Viscosity Solution

In the world of optimization and dynamic systems, from navigating a spacecraft to pricing a financial asset, the central goal is to make the best possible decision at every moment. For centuries, calculus, with its concept of the derivative, has been our primary guide, allowing us to find optimal paths by following the steepest descent. This logic is enshrined in powerful mathematical tools like the Hamilton-Jacobi-Bellman (HJB) equation, the master equation of optimal control. However, this classical approach faces a critical breakdown when the very landscape of "value" we are trying to navigate is not smooth but contains sharp "kinks" or corners, points where derivatives cease to exist. This common occurrence in real-world problems renders our classical tools useless and the HJB equation meaningless.

This article explores the revolutionary workaround to this fundamental problem: the theory of ​​viscosity solutions​​. Introduced by Michael Crandall and Pierre-Louis Lions, this framework provides a new and robust way to interpret and solve nonlinear PDEs even when their solutions are not differentiable. It replaces the impossible task of direct differentiation with an elegant "touching" principle, fundamentally changing how we understand these equations. In the chapters that follow, you will discover the core ideas behind this powerful theory. The first chapter, "Principles and Mechanisms," will unpack the intuitive definition of viscosity solutions and explain the three pillars—consistency, uniqueness, and existence—that make the theory so effective. The subsequent chapter, "Applications and Interdisciplinary Connections," will showcase the astonishing breadth of this idea, demonstrating how it provides a unifying language for problems in optimal control, differential games, geometric flows, and mathematical finance.

Imagine you're an intrepid explorer in a vast, hilly landscape, and your mission is to find the quickest path from your current location to a faraway destination. Better yet, let's say every point in this landscape has an intrinsic "cost" or "value"—perhaps representing the total effort required to reach the final goal from that point. Your job is to make the locally optimal decision at every step to minimize your total cost. How do you do it?

For centuries, the primary tool for such problems has been calculus. In this landscape of value, calculus provides a marvelous compass: the ​​derivative​​. The gradient, or derivative, at any point tells you the direction of steepest ascent. To minimize your cost, you simply need to head in the opposite direction. This fundamental idea is the heart of countless optimization methods, and when applied to dynamic problems in time, it gives birth to powerful descriptions like the ​​Hamilton-Jacobi-Bellman (HJB) equation​​. The HJB equation is supposed to be the master equation of optimal control, a differential equation whose solution, the ​​value function​​ $V(t,x)$, contains all the information needed to navigate any situation optimally.

But here lies a profound and beautiful difficulty. What if your landscape isn't smooth? What if it has sharp ridges, V-shaped valleys, or cliffs? Stand on the peak of a pyramid—which way is "down"? There are infinitely many downhill directions. At such "kinks," the very idea of a single, well-defined slope breaks down. The derivative does not exist.

As it turns out, the value functions of many real-world control problems are precisely like this. The very act of making an optimal choice—of picking the single best path among many alternatives—can create these kinks and non-differentiable points in the value landscape. For example, if two different strategies yield the exact same minimal cost from a certain point, that point might lie on a "seam" where the value function is not smooth.

This presents a catastrophic problem for the classical HJB equation. The equation is written in terms of derivatives of the value function, like $\partial_t V$, $\nabla V$, and the Hessian matrix $D^2 V$. If these derivatives don't exist at a point, the equation becomes meaningless, like asking for the color of the number nine. Our calculus compass is broken. We can't apply the tools of classical analysis, such as Itô's formula, directly to the value function to derive the HJB equation, because the required smoothness is simply not there. We are, in a sense, lost in a landscape we can't navigate. We need a new kind of compass.

When a problem seems unsolvable, a common tactic in mathematics is to change the question. If we cannot measure the slope of our kinky landscape directly, perhaps we can understand it by comparing it to things we do understand. This is the breathtakingly clever idea behind ​​viscosity solutions​​, a concept developed in the early 1980s by Michael Crandall and Pierre-Louis Lions that revolutionized the study of nonlinear partial differential equations.

The idea is this: instead of trying to differentiate our potentially non-smooth value function $V$, we "test" it using infinitely smooth functions, which we'll call $\varphi$. Imagine bringing a perfectly smooth, curved sheet of glass ($\varphi$) and touching it to our landscape ($V$) at a single point, say $(t_0, x_0)$.

