🔢 mathematics

Reflected Optimal Stopping with a Max-Type Payoff: Measure-Valued Stopping Gains and Killed Resolvent Representation

This paper analyzes an infinite-horizon optimal stopping problem for a two-dimensional reflected diffusion with a max-type payoff, demonstrating that the associated stopping gain is a signed measure with a singular diagonal component and establishing that the value function is correctly represented via a killed resolvent formula rather than an unrestricted one.

Original authors: Ye Liang

Published 2026-07-14

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ye Liang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are playing a high-stakes video game where you control a glowing ball bouncing around inside a giant, invisible box. This box has two walls: a floor and a left wall. If your ball hits the floor, it bounces straight up. If it hits the left wall, it bounces straight right. It never slides along the wall; it always gets a perfect, bouncy push back into the center. This is what mathematicians call a "normally reflected diffusion."

Your goal in the game is to decide the perfect moment to stop the ball and cash out. But here's the twist: the prize money isn't just based on where the ball is. The reward is the maximum of two things: your horizontal distance from the start ( $x_1$ ) OR a scaled version of your vertical distance ( $\alpha x_2$ ). Think of it like this: you get paid based on whichever coordinate is bigger. If you are far to the right, you get paid for that. If you are high up, you get paid for that. But there's a sharp, jagged edge where these two rules meet—a "kink" in the middle of the map.

This paper, written by Ye Liang, tackles the tricky math of finding that perfect stopping moment in this bouncy, two-dimensional world.

The Big Surprise: The "Ghost" Penalty

In simpler games (one-dimensional ones), you can usually calculate the value of stopping by looking at a smooth curve. But in this game, because the reward function has a sharp corner (the kink where $x_1 = \alpha x_2$ ), the math gets messy.

The authors discovered that the "gain" you get from stopping isn't just a simple number or a smooth function. It's actually a signed measure. To use a metaphor: imagine the reward map has a hidden, invisible "ghost penalty" running along that sharp diagonal line where the two reward rules meet.

Usually, you might think, "If I stop here, I gain this much." But the paper proves that right on that diagonal line, there is a sudden, negative spike in the math—a "signed measure" that acts like a sudden drop in value. This isn't a smooth slope; it's a sharp, singular tear in the fabric of the calculation. The authors explicitly show that if you try to ignore this ghost penalty and treat the reward as a smooth function, your math breaks. You must account for this invisible, negative spike along the diagonal line to get the right answer.

The Trap of the "Unrestricted" View

Another major finding is about how we look at the future. A common mistake in these types of problems is to calculate the value of the game as if the ball keeps bouncing forever, even after you've decided to stop.

The paper argues against this "unrestricted" view. They prove that the correct way to calculate the value is to use a "killed" resolvent. Think of it like this: once you decide to stop the game, the ball doesn't just keep bouncing around in the background, potentially coming back to a "good" spot later. No, the moment you stop, the game ends. The ball vanishes.

The authors show that if you use the "unrestricted" method (pretending the ball keeps bouncing), you might accidentally count points the ball would have earned if it kept playing. But since you stopped, those points don't exist. The correct formula is: Value = Current Reward minus the "Killed" Potential. This "killed" potential only counts the time the ball spent before you stopped. It's a strict "stop and count" rule, not a "keep watching" rule.

What They Don't Claim

It's important to know what this paper doesn't say. The authors are very careful not to promise that they can always draw a perfect, smooth line for where you should stop. They don't claim to have solved the mystery of exactly how smooth that stopping line is. Instead, they provide a rigorous verification theorem.

This means they say: "If you guess a stopping line, and you can prove it meets these specific, strict conditions (like checking that the ghost penalty is handled correctly and the ball doesn't bounce back in after stopping), then your guess is the winner." They don't tell you exactly what the line looks like for every possible game; they give you the checklist to prove that your specific line is the right one.

The "Epigraph" Shape

The paper also suggests that under certain conditions (like if the reward behaves nicely as you move up or down), the area where you should stop looks like an "epigraph." In plain English, this means the "stop zone" is everything above a certain wavy line. If you are above the line, stop. If you are below it, keep playing. The authors show that if the "advantage" of stopping behaves in a specific way (getting smaller as you go up), then this "stop above the line" shape is guaranteed.

