variation, exploring the line where smooth functions end and infinitely complex ones begin. We will construct these mathematical creatures and meet their most famous representative: the path of Brownian motion. We will discover how its unique properties break the rules of ordinary calculus. Following this, the "Applications and Interdisciplinary Connections" chapter will reveal how this seemingly abstract idea is a cornerstone of modern science, from the limits of signal processing to the very foundations of stochastic calculus and the modeling of financial markets.

Imagine walking along a path. Some paths are smooth and gently rolling; if you were to measure your total distance traveled, you'd get a perfectly reasonable, finite number. Now, imagine trying to measure the coastline of Norway. As you zoom in, you find more and more intricate fjords and inlets. The closer you look, the longer the coastline seems to get. It’s as if the path has an infinite amount of "jiggle" packed into a finite space.

In mathematics, we have a precise way to talk about this "jiggleness": the concept of ​​total variation​​. For any function $f(x)$ on an interval $[a, b]$, we can approximate its path length. We pick a set of points along the x-axis, say $a = x_0 x_1 \dots x_n = b$, and connect the corresponding points on the graph, $(x_i, f(x_i))$, with straight lines. The length of this "connect-the-dots" path is simply the sum of the absolute changes in height: $\sum_{k=1}^{n} |f(x_k) - f(x_{k-1})|$. The ​​total variation​​, $V_a^b(f)$, is the ultimate, most detailed path length we can find—it's the supremum (the least upper bound) of these sums over all possible choices of points.

If this total variation $V_a^b(f)$ is a finite number, we say the function is of ​​bounded variation​​. This means the total "up and down" travel of the function is limited. Monotonic functions (ones that only ever go up or only ever go down) are the simplest examples; their total variation is just the total change from start to finish, $|f(b) - f(a)|$. But many other functions, like a smooth sine wave or a parabola, also have bounded variation.

There is a beautiful, deeper meaning here. The ​​Jordan Decomposition Theorem​​ tells us that a function has bounded variation if and only if it can be written as the difference of two non-decreasing (always increasing or staying flat) functions, say $f(x) = g(x) - h(x)$. You can think of this as separating the function's journey into a total "up" trip ($g(x)$) and a total "down" trip ($h(x)$). Functions of bounded variation are "tame" enough to allow this separation. Their jiggles, while perhaps numerous, are not infinitely wild.

So, could a function that is continuous—an unbroken, single curve you can draw without lifting your pen—have infinite total variation? It seems paradoxical. How can you pack an infinitely long path into a finite segment of the plane without breaking the curve?

Let's build such a creature. Consider a function that tries to wiggle faster and faster as it approaches a point, say, $x=0$. A function like $\sin(1/x)$ does this beautifully. As $x$ gets smaller, $1/x$ rockets to infinity, and the sine function oscillates between $-1$ and $1$ with ever-increasing frequency. The problem is, $\sin(1/x)$ isn't continuous at $x=0$; its limit doesn't exist.

To tame it, let's multiply it by a term that goes to zero, like $x^p$. This gives us functions of the form $f(x) = x^p \sin(1/x^q)$ (and we'll define $f(0)=0$ to plug the hole). Here we have a battle: the term $x^p$ tries to "dampen" the oscillations, squashing them to zero, while the $\sin(1/x^q)$ term tries to wiggle infinitely fast. Who wins?

The answer depends on the exponents $p$ and $q$. Through careful analysis, we find a simple, elegant rule: the function is of ​​unbounded variation​​ if the damping power is less than or equal to the oscillation power, that is, $p \le q$.

Consider the famous borderline case: $f(x) = x \sin(1/x)$, where $p=q=1$. This function is continuous everywhere, including at $x=0$ where it's squeezed to zero. Yet, the damping is just barely insufficient. If you try to sum the lengths of its wiggles, the sum adds up like the harmonic series ($1 + 1/2 + 1/3 + \dots$), which famously diverges to infinity. So, here we have it: a perfectly continuous curve with an infinite path length on a finite interval. It's a function so "jagged" at the microscopic level that it cannot be decomposed into a simple "up" trip and "down" trip; it does not have a Jordan decomposition.

You might think these functions are just mathematical curiosities, oddities confined to the chalkboard. But nature is full of such beautiful complexity. The classic example is ​​Brownian motion​​, the random, jittery dance of a pollen grain suspended in water, kicked about by invisible water molecules.

The path of this pollen grain can be modeled by a mathematical object called a ​​Wiener process​​, which we'll denote by $W_t$. Its sample paths are, with probability one, continuous everywhere. Yet, they are also of ​​unbounded variation​​ on any time interval, no matter how small. A Brownian path is not just a little jagged; it is infinitely, relentlessly jagged, everywhere.

The secret to this infinite ruggedness lies in its fundamental ​​scaling law​​. For a "normal," smooth, differentiable function $f(t)$, if we look at a tiny time step $\Delta t$, the change in the function's value is proportional to the time step itself: $\Delta f \approx f'(t)\Delta t$. If you halve the time step, you halve the change.

Brownian motion is radically different. The change in position over a small time step $\Delta t$, which we write as $\Delta W_t = W_{t+\Delta t} - W_t$, is a random variable. Its average value is zero, but its typical size, measured by its standard deviation, is proportional to the ​​square root​​ of the time step: $\sqrt{\Delta t}$. If you cut the time step by a factor of four, the typical displacement only halves. This means that on small scales, Brownian motion is far more volatile and "jerky" than any smooth function.

This peculiar $\sqrt{\Delta t}$ scaling law has a stunning consequence. Let’s try to measure the "length" of a Brownian path from time $0$ to $T$. We'll do it the same way as before: slice the interval $[0,T]$ into $n$ tiny pieces, each of duration $\Delta t = T/n$, and sum the sizes of the steps.

First, let's try the total variation, which we'll call the ​​first-order variation​​. We sum the absolute values of the increments: $V^{(1)}_n = \sum_{i=1}^{n} |W_{t_i} - W_{t_{i-1}}| = \sum_{i=1}^{n} |\Delta W_i|$ We are adding up $n$ terms. The typical size of each term, $|\Delta W_i|$, is on the order of $\sqrt{\Delta t}$. So, our sum is roughly: $V^{(1)}_n \approx n \times (\text{typical size of } |\Delta W_i|) \approx n \times \sqrt{\Delta t} = \frac{T}{\Delta t} \times \sqrt{\Delta t} = \frac{T}{\sqrt{\Delta t}}$ As we take our ruler to be finer and finer (i.e., as $\Delta t \to 0$), this sum blows up to infinity! The path length is truly infinite, just as our intuition about the jagged coastline suggested.

Now for a stroke of genius. What if, instead of summing the steps, we sum the squares of the steps? Let's call this the ​​second-order​​ or ​​quadratic variation​​: $V^{(2)}_n = \sum_{i=1}^{n} (W_{t_i} - W_{t_{i-1}})^2 = \sum_{i=1}^{n} (\Delta W_i)^2$ Again, we are adding $n$ terms. But now, the typical size of each term, $(\Delta W_i)^2$, is on the order of $(\sqrt{\Delta t})^2 = \Delta t$. So, this sum is roughly: $V^{(2)}_n \approx n \times (\text{typical size of } (\Delta W_i)^2) \approx n \times \Delta t = n \times \frac{T}{n} = T$ Look at that! The $\Delta t$ terms cancel out. As we refine our partition, this sum does not blow up, nor does it vanish. In fact, it can be proven rigorously that as $n \to \infty$, this sum converges to the total time elapsed, $T$. This remarkable property is the fingerprint of Brownian motion. While its length is infinite (infinite first-order variation), its "accumulated squared jiggle" (its quadratic variation) is finite and meaningful. For a normal smooth path, the quadratic variation would be zero, because the sum would be on the order of $n \times (\Delta t)^2 = T \Delta t \to 0$.

This distinction is not just a mathematical curiosity; it is the foundation of a whole new branch of mathematics. The calculus we all learn in school, with its derivatives and integrals, is built on the assumption of local smoothness—assumptions that are violated by functions of unbounded variation.

• ​​Non-differentiability:​​ A function that is differentiable at a point must be "locally linear"; it can't be too jagged. This implies it must have bounded variation in a small neighborhood of that point. Since a Brownian path has unbounded variation on every interval, no matter how tiny, it cannot be differentiable at any point. It is a continuous curve with no well-defined tangent anywhere.

• ​​Breakdown of Classical Integration:​​ The classical rules of calculus, like the integration by parts formula, implicitly assume that the quadratic variation term is zero. The formula $\int f dg + \int g df = f(T)g(T) - f(0)g(0)$ holds beautifully for functions of bounded variation because the leftover term, which looks like $\sum (\Delta f)(\Delta g)$, vanishes in the limit. When we try to integrate with respect to a Brownian path, this term does not vanish.

This breakdown is not a disaster; it is an opportunity. It tells us that to work with the jagged reality of random processes, we need a new set of rules. This new set of rules is ​​Itô Calculus​​. It redefines integration and differentiation to account for the non-zero quadratic variation of processes like Brownian motion. The old formulas of calculus are reborn with new "correction terms" that precisely capture the contribution of this infinite, microscopic jiggling. The failure of old tools, like the Riemann-Stieltjes integral and the Dominated Convergence Theorem, in this context is the very reason stochastic calculus had to be invented. This is where our journey must head next, into the fascinating world of Itô's formula and stochastic differential equations.

In our previous discussion, we met a strange and wonderful creature from the mathematical zoo: the function of unbounded variation. We saw that such a function, even if perfectly continuous, can be so furiously wrinkled and jagged that its total "length" over any interval is infinite. At first glance, this might seem like a mere curiosity, a pathological case cooked up by mathematicians for their own amusement. But nothing could be further from the truth.

The discovery of functions with unbounded variation was not an endpoint; it was a doorway. It marked the boundary of the peaceful, smooth world of classical calculus and the beginning of a wilder, more realistic landscape. In this chapter, we will embark on a journey to see where these "jagged" functions live. We will find them not just in abstract equations, but at the heart of signal processing, physics, finance, and the very definition of randomness. We will see that grappling with the consequences of unbounded variation has been one of the great catalysts of modern mathematics, forcing us to forge powerful new tools to understand the world as it truly is: beautifully, irreducibly rough.

One of the crown jewels of 19th-century mathematics is the Fourier series. The idea is wonderfully simple and powerful: any reasonably well-behaved signal—be it the sound of a violin, the temperature fluctuations of a room, or the brightness of a distant star—can be broken down into a sum of simple, pure sine and cosine waves. This decomposition is the foundation of modern signal processing. But what, precisely, does "reasonably well-behaved" mean?

The conditions that guarantee a Fourier series will faithfully reconstruct the original function are known as the Dirichlet conditions. They are quite generous, allowing for functions with a finite number of jumps and wiggles. One of these crucial conditions, however, is that the function must be of ​​bounded variation​​. This means that its total up-and-down movement must be finite. For a long time, this seemed like a minor technicality. Surely any real-world signal we might care about would satisfy this?

It turns out, the world is more subtle. Consider a function like $f(x) = x^2 \sin(x^{-2})$ near the origin. This function is not only continuous, it's perfectly differentiable everywhere, even at the notoriously tricky point $x=0$. Yet, as you approach the origin, its oscillations become infinitely fast. The wiggles get smaller in amplitude (thanks to the $x^2$ factor, which pulls the function toward zero), but their frequency explodes. If you were to try and trace the path of this function, your pen would have to travel an infinite distance to get from any point $x \gt 0$ to the origin, even though the straight-line distance is finite. The function has unbounded variation. Another famous example is the function $\sin(1/x)$ itself, which oscillates with full amplitude an infinite number of times as it nears the origin.

The existence of such functions shows that the classical toolkit of Fourier analysis has its limits. It warns us that continuity and even differentiability are not enough to guarantee the kind of "tameness" that these tools were built for. This realization spurred mathematicians to develop more general theories. For instance, one can define "Fourier coefficients" for a broader class of objects using the Riemann-Stieltjes integral, which can handle functions of bounded variation that aren't necessarily smooth. But even here, the ghost of variation haunts the results. For a function with a sudden jump—a type of bounded variation function—the generalized Fourier coefficients may not fade away to zero at high frequencies as they do for smooth functions. This stubborn, non-decaying signal at high frequencies is, in a sense, the signature of the singularity, a permanent echo of the jump. Bounded variation, it turns out, is not just a technical condition; it's a deep classifying principle that governs the very behavior of a function's frequency content.

The real revolution came when we turned our attention from the deterministic world of signals to the chaotic dance of randomness. In 1905, Albert Einstein developed a mathematical theory for the incessant, jittery motion of a pollen grain suspended in water, bombarded by unseen water molecules. This phenomenon, known as Brownian motion, became the quintessential model for random walks everywhere, from the diffusion of heat to the fluctuations of stock prices.

When mathematicians began to study the sample path of a single Brownian particle, they discovered something astonishing. The path is continuous—the particle doesn't magically teleport from one point to another. But it is so erratic, so full of instantaneous zig-zags at every possible scale, that it is ​​nowhere differentiable​​. And, most importantly for our story, its path over any interval of time has ​​unbounded variation​​. It is the ultimate "jagged" function.

This is not a mathematical trick. It is the very essence of true randomness. And it has profound consequences. Suppose we try to tame a Brownian path. We can pick a series of points along its path and connect them with straight lines, creating a polygonal approximation. This new path, for any finite number of segments, is piecewise linear and therefore has bounded variation. We can apply all the rules of ordinary calculus to this approximation. We can calculate its rate of change (which is constant on each segment) and integrate functions against it using the classical Riemann-Stieltjes integral.

But what happens when we let our approximation become more and more faithful to the true path, by taking smaller and smaller time steps? The answer is that our classical calculus breaks down spectacularly. The limit of these well-behaved approximations is a path of unbounded variation, and for such a path, the very definitions of derivative and Riemann-Stieltjes integral collapse. You simply cannot do ordinary calculus with a Brownian motion.

