Sublevel Sets

How can we understand the complex, multi-dimensional landscapes defined by mathematical functions? A surprisingly simple yet powerful tool for this task is the concept of a ​​sublevel set​​—the collection of all points where a function's value lies at or below a given threshold. While abstract at first glance, this idea provides a unified language to analyze and solve problems across a vast range of scientific disciplines. This article addresses the gap between the formal definition of sublevel sets and their practical, far-reaching implications. It bridges this gap by providing an intuitive exploration of their properties and a tour of their most significant applications.

The article begins by establishing the foundational principles and mechanisms, exploring the crucial connections between sublevel sets, convexity, and topology. From there, it ventures into the real world, showcasing how this single concept is used to find optimal solutions, guarantee the stability of dynamic systems, and reveal the hidden shape of complex data.

Imagine you are a hiker exploring a vast, mountainous landscape. The elevation at any point $(x, y)$ on your map is given by a function, $f(x, y)$. Now, suppose you want to identify all the regions on the map that are at or below a certain altitude, say, 1000 meters. The set of all these points on your map constitutes the ​​sublevel set​​ for the level $\alpha = 1000$. This simple idea of "coloring in" all the parts of a domain where a function's value is below a certain threshold is one of the most powerful concepts in mathematics, with profound implications in fields from optimization to topology.

Let's make this idea more precise. For a function $f$ that takes a point $x$ from some space and returns a real number, the $\alpha$-sublevel set, which we can call $S_\alpha$, is simply the collection of all points $x$ in the domain where the function's value is less than or equal to $\alpha$. In mathematical notation:

There is a wonderfully intuitive way to visualize this. Think about the graph of the function. For a simple function like $f(x) = x^2$, the graph is a familiar parabola in the 2D plane. Now, imagine not just the curve of the graph, but the entire region above it. This region, called the ​​epigraph​​, is the set of all points $(x, y)$ such that $y \ge f(x)$.

How do we find the sublevel set $S_\alpha$ from this picture? We take a horizontal "slicer" and cut through the epigraph at a height of $y = \alpha$. The intersection of this plane with the epigraph gives us a set of points. If we then take this slice and project it straight down onto the x-axis, the shadow it casts is the sublevel set. For $f(x) = x^2$ and some positive level $\alpha$, slicing the epigraph at $y = \alpha$ gives us the line segment of points $(x, \alpha)$ where $x^2 \le \alpha$. Projecting this onto the x-axis gives us the interval $[-\sqrt{\alpha}, \sqrt{\alpha}]$, which is the sublevel set.

The nature of these sets can be surprisingly varied. For a constant function, like a perfectly flat plain at elevation $c$, the sublevel set $S_\alpha$ is either the entire map (if $\alpha \ge c$) or a completely empty map (if $\alpha \lt c$). For a wavy, periodic function like $f(x) = \sin(x)$, the character of the sublevel set changes dramatically depending on our chosen altitude $\alpha$. If $\alpha$ is 1 or greater, the condition $\sin(x) \le \alpha$ is always true, so the sublevel set is the entire real line $\mathbb{R}$. If $\alpha$ is less than -1, the condition is never true, and the sublevel set is the empty set $\emptyset$. For any level in between, say $\alpha = 0.5$, the sublevel set becomes an infinite collection of disjoint intervals—all the regions where the sine wave dips below the line $y=0.5$.

One of the most celebrated properties of sublevel sets relates to the concept of ​​convexity​​. A set is convex if, for any two points in the set, the straight line segment connecting them is also entirely contained within the set. A disk is convex; a donut shape is not. A function is convex if its epigraph is a convex set—geometrically, its graph "holds water."

Here lies a beautiful and fundamental connection: ​​every sublevel set of a convex function is a convex set​​. The logic is almost self-evident from the geometric picture. If a function's graph curves upwards everywhere, and you take two points $x_1$ and $x_2$ that are both below a certain altitude $\alpha$, it's impossible for the path between them, $f((1-t)x_1 + tx_2)$, to "bulge up" and cross above $\alpha$. The definition of convexity guarantees that the function's value along the line segment stays below the line segment connecting the heights at the endpoints, and since both heights are below $\alpha$, the whole path must be too.

This property is a one-way street, and the distinction is crucial. Does a function having all convex sublevel sets imply that the function itself must be convex? The answer is no. Consider the function $f(x) = \sqrt{\|x\|}$ for a point $x$ in $\mathbb{R}^n$. Its sublevel sets for $\alpha \ge 0$ are sets where $\|x\| \le \alpha^2$, which are just balls centered at the origin—perfectly convex sets. Yet, the function $f(x) = \sqrt{\|x\|}$ is not convex; its graph does not curve upwards sharply enough. Functions like this, which are not necessarily convex but have convex sublevel sets, are called ​​quasiconvex​​. This reveals a subtle hierarchy: all convex functions are quasiconvex, but not all quasiconvex functions are convex. The sublevel set is the key that unlocks this distinction.

Why is the convexity of sublevel sets so important? It's a cornerstone of modern optimization theory—the science of finding the best possible solution to a problem. Imagine you want to find the lowest point in a landscape described by a function $f(x)$.

If the landscape extends infinitely and generally slopes downwards in some direction, you might never find a minimum; you could just walk forever and keep going lower. But what if the landscape is ​​coercive​​—that is, no matter which direction you walk, if you go far enough away from the origin, the elevation $f(x)$ eventually goes up towards infinity?. Think of a bowl shape. In this case, a minimum must exist.

Sublevel sets provide the rigorous argument. If a function is continuous, its sublevel sets are ​​closed​​ sets (they include their boundaries). If the function is also coercive, its sublevel sets must also be ​​bounded​​. Why? Suppose a sublevel set $S_\alpha$ were unbounded. This would mean you could find points in it that are arbitrarily far from the origin. But by coercivity, the function's value at these far-out points must be tending to infinity. This creates a contradiction, because for every point in $S_\alpha$, the function's value must be less than or equal to $\alpha$. Therefore, the sublevel set must be confined to a finite region.

In the language of topology, a set in Euclidean space that is both closed and bounded is called ​​compact​​. The Extreme Value Theorem, a pillar of analysis, states that any continuous function on a compact set is guaranteed to attain its minimum and maximum values on that set.

This gives us a powerful strategy for optimization. To find the global minimum of a coercive function $f$, we can just pick any starting point $x_0$. The global minimum, wherever it is, must have a value less than or equal to $f(x_0)$. This means the minimum must lie within the sublevel set $S_{f(x_0)}$. Since $f$ is coercive and continuous, this sublevel set is compact. We have successfully trapped the solution within a bounded, closed region, transforming an infinite search into a finite one. In contrast, for a non-coercive function like $f(x) = \tanh(x)$, the sublevel set for $\alpha=1$ is the entire real line, which is not compact, and indeed the function approaches but never reaches the value of 1.

Sublevel sets do more than just tell us about convexity and minima; they can describe the entire topological "skeleton" of a function. Imagine flooding our mountainous landscape by slowly raising the water level, $\alpha$. The sublevel set $S_\alpha$ is the region of land that is underwater.

As we raise the water level, nothing topologically interesting happens—the flooded region just expands smoothly. But everything changes when the water level reaches a ​​critical point​​: a point where the gradient of the function is zero (a peak, a valley floor, or a saddle/pass). It is only at these critical altitudes that the shape, or topology, of the flooded region can change.

Let's consider a 2D landscape:

• When the water level $\alpha$ reaches a local minimum (the bottom of a valley), a new body of water appears out of nowhere. A new ​​connected component​​ is born in the sublevel set.
• When $\alpha$ reaches a ​​saddle point​​ (a mountain pass), a dramatic event occurs. Two separate bodies of water, previously in two different valleys, might suddenly merge into one. As the water crests the pass, what were two distinct components of the sublevel set become a single, larger component. The number of connected components decreases by one. This is like attaching a "handle" or a bridge between two islands.
• When $\alpha$ reaches a local maximum (a peak), the last bit of dry land on an island is submerged. A "hole" in the sublevel set is filled in.

This perspective, a cornerstone of ​​Morse theory​​, transforms our static view of a function into a dynamic story. By observing how the topology of the sublevel sets changes as we sweep the level $\alpha$ from $-\infty$ to $+\infty$, we can reconstruct the entire structure of the function's landscape, piece by piece, critical point by critical point. From a simple tool for identifying "low ground," the sublevel set becomes a sophisticated probe into the very fabric of a function's shape.

Now that we have acquainted ourselves with the definition of a sublevel set, you might be asking a perfectly reasonable question: “So what?” Is this just another clever piece of mathematical abstraction, a sterile concept destined for the dusty shelves of an ivory tower?

Far from it! It turns out this simple idea of “everything below a certain line” is a remarkably powerful lens. It is one of those wonderfully simple, yet profound, concepts that illuminates deep connections between seemingly disparate worlds. With this single tool, we can journey from the abstract certainty of a mathematical proof to the messy reality of biological molecules, from the quest for the best possible design to the challenge of keeping a spaceship on its course. It is a testament to the inherent unity of scientific thought that such a simple construction can act as a skeleton key, unlocking secrets across the disciplines. Let us begin our tour of some of these fascinating landscapes.

At the heart of optimization lies a very basic question: can we find the best solution? This could mean the cheapest way to build a bridge, the most accurate model to predict the weather, or the most efficient route for a delivery truck. Often, these problems involve searching through an infinite space of possibilities. How, then, can we be sure that a “best” solution even exists? Perhaps we can get better and better, forever approaching some ideal without ever reaching it.

This is where sublevel sets provide a foothold. Consider the common problem of fitting a line to data, which often boils down to minimizing a function like $f(x) = \|Ax-b\|_2^2$. The vector $x$ represents the parameters of our model, and $f(x)$ is the error. We want to find the $x$ that makes the error as small as possible. The challenge is that $x$ can be any vector in $\mathbb{R}^n$. If our function $f(x)$ just gets flatter and flatter as we go out to infinity in some direction, its sublevel sets $\{x : f(x) \le \alpha\}$ would be unbounded, and our search for a minimum might send us on a wild goose chase to infinity.

But we can perform a clever change of scenery. Instead of looking at the space of parameters $x$, let's look at the space of outcomes, $y = Ax$. The problem is now to find the point $y$ in the subspace spanned by the columns of $A$ (the range of $A$) that is closest to our target data $b$. For any candidate error level $\alpha$, the sublevel set in this outcome space is $\{y \in \operatorname{range}(A) : \|y-b\|_2^2 \le \alpha\}$. This is the intersection of a closed ball (a compact set) and a subspace (a closed set). The resulting set is therefore compact! The Weierstrass theorem tells us that a continuous function on a compact set must attain its minimum. So, a best outcome $y^*$ must exist. And if a best outcome exists, there must be at least one parameter vector $x^*$ that produces it ($Ax^*=y^*$). Voilà! The existence of a minimum is guaranteed, not by coercivity, but by looking at the compact sublevel sets in the right space.

Alright, so a minimum exists. But is it unique? Is there only one “best” answer? One might guess that if the function is convex (shaped like a bowl), the minimum should be unique. Convex functions have convex sublevel sets. Is that enough? Let’s consider a function shaped like a flat-bottomed bowl, for instance $f(x) = (\max\{0, \|x\|-1\})^2$. Its sublevel sets for any $\alpha \ge 0$ are just closed balls, which are perfectly nice strictly convex sets. Yet, the minimum value of this function is $0$, and it is achieved for any point in the entire unit ball $\|x\| \le 1$. We have an infinity of minimizers! The sublevel sets reveal the nuance: their shape tells us about uniqueness. What we truly need is a condition that forbids these "flat spots" at the minimum value. This condition is called strict quasiconvexity, which ensures that on any line segment between two points, the function value in the middle is strictly lower than the higher of the two endpoints. This subtle refinement, naturally expressed in terms of function values on line segments within sublevel sets, is precisely what guarantees a unique global minimum.

This geometric insight is not just for proofs; it builds powerful algorithms. Some problems are not convex, but their sublevel sets are. These are called quasiconvex problems. We can solve them using a bisection method that feels like a detective narrowing down a search. The problem of minimizing $f(x)$ is equivalent to finding the smallest value $\tau$ for which the sublevel set $S_\tau = \{x : f(x) \le \tau\}$ is not empty. We can search for this optimal $\tau^*$. We start with a range $[\tau_L, \tau_U]$ where we know $\tau^*$ must lie. We pick a test value $\tau$ in the middle. We then ask a simple, geometric question: "Is the set $S_\tau$ empty?" Because the sublevel set is convex, this is a convex feasibility problem, which is very efficient to solve. If it's not empty, it means we can achieve a value of $\tau$ or better, so we update our upper bound: $\tau_U = \tau$. If it is empty, we aimed too low, so we update our lower bound: $\tau_L = \tau$. We repeat this, cutting our interval of uncertainty in half at each step, until we have zeroed in on the true minimum value. A complex optimization is thus reduced to a sequence of simple geometric "yes/no" questions.

Sublevel sets don't just help us find static optima; they are also indispensable for describing the dynamic behavior of systems over time. A central question in control theory is stability: if we nudge a system—a pendulum, a satellite, a chemical reactor—will it return to its stable equilibrium, or will it fly off course?

The brilliant Russian mathematician Aleksandr Lyapunov had an idea of genius, one that avoids the nearly impossible task of solving the system's equations of motion directly. He imagined the state of the system as a point on a landscape, where the height is given by an "energy-like" function $V(x)$. The stable equilibrium we care about, say the origin $x=0$, sits at the bottom of a valley. The sublevel sets, $\Omega_c = \{ x : V(x) \le c \}$, are then all the points in the landscape below a certain altitude $c$.

Now, let's watch how the system moves on this landscape. We can calculate the rate of change of energy along a trajectory, $\dot{V}(x)$. If we can show that the system's dynamics are always pointing "downhill" or, at the very least, not uphill ($\dot{V}(x) \le 0$) everywhere on the boundary of a sublevel set $\Omega_c$, then that boundary acts as an invisible fence. A trajectory that starts inside $\Omega_c$ can never climb past the altitude $c$, so it can never get out. The sublevel set becomes a trapping region, a provably safe zone from which the system can never escape. If the function $V(x)$ is radially unbounded (it grows to infinity in all directions), then these sublevel sets are also bounded and thus compact.

This method gives us a powerful recipe for certifying safety and stability. For a given system, we can propose a Lyapunov function, like the simple quadratic energy $V(x) = x_1^2 + x_2^2$, whose sublevel sets are circles (or spheres). We then calculate the region of state space where the dynamics are guaranteed to be dissipative, i.e., where $\dot{V}(x) 0$. The game is then to find the largest possible circle that fits entirely inside this dissipative region. Any trajectory starting inside this circle is guaranteed to be trapped and, if set up correctly, will spiral down into the stable equilibrium. The radius of this largest circle is determined by the "first-exit" points—the points on the boundary of the dissipative region (where $\dot{V}(x)=0$) that are closest to the origin.

The method is even more powerful. What if $\dot{V}(x)$ isn't strictly negative, but can be zero on some paths away from the origin? LaSalle's Invariance Principle extends Lyapunov's idea. A trajectory might drift along a path of constant "energy," but as long as it cannot stay on such a path forever (unless that path is just the equilibrium point itself), it must eventually descend. The strategy remains to find the largest sublevel set $\Omega_c$ that avoids any "uphill" regions. Even if this set contains paths where $\dot{V}(x)=0$, we can check if the system dynamics would force it to leave those paths. If so, LaSalle's principle still guarantees that everything starting in $\Omega_c$ will eventually reach the equilibrium. This provides a rigorous way to estimate the region of attraction—the set of all initial states that will lead to stability.

So far, we have used the sublevel sets of a function to understand its optima or the dynamics it governs. But we can flip this perspective entirely. What if we use the changing topology of the sublevel sets to reveal the shape of the underlying space itself?

This is the central idea of Morse Theory. Imagine a rugged island, and let our function $f$ be the height above sea level. The sublevel set $M_c = \{p : f(p) \le c\}$ is simply the part of the island that is underwater when the sea is at level $c$. As we slowly raise the water level, the topology of this underwater region changes, but only in very specific, predictable ways. Nothing much happens, until the water level hits a critical point of the height function.

• At a local minimum (the bottom of a basin), a new body of water—a new connected component—appears out of nowhere. This is a topological "birth".
• At a saddle point (a mountain pass), two previously separate bodies of water may merge into one. This is a "death" of one component.
• At a local maximum (a peak), the last bit of dry land is submerged, potentially filling in a hole (like a lagoon) in the water.

Morse theory gives us a precise formula for these events. The change in a topological counter called the Euler characteristic, $\chi$, as the level $c$ crosses a critical value is given by $\Delta\chi = (-1)^\lambda$, where $\lambda$ is the Morse index (the number of "downhill" directions) at the critical point. For a saddle point in 2D, $\lambda=1$, so $\Delta\chi = -1$. This makes perfect intuitive sense: two components merge into one, so the number of connected components decreases by one. We can see this in action by considering a function like $f(x,y) = \cos(2x) + \cos(y)$ on a torus. This function has two global minima. If we choose a level just above the minimum value but below the value at the saddles, the sublevel set consists of two small, disconnected "puddles" centered at these minima.

This powerful idea—of tracking the evolution of sublevel sets—is the engine behind a modern and exciting field called Topological Data Analysis (TDA). The core technique, persistent homology, turns a function on a space into a topological signature. We build a filtration by sweeping a level across the function's range and recording the birth and death of topological features.

Let's take a trefoil knot in 3D space and use the simple height function $z$. As we sweep a plane from bottom to top, we observe which parts of the knot are below the plane. At each local minimum of the height function, a new piece of the knot appears, giving birth to a new connected component. As the plane rises, these pieces grow until they meet at a local maximum, where they merge, causing the "death" of one component. The crucial insight of TDA is to measure the persistence of each feature: its lifespan, death time - birth time. A small, insignificant wiggle in the knot will create a component that is born and dies almost immediately—it has low persistence. A major fold of the knot, however, will create a component that survives for a long range of height values—it has high persistence.

This ability to separate significant features from noise is revolutionary for real-world data. Consider the problem of finding binding sites on a protein, which are often pockets or clefts on its surface. A protein is a fantastically complex jumble of atoms. We can define a function on its surface, such as the electrostatic potential. By running a sublevel set filtration on this potential function, we can track the birth and death of components. A shallow dimple on the surface will be a low-persistence feature. But a deep pocket, a candidate for drug binding, will appear as a highly persistent feature—a component that is born at a low potential value and "survives" for a long time before merging with the rest of the surface. TDA provides an automated, rigorous way to see the true "shape" of the data, filtering out the noise to reveal the essential structure.

From the abstract foundations of optimization to the tangible challenges of engineering and biology, the sublevel set provides a simple, yet profound, unifying language. By simply asking "what lies below this level?", we gain a powerful new perspective, revealing the hidden geometry that governs the world around us.