The Bottom Line

Ye Liang's work is a rigorous correction of how we do the math for these bouncy, two-dimensional games with sharp corners.

The Kink is Real: You cannot ignore the sharp diagonal line where the reward rules switch. It creates a "signed measure"—a mathematical ghost penalty that must be calculated explicitly.
Stop the Clock: You must calculate the value based on the time before you stop, not the time after. The "killed" formula is the only correct one; the "unrestricted" one is wrong.
Verification, Not Magic: The paper doesn't magically draw the perfect stopping line for you. Instead, it gives you a strict, step-by-step checklist to verify if a proposed stopping line is actually the best one.

The authors prove these points using advanced tools like "Itô–Krylov–Tanaka" formulas (which are like super-charged versions of the rules for how things move and change) and "viscosity solutions" (a way of solving equations even when they have sharp corners). They don't just suggest these ideas; they provide a formal proof that these specific corrections are necessary to get the right answer in this complex, bouncy world.

Technical Summary: Reflected Optimal Stopping with a Max-Type Payoff

Problem Statement
This paper investigates an infinite-horizon optimal stopping problem for a two-dimensional normally reflected diffusion $X = (X_1, X_2)$ evolving in the quadrant $\mathbb{R}^2_+ = [0, \infty)^2$ . The objective is to maximize the expected discounted payoff minus the accumulated running cost:
$V(x) = \sup_{\tau \in \mathcal{T}} \mathbb{E}_x \left[ e^{-r\tau}G(X_\tau) - \int_0^\tau e^{-rs}c(X_s) ds \right]$
where the payoff function is of the "max-type" form $G(x_1, x_2) = x_1 \vee \alpha x_2$ (with $\alpha > 0$ ), $r > 0$ is the discount rate, and $c \ge 0$ is a running cost. The process $X$ is constrained by normal reflection on the coordinate axes.

The problem presents three specific analytical challenges that complicate standard free-boundary analysis:

Reflection: The process is constrained to the quadrant by normal reflection, requiring Neumann boundary conditions.
Nonsmooth Payoff: The reward function $G$ has a "kink" along the diagonal set $\Delta := \{x_1 = \alpha x_2\}$ , meaning it is not $C^2$ across this line.
Multidimensionality: The stopping region is genuinely two-dimensional, unlike many solvable one-dimensional cases.

Methodology and Framework
The authors formulate the problem using a reflected obstacle problem framework, interpreted in the viscosity sense. However, the paper emphasizes that viscosity supersolution status alone is insufficient for rigorous verification in this context. Instead, the methodology relies on:

Measure-Valued Stopping Gains: The authors define the stopping-gain object $\Gamma := c + rG - LG$ . Due to the kink in $G$ , $\Gamma$ is not a function but a signed Radon measure. It decomposes into an absolutely continuous part ( $\Gamma_{ac}$ ) away from the diagonal and a singular surface measure ( $\Gamma_\Delta$ ) supported on $\Delta$ .
Itô–Krylov–Tanaka Admissibility: Verification theorems are established under explicit assumptions that the candidate value function belongs to a class where a generalized Itô formula (incorporating Tanaka's formula for the local time of the kink) holds.
Killed Resolvents: The potential representation of the value function is derived using a "killed" resolvent, which accumulates the stopping-gain measure only up to the first entry time into the stopping set, rather than over the entire infinite horizon of the unrestricted process.

Key Contributions and Results

Verification Theorem under Measure Superharmonicity:
The paper proves a verification theorem (Theorem 3.2) that separates the verification argument from hidden regularity theorems. It requires the candidate function $u$ to satisfy:
- Global measure-superharmonicity: $(L-r)u - c \le 0$ in the sense of signed measures.
- Explicit Itô–Krylov–Tanaka admissibility, ensuring the generalized Itô formula applies.
- Normal reflection boundary conditions.
  The theorem clarifies that viscosity status does not automatically imply the existence of a signed-measure representative or the validity of the generalized Itô formula.
Singular Stopping-Gain Measure:
A central technