Viscosity Solutions

In fields ranging from optimal control to financial engineering, the quest for the 'best' strategy often leads to mathematical models whose solutions are not perfectly smooth. Like a GPS route that involves sharp turns, these "value functions" possess 'kinks' or 'corners' where classical calculus breaks down, creating a crisis for traditional methods based on differentiation. How can we make sense of a differential equation when its solution isn't differentiable? The theory of viscosity solutions, a revolutionary concept developed by Michael Crandall and Pierre-Louis Lions, provides a profound and elegant answer to this very question by redefining what it means to be a "solution."

This article delves into the elegant theory of viscosity solutions. In the first chapter, "Principles and Mechanisms," we will explore the intuitive idea behind the definition and its key properties of existence, uniqueness, and stability that make it so robust. Following this, the "Applications and Interdisciplinary Connections" chapter will showcase the theory's remarkable impact, revealing its role as a unifying language across fields like optimal control, geometric analysis, fluid dynamics, and mathematical finance.

Imagine you are using a GPS to find the fastest route through a city. The path it suggests isn't always a smooth, gentle curve. It often involves sharp turns, sudden stops, and abrupt changes in direction. If you were to plot a function representing the "optimal time to destination" from any point on the map, you would find that this function has "kinks" or "corners" precisely at the intersections where your optimal path makes a sharp turn. At these kinks, the function is not differentiable. You cannot use the familiar tools of calculus to describe its local behavior.

This seemingly simple problem of a non-smooth "value function" is a microcosm of a deep crisis that arises in many fields of science and engineering, from optimal control and economics to fluid dynamics and geometric flows. The fundamental laws governing these systems are often expressed as partial differential equations (PDEs), and for centuries, we sought "classical" solutions—that is, functions that are smooth enough to be plugged directly into the equation. But reality, like our GPS route, is often not smooth. The value of an option in finance, the shape of a melting ice crystal, and the dynamics of a system that can switch between different modes all lead to solutions that are continuous but not everywhere differentiable. When a function isn't differentiable, how can we even say what it means to be a "solution" to a differential equation? This is where the beautiful and powerful idea of ​​viscosity solutions​​ enters the stage.

The classical approach to solving a PDE, say the Hamilton-Jacobi-Bellman (HJB) equation that governs optimal control problems, relies on a formal derivation using tools like the ​​Dynamic Programming Principle​​—the common-sense idea that any sub-path of an optimal path must itself be optimal. This derivation involves applying calculus, such as Itô's formula in a stochastic setting, directly to the value function $V$. But this step brazenly assumes that $V$ is smooth enough to have one time derivative and two spatial derivatives ($V \in C^{1,2}$). When faced with the reality that $V$ often has kinks, this derivation breaks down completely. We are left with an equation we cannot even evaluate.

The breakthrough, pioneered by Michael Crandall and Pierre-Louis Lions in the early 1980s, was to change the very philosophy of what it means to be a solution. The idea is wonderfully intuitive: if you cannot directly measure the properties of your wrinkly, non-smooth function, you can learn about it by seeing how it interacts with a universe of perfectly smooth functions.

Imagine our non-smooth value function $V$ as a rugged mountain range. We want to check if it satisfies our PDE, but we cannot measure the "slope" and "curvature" at every point because of the sharp peaks and crags. Instead, we use smooth "hills" and "bowls"—differentiable ​​test functions​​—to probe the landscape.

The viscosity definition consists of two conditions:

1. ​​The Supersolution Condition (No Peaks are Too Sharp):​​ Consider any point $(x_0, t_0)$ on our mountain range $V$. If we can find a smooth hill $\varphi$ that just touches $V$ from below at this point (meaning $V-\varphi$ has a local minimum), then the PDE inequality for a "supersolution" must hold for the smooth function $\varphi$. In essence, we are saying that at any point, the mountain range cannot be "sharper" or curve down more steeply than what the law of the system allows.

2. ​​The Subsolution Condition (No Valleys are Too Flat):​​ Symmetrically, if we can find a smooth bowl $\varphi$ that just touches $V$ from above at some point $(x_0, t_0)$ (meaning $V-\varphi$ has a local maximum), then the PDE inequality for a "subsolution" must hold for $\varphi$. This means the mountain range can never be "flatter" or curve up less steeply in any valley than what the governing equation dictates.