There are two ways we can do this:

1. ​​Touching from Above (The Subsolution Condition):​​ We can place our smooth sheet of glass on top of the landscape, so it rests on it, touching at $(t_0, x_0)$. At this point, the landscape $V$ must be "flatter" than, or at most as curved as, our test sheet $\varphi$. It cannot be more 'peaked' than the sheet it is supporting. This geometric intuition is translated into a precise mathematical inequality. Since we know the derivatives of our smooth sheet $\varphi$ everywhere, we plug its derivatives into the HJB equation. The ​​viscosity subsolution​​ condition demands that at this touching point, the equation holds in a one-sided way:

   $$-\partial_t \varphi(t_0,x_0) + H(x_0, D\varphi(t_0,x_0), D^2\varphi(t_0,x_0)) \le 0.$$

   Here, $H$ is the Hamiltonian, the part of the HJB equation that involves the control choices and derivatives. This inequality must hold for any smooth function $\varphi$ that touches $V$ from above at any point.

2. ​​Touching from Below (The Supersolution Condition):​​ We can also bring our smooth sheet of glass underneath the landscape, pushing it up until it just kisses the landscape at $(t_0, x_0)$. Now, the landscape $V$ must be "steeper" than, or at least as curved as, our test sheet $\varphi$. This gives us the opposite inequality. The ​​viscosity supersolution​​ condition demands that for any such test function, we have:

   $$-\partial_t \varphi(t_0,x_0) + H(x_0, D\varphi(t_0,x_0), D^2\varphi(t_0,x_0)) \ge 0.$$

A function $V$ is then defined as a ​​viscosity solution​​ if it is simultaneously a subsolution and a supersolution. It is a function that passes every single one of these "touching tests," from above and below, at every single point in its domain. This definition beautifully sidesteps the problem of non-differentiability. It never attempts to compute a derivative of $V$. Instead, it uses a whole family of smooth proxies to enforce the physical and geometric constraints of the original problem everywhere. It's a generalization of the classical notion of a solution, and it turns out that if a classical (smooth) solution exists, it is also a viscosity solution.

This "touching" principle might at first seem like a purely formal trick. But it is far more. The viscosity framework stands on three pillars that establish it as the natural, correct, and powerful way to understand these equations.

Pillar 1: Consistency with Reality

The definition is not arbitrary. One can rigorously prove that the value function of the optimal control problem—the "true" answer we seek—is indeed a viscosity solution to the HJB equation. The subsolution and supersolution inequalities can be derived directly from the ​​Dynamic Programming Principle​​, the fundamental statement that an optimal path is made of smaller optimal segments. This ensures that the mathematical object we have defined is the same one the underlying physical or economic problem describes.

Pillar 2: Uniqueness via Comparison

A theory that provides multiple answers to the same well-posed question isn't very useful. The viscosity framework provides a definitive answer through a powerful result known as the ​​comparison principle​​. The principle states that if you have a viscosity subsolution $u$ and a viscosity supersolution $v$, and if $u$ starts out below $v$ on the boundary of the problem (e.g., at the terminal time), then $u$ must remain below $v$ everywhere.

This has a monumental consequence. If you have two different viscosity solutions, $V_1$ and $V_2$, for the same problem with the same boundary data, you can apply the principle twice. First, treat $V_1$ as the subsolution and $V_2$ as the supersolution; you get $V_1 \le V_2$. Then, swap their roles; you get $V_2 \le V_1$. The only way both can be true is if $V_1 = V_2$. The solution is unique!. This property relies on a structural feature of the equation called ​​degenerate ellipticity​​—essentially, a monotonicity condition that ensures the "touching" inequalities point the right way.

Pillar 3: Existence through Stability

The final pillar is the guarantee that a solution actually exists. Classical theories often struggle here, but viscosity solutions possess a remarkable property: ​​stability under uniform limits​​. If you have a sequence of viscosity solutions to a sequence of problems, and these solutions converge to some limit function (even just in the 'looks like' sense of uniform convergence, without any information about derivatives), then the limit function is itself a viscosity solution to the limit problem!.

This stability is a game-changer. It allows us to prove that a solution exists for a very difficult, degenerate problem by first solving a sequence of "nicer," non-degenerate approximations (for example, by adding a tiny bit of extra randomness, a method called "vanishing viscosity"). We can then show that the solutions to these easier problems converge, and the stability property guarantees that their limit is the solution we were looking for. It also provides a rigorous foundation for numerical schemes that approximate a complex domain with a series of simpler ones.

The theory of viscosity solutions did more than just fix a "bug" in the classical theory for HJB equations. It provided a completely new and unifying lens through which to view a vast class of nonlinear partial differential equations.

It isn't confined to problems where randomness is everywhere. It seamlessly handles deterministic control problems (where the diffusion matrix $\sigma$ is zero) and degenerate problems where noise only affects the system in certain directions. It also provides a framework for handling difficult boundary conditions. For instance, if a problem has a discontinuous terminal reward (e.g., you get a prize only if you land on an exact spot), the viscosity framework handles this with grace by relaxing the terminal condition to involve the upper and lower limits of the discontinuous function, once again ensuring a unique and stable solution exists.

Prior to this theory, mathematicians often used a different concept of "weak solutions" (or distributional solutions), which works beautifully for linear equations by using integration by parts. However, for the fully nonlinear, non-divergence form equations that arise in control theory and geometry, that approach fails because you cannot sensibly define what it means to multiply nonlinear functions of derivatives that are not even functions themselves. Viscosity solutions provided the right way forward.

In the end, the journey from a broken compass to a robust, unifying theory is a testament to the power of asking the right questions. By stepping back from the impossible demand of direct differentiation and instead asking what we can learn by comparison to smooth objects, mathematicians uncovered a deep and beautiful structure that was there all along, waiting to be seen.

What does the flight of an optimal rocket have in common with an evolving soap bubble? How can the mathematics describing a financial market under attack also explain the most likely way for a molecule to cross an energy barrier? It may seem that these problems live in completely different universes. Yet, as we so often find in science, a powerful idea can slice through the disciplinary boundaries and reveal a stunning, hidden unity.

In the previous chapter, we delved into the machinery of viscosity solutions. It might have felt like a rather abstract exercise in taming the wild beasts of non-differentiable functions. But this framework is no mere mathematical curio. It is a master key, one that unlocks a vast and diverse range of problems in science, engineering, and even economics. Its secret lies in its remarkable ability to identify the one "correct," physically stable solution in situations where classical approaches fail, typically when things develop "kinks," "corners," or other messy features. Let’s embark on a journey to see this key in action.

Perhaps the most natural home for viscosity solutions is in the world of optimal control theory—the science of finding the best way to do something. Imagine you are trying to fly a probe to Mars using the least amount of fuel, or you are an economist trying to set a tax policy to maximize social welfare over the next century. In each case, there is a "value" associated with your current situation: the minimum fuel you'll need from here on out, or the maximum possible future welfare. This "value" function, it turns out, must obey a specific law—a partial differential equation known as the ​​Hamilton-Jacobi-Bellman (HJB) equation​​.

The trouble is, this value function is often not smooth. Think of the shortest path to an exit in a room with a large pillar in the middle. Your optimal path will go straight, then bend sharply around the pillar. The "cost-to-go" function associated with this path develops a kink—a point of non-differentiability—right at the edge of the pillar's shadow. The classical theory of differential equations throws its hands up at such points. This is where viscosity solutions make their grand entrance. They provide a robust way to make sense of the HJB equation even when its solution has corners and kinks.

This handles a whole class of problems with breathtaking generality. For instance, we can solve "exit-time" problems, where we want to control a system optimally until it leaves a predefined "safe" domain. This is akin to a game that ends when your player goes out of bounds, and it's the mathematical basis for pricing certain financial instruments that expire if a stock price hits a certain barrier. We can also tackle "infinite-horizon" problems, where we manage a system forever, like sustaining a fishery or regulating a power grid. The viscosity framework elegantly handles the complexities of asserting conditions "at infinity."

The theory is so flexible that it can even manage situations where the rules of the game are about staying inside the lines. In "viability theory," we ask: what actions can I take to guarantee my system (a robot, an airplane, a biological population) remains within a safe operating region? The viscosity solution to the corresponding HJB equation cleverly encodes this constraint, not as a simple boundary condition, but by requiring the equation's inequalities to hold on the entire closed domain, boundary and all.

The story gets even more exciting when you're not just playing against randomness, but against a competitor or an adversarial "nature." This is the realm of differential games and robust control. What is the best investment strategy if you know the market might move against you in the worst possible way? The governing equation is no longer a simple HJB, but a ​​Hamilton-Jacobi-Bellman-Isaacs (HJBI) equation​​, which involves a complex minimax inf-sup structure. Once again, the viscosity solution framework takes this in stride, providing a solid foundation for finding the value of the game, provided certain structural conditions—like the so-called Isaacs condition—are met.

Let's now pivot from the abstract world of value functions to something we can literally see: the evolution of shapes. Imagine a soap bubble. Surface tension pulls it into a sphere to minimize its surface area. This process, where a surface moves in the direction of its mean curvature, is called ​​Mean Curvature Flow​​. It's nature's way of smoothing things out. This idea is not just for bubbles; it's a critical tool in materials science for modeling grain growth, in computer graphics for smoothing 3D models, and in medical imaging for segmenting organs.

But there's a problem. As a shape evolves, its topology can change. A dumbbell shape might pinch off in the middle to become two separate spheres. At the moment of pinching, the surface is no longer smooth, and a classical description of the evolving boundary breaks down.

The brilliant solution is the "level-set method." Instead of tracking the moving boundary itself, we imagine it as the coastline of a mysterious, higher-dimensional landscape defined by a function $u(x,t)$. The coastline is simply the set of points where the altitude is zero—the zero level set. As the landscape $u$ evolves according to a specific PDE, the coastline moves with it. The beauty of this is that the landscape can remain perfectly well-behaved even as its coastline pinches off, merges, or develops sharp corners.

The PDE that governs the landscape's evolution is a tricky, "degenerate parabolic" equation. And the right way to understand its solutions is—you guessed it—through the theory of viscosity solutions. One of the most elegant consequences of this is the ​​avoidance principle​​. The comparison principle for viscosity solutions, which states that if one solution starts below another it must stay below, has a startling geometric meaning. If you start with two separate, non-intersecting bubbles and let them both evolve by mean curvature flow, their level-set functions will be ordered, and as a result, the bubbles will never collide. This is by no means obvious from just looking at the flow, but it falls out directly from the robust structure of viscosity solutions.

The true power of a great idea is measured by the unexpected connections it reveals. Viscosity solutions form a nexus, a bridge between optimal control, geometry, and other deep theories in physics and probability.

The Ghost of Viscosity and the Path of Light

The very name "viscosity solution" comes from a physical idea. Often in physics, a difficult, idealized problem (like the flow of a frictionless fluid) can be understood by first solving a more realistic problem with a bit of friction or "viscosity," and then seeing what happens as this viscosity term is driven to zero. The same trick works for PDEs. A notoriously hard first-order equation can be "regularized" by adding a tiny second-order term, $-\epsilon \Delta u$, which acts like mathematical viscosity. As you let $\epsilon \to 0$, the sequence of "nice" solutions $u_\epsilon$ converges to a unique, and often not-so-nice, limit function $u$. This very limit is the viscosity solution.

One of the most famous equations that arises this way is the ​​Eikonal equation​​, $|\nabla u|^2 = 1$. This equation is ancient; it describes the propagation of wavefronts in optics. When light passes through a lens, the wavefronts can focus, cross, and form singularities called caustics—the bright, sharp lines of light you see at the bottom of a swimming pool. The viscosity solution of the Eikonal equation correctly describes the wavefront even after it has passed through these singularities, capturing the physically correct, multi-valued solution in a single function.

The Unlikely Path and the Calculus of Randomness

Consider a system governed by random fluctuations, like a particle jiggling in a warm fluid. Most of the time it stays put, but there's a tiny chance it could make a large, improbable journey across the container. What is the "most likely" way for such an unlikely event to happen? The theory of ​​Large Deviations​​, pioneered by Varadhan with roots in the work of Freidlin and Wentzell, gives a stunning answer. The most probable path is the one that minimizes a certain "action" or "cost." This means the question of the most likely improbable path is secretly an optimal control problem!

The value function for this control problem, which tells you the minimum cost (or log-probability) to get from one point to another, satisfies an HJB equation. And its solution is, of course, a viscosity solution. This profoundly connects the statistical mechanics of random systems to the deterministic world of optimal control, all refereed by the theory of viscosity solutions.

Looking Backwards in Time: Finance and Path-Dependence

Finally, let us consider a truly strange idea: the ​​Backward Stochastic Differential Equation (BSDE)​​. A normal differential equation starts with a known initial condition and evolves forward into an unknown future. A BSDE is defined by a known terminal condition, and it evolves backward in time to find the unknown state today.

This might sound like science fiction, but it is the natural language of modern mathematical finance. The price of a financial option contract is a perfect example. You know its value at the expiration date—it's determined by the stock price at that moment. The fundamental problem is to determine its fair price today. This is a backward-in-time problem. The celebrated ​​nonlinear Feynman-Kac formula​​ establishes a deep duality: the solution to a BSDE can be represented as the value of a viscosity solution to a related (semilinear) PDE. This bridge is one of the most powerful tools in quantitative finance for pricing and hedging complex derivatives.

And the theory does not stop there. What if the final payoff depends not just on the final stock price, but on its entire history—say, its average price over the last month? This requires a radical generalization to path-dependent PDEs, where the value function depends on the entire past trajectory. This is the frontier of research, yet the core ideas of viscosity solutions are once again being extended to bring mathematical rigor to this incredibly complex world.

Our journey is complete. We have seen the same fundamental idea—the viscosity solution—appear in the design of optimal rocket trajectories, the evolution of geometric shapes, the propagation of light, the statistical physics of rare events, and the pricing of financial securities. In each case, it plays the same crucial role: it selects the unique, stable, and physically meaningful solution precisely when the world becomes messy and classical tools fail. It is a beautiful testament to how a single, powerful piece of mathematics can impose order and reveal unity across a vast landscape of scientific inquiry.