This crisis was the birth of a new kind of mathematics: ​​stochastic calculus​​. Forged primarily by Kiyosi Itô, this new calculus was built from the ground up to handle functions of unbounded variation. The rules are different, and wonderfully strange. In ordinary calculus, if we take a small step $\Delta t$, the change in a smooth function is roughly proportional to $\Delta t$, and the square of the change is proportional to $(\Delta t)^2$, a much smaller number. For Brownian motion, a step in time $\Delta t$ produces a random change $\Delta W$. The key insight of Itô was that the square of this change, $(\Delta W)^2$, is not of order $(\Delta t)^2$. On average, it is of order $\Delta t$. This "anomalous" scaling, where the square of a small change is as significant as the change itself, is the mathematical signature of unbounded variation. It is called ​​quadratic variation​​, and its non-zero value is what distinguishes the calculus of random walks from the calculus of smooth paths.

The discovery that Brownian motion has unbounded variation opens up a new set of questions. Is all randomness this "rough"? Are there different kinds of random processes? The answer is a resounding yes. The world of stochastic processes is a veritable zoo, and the property of bounded variation is one of the key markers we use to classify its inhabitants.

A vast and important class of random processes are the Lévy processes, which are the building blocks for any process with stationary and independent increments. Think of them as the "elementary particles" of randomness. Brownian motion is one. Another is the ​​compound Poisson process​​, which models phenomena like the total claims arriving at an insurance company or the cumulative effect of meteorite impacts. This process stays constant for random periods of time and then suddenly jumps by a random amount. Unlike Brownian motion, its path is a series of steps. Over any finite time, it makes only a finite number of jumps. If we sum the absolute sizes of these jumps, we get a finite number. Thus, a compound Poisson process has ​​bounded variation​​.

Here we have a beautiful dichotomy: two fundamental random processes, one continuous but with infinite variation, the other discontinuous but with finite variation. This insight can be generalized dramatically. The great ​​Lévy-Itô decomposition theorem​​ tells us that any Lévy process can be broken down into three independent parts: a smooth, deterministic drift; a continuous, jittery Brownian motion; and a pure jump process. A Lévy process will have unbounded variation for one of two reasons: either it has a Brownian motion component, or its jump part consists of an "infinite swarm" of infinitesimally small jumps that are so numerous that their total absolute size is infinite.

This dichotomy has profound implications. For example, in financial modeling, a martingale represents a "fair game" where, on average, your future wealth is equal to your current wealth. A fundamental theorem of stochastic calculus states that any non-constant, continuous martingale must have paths of unbounded variation. This means that the price of a stock, if it is to be modeled as a continuous and unpredictable "fair game," cannot follow a smooth path. It must be an infinitely jagged function, like a Brownian motion. A smooth path would imply a degree of predictability that the market, in its ceaseless, random fluctuations, simply does not possess.

By now, it should be clear that the distinction between bounded and unbounded variation is fundamental. But we can be even more precise. "Unbounded variation" is not a single category; there is a whole spectrum of roughness. To measure this, mathematicians use the concept of ​​p-variation​​. Total variation is just 1-variation. The quadratic variation we discussed earlier is, in essence, 2-variation.

For the jump part of a Lévy process, its roughness is beautifully captured by a single number called the ​​Blumenthal-Getoor index​​, denoted by $\beta$. This index, which ranges from $0$ to $2$, measures the intensity of small jumps. A process has bounded variation (finite 1-variation) if its jump activity is low enough that $\beta \le 1$. If $\beta \ge 1$, the paths have infinite total length.

A canonical example is the family of symmetric $\alpha$-stable processes, which are pure jump processes used to model phenomena with occasional extreme events. For these processes, the Blumenthal-Getoor index is simply $\beta = \alpha$, where $\alpha$ is a parameter between $0$ and $2$. When $\alpha$ is close to $2$, the process resembles Brownian motion. When $\alpha$ is small, the process is dominated by large, infrequent jumps. The path of such a process has finite p-variation if and only if $p \gt \alpha$. This gives us a tunable dial for roughness: by changing $\alpha$, we can move smoothly through a whole universe of processes with different degrees of jaggedness.

Our journey has taken us far from the smooth, predictable world of elementary calculus. We have seen that unbounded variation is not a mathematical monster to be locked away, but a fundamental feature of the functions and processes that describe our universe. It is the signature of certain complex signals, the defining characteristic of random walks, the engine of financial markets, and the reason we needed to invent entirely new forms of calculus.

By embracing the complexity of the "jagged edge," we have gained a much deeper and more accurate understanding of nature. We learned that the world is not always smooth, and that in its infinite, intricate roughness, there lies a profound and challenging beauty. The story of unbounded variation is a testament to the power of mathematics to not only describe what we see, but to reveal the hidden structures that govern the seemingly chaotic world around us.