A function is a ​​viscosity solution​​ if it is simultaneously a subsolution and a supersolution everywhere. It's a brilliant workaround. We never differentiate the non-smooth function $V$. We only ever differentiate the smooth test functions $\varphi$ that touch it, and use them as proxies to ensure that $V$ obeys the PDE in a "weak" but profoundly meaningful way. This concept works even when the underlying process has randomness that can vanish in certain directions, leading to what are known as ​​degenerate parabolic​​ equations—a nightmare for classical methods but familiar territory for viscosity solutions.

This clever definition would be a mere curiosity if not for three powerful properties that make it the bedrock of modern nonlinear PDE theory. These properties are often called the "three pillars": existence, uniqueness, and stability.

Pillar 1: Existence

The first question is, do such solutions even exist? The answer is a resounding yes. For a vast class of problems in optimal control, differential games, and geometric analysis, one can prove that a viscosity solution exists. Often, the value function of a control problem or the solution to a backward stochastic differential equation (BSDE) can be shown to be the very viscosity solution we are looking for, providing a beautiful link between probability and PDEs.

Pillar 2: Uniqueness via the Comparison Principle

Having a solution is good, but having the solution is what makes a theory predictive. The crown jewel of viscosity theory is the ​​comparison principle​​. In its simplest form, it states that if a viscosity subsolution $u$ starts out below a viscosity supersolution $v$ (e.g., at an initial or terminal time), then it must remain below it for all time. That is, if $u(x,0) \le v(x,0)$ for all $x$, then $u(x,t) \le v(x,t)$ for all later $t$.

This seemingly modest principle has a monumental consequence: it implies uniqueness. If you had two different continuous viscosity solutions, $V_1$ and $V_2$, starting with the same initial data, you could apply the principle once with $u=V_1, v=V_2$ to get $V_1 \le V_2$, and a second time with $u=V_2, v=V_1$ to get $V_2 \le V_1$. The only possibility is that $V_1=V_2$. Therefore, if we can find a viscosity solution—for instance, by constructing it as the value function of a control problem—the comparison principle guarantees it is the one and only solution.

The power of comparison is beautifully illustrated in geometry. Imagine two disjoint soap bubbles floating in space. Their surfaces evolve according to mean curvature flow. The ​​avoidance principle​​ states that these two bubbles will never touch each other as they evolve. This physical intuition is a direct mathematical consequence of the comparison principle applied to the level-set PDE that describes the flows. The functions describing the two bubbles are ordered, and this ordering prevents their zero-level sets (the surfaces of the bubbles) from ever intersecting.

Pillar 3: Stability

The final pillar is stability, which makes the theory robust and practical. The ​​stability theorem​​ for viscosity solutions is an engineer's dream: it guarantees that solutions depend continuously on the data of the problem. If you have a sequence of problems whose coefficients and costs $(b_n, \sigma_n, f_n, g_n)$ converge to a limiting set of data $(b, \sigma, f, g)$, then the corresponding value functions $V_n$ will converge to the value function $V$ of the limiting problem.

This property is what breathes life into numerical methods and approximation schemes. For example, to solve a problem on a complex domain $D$, we can solve it on a sequence of much simpler, expanding domains $D_n$ that eventually fill $D$. The stability theorem ensures that the sequence of simple solutions $u_n$ will converge to the true solution $u$ on the complex domain $D$. It provides the rigorous justification for why our approximations work.

Stability also allows the theory to gracefully handle "rough" data. What if the terminal cost function $g(x)$ in our control problem is discontinuous, perhaps representing a fixed reward for landing in a specific winning region and zero elsewhere? The viscosity framework doesn't break down. It naturally incorporates the discontinuity by relaxing the terminal condition. Instead of requiring the solution $V$ to equal $g$ at the boundary, it requires the upper limit of $V$ to be less than or equal to the upper envelope of $g$, and the lower limit of $V$ to be greater than or equal to the lower envelope of $g$. This relaxed condition is exactly what's needed for the comparison principle—and thus uniqueness—to hold, even in this rough setting.

In the end, the journey into viscosity solutions starts with a crisis—the breakdown of classical calculus in the face of non-smooth reality. But it leads us to a theory of remarkable depth and elegance. By redefining what it means to be a solution, we unlock a unified framework that not only solves the original problem but also provides guarantees of existence, uniqueness, and stability, revealing deep and unexpected connections between fields as disparate as stochastic control, geometric analysis, and finance. It is a testament to the power of finding the right perspective—the right "viscosity"—with which to view the world.

In our previous discussion, we meticulously assembled the theoretical machinery of viscosity solutions. We saw why such a concept is necessary—to make sense of equations whose solutions are not content to be smooth and well-behaved. We now stand at a delightful vantage point. From here, we can look out over the vast landscape of science and engineering and see where this machinery, which might have seemed abstract, makes a profound and tangible impact. You will see that viscosity solutions are not merely a technical fix; they are a unifying language that reveals deep and often surprising connections between seemingly disparate fields, from the flow of water to the flow of financial markets.

Let's begin with the name itself: "viscosity" solution. Is this just a poetic label? Not at all. It hints at a deep physical intuition. Many of the first-order partial differential equations we are interested in, like the Hamilton-Jacobi equations, arise as idealizations of physical systems where dissipative forces like friction or diffusion have been ignored. A beautiful and powerful idea is to reintroduce a tiny amount of this "viscosity" or diffusion back into the equation. This typically involves adding a small second-order term, often a Laplacian like $\epsilon \Delta u$, which has a smoothing effect. The resulting equation is now better behaved and often possesses a unique, smooth classical solution. The magic happens when we then ask: what happens to this smooth solution as we slowly dial the viscosity term $\epsilon$ back down to zero? The limit we obtain is precisely the physically relevant, and often non-smooth, viscosity solution of the original, idealized equation. The solution "remembers" the ghost of the friction that was taken away.

This idea finds its most dramatic expression in the study of conservation laws, which govern everything from traffic flow to fluid dynamics. Consider the famous Burgers' equation, $u_t + (\frac{1}{2}u^2)_x = 0$, a simplified model for the motion of a gas. If you start with a smooth profile, it will often steepen over time and try to become multi-valued—an obvious physical impossibility. What happens in reality is that a "shock wave" forms, a sharp discontinuity where the solution jumps. The trouble is, the mathematics of weak solutions permits infinitely many possible shock solutions, yet nature chooses only one. Which one? The one that respects thermodynamics—the one that dissipates energy correctly. This physical selection rule is called an "entropy condition." The breathtaking insight, a cornerstone of the whole theory, is that the viscosity solution of the corresponding Hamilton-Jacobi equation, $v_t + \frac{1}{2}(v_x)^2 = 0$, provides exactly this entropy condition when we relate the two by $u = v_x$. The purely mathematical definition of a viscosity solution—a definition of touching test functions from above and below—perfectly singles out the one shock wave that is physically correct.

The reach of viscosity solutions extends far beyond fluids and into the very geometry of the world around us. One of the most fundamental geometric equations is the Eikonal equation, $|\nabla u| = n(x)$, where $u(x)$ represents the time it takes for a wave to travel from a source to a point $x$, and $n(x)$ is the local refractive index (or slowness) of the medium. This equation is the foundation of geometric optics. When you see the shimmering distortion above a hot road or the intricate patterns of light at the bottom of a swimming pool (caustics), you are witnessing the complexities of the Eikonal equation. Classical methods like ray tracing break down at these caustics, but the viscosity solution framework gives us a global, continuous solution for the wavefront's arrival time, even in a wildly inhomogeneous medium like the Earth's crust for a seismologist, or a complex biological tissue for medical imaging.

From static shapes, we can leap to dynamic ones. Imagine a cluster of soap bubbles. To minimize surface energy, their surfaces are constantly in motion, a process called mean curvature flow. As bubbles merge or pop, they create singularities—sharp corners and topological changes that are anathema to classical PDE theory. The level-set method, pioneered by Stanley Osher and James Sethian, brilliantly sidesteps this by representing the moving surface as the zero-level set of a higher-dimensional function $u(x,t)$. The evolution of this function $u$ is governed by a complex, degenerate elliptic PDE. The theory of viscosity solutions provides the perfect, robust framework for this equation. It allows for the formation of singularities and changes in topology with utter mathematical grace. A beautiful consequence of the theory's comparison principle is the "avoidance principle": if two evolving surfaces start out separate, their level-set functions ensure they will never touch. This powerful technique is now a workhorse in fields like image processing (for object segmentation), computer graphics (for fluid simulation), and materials science (for crystal growth).

Perhaps the most natural home for viscosity solutions is in the world of optimal control theory. This is the science of making the best decisions over time to achieve a goal. Whenever you use your phone's GPS to find the fastest route, you are solving an optimal control problem. The master equation of this field is the Hamilton-Jacobi-Bellman (HJB) equation. Its solution, the "value function," represents the best possible outcome you can achieve starting from any given state.

However, the value function is almost never smooth. It typically has "kinks" or "corners" at points where the optimal strategy abruptly changes (e.g., "turn left now!"). For decades, this lack of smoothness was a major stumbling block. The arrival of viscosity solutions was a revolution. They provided a framework in which the value function was always guaranteed to exist and be unique. The theory was so perfectly suited to the problem that one could say the HJB equation was waiting for viscosity solutions to be invented.

This power becomes even more apparent when we introduce uncertainty. Imagine a robot navigating a cluttered room with noisy sensors, or a central bank setting interest rates in a volatile economy. These are problems of stochastic optimal control. The HJB equation gains a second-order term representing the diffusion or randomness. Here again, viscosity solutions provide a rigorous a framework for finding the optimal strategy, even when the randomness is degenerate—that is, it doesn't affect the system in all directions. A related and equally important application is not just in finding optimal paths, but in proving that a system is stable. Classical Lyapunov theory required finding a smooth "energy-like" function that always decreases. Viscosity solutions allow us to generalize this powerful idea to non-smooth Lyapunov functions, greatly expanding the class of complex engineering and biological systems whose stability we can rigorously verify.

The journey does not end there. In the most advanced realms of mathematics, viscosity solutions serve as a bridge to even deeper, more abstract structures. One of the most profound discoveries is the connection between PDEs and Backward Stochastic Differential Equations (BSDEs). A BSDE is a strange and wonderful object: an SDE that is specified by a terminal condition and solved backwards in time. The "nonlinear Feynman-Kac formula" establishes that the viscosity solution to a large class of semilinear parabolic PDEs has a purely probabilistic representation in terms of a coupled system of forward and backward SDEs. This duality is now a cornerstone of modern mathematical finance, where it is used to price and hedge complex financial derivatives in the presence of market frictions or default risk.

The robustness of the theory is such that it can even be pushed into infinite dimensions. Many real-world systems, particularly in finance, have "memory"—their future evolution depends not just on their current state, but on their entire past history (think of an option whose payoff depends on the average price of a stock over a month). The governing equations become path-dependent PDEs (PPDEs), defined on a space of functions. Astonishingly, the entire framework of viscosity solutions has been successfully extended to this mind-bogglingly abstract setting, taming these infinite-dimensional beasts and providing a rigorous foundation for pricing path-dependent financial instruments.

Finally, we come to a connection that feels like a piece of pure magic: Large Deviation Theory. This theory studies the probability of extremely rare events in systems subject to small random perturbations. Think of a particle in a potential well; quantum mechanics allows it to tunnel out, but classical mechanics with a bit of random noise (temperature) also allows it to be "kicked" over the barrier. What is the probability of this happening? Freidlin-Wentzell theory shows that this probability vanishes exponentially as the noise level $\epsilon$ goes to zero, and the rate of decay is governed by the value function of a related deterministic optimal control problem. This value function, which tells you the "cheapest" way to force the rare event to happen, is—you guessed it—the unique viscosity solution to a Hamilton-Jacobi equation. This beautiful result creates a three-way bridge between probability (rare events), deterministic mechanics (optimal control), and analysis (PDEs), showing how the most likely way for an unlikely event to occur is, in a deep sense, the most optimal path.

From the crash of a wave to the path of a photon, from the choice of an investor to the stability of a robot, viscosity solutions have revealed themselves to be an essential part of the mathematical toolkit. They are a testament to the power of a good definition—one that not only solves a problem, but reveals a new layer of unity and beauty in the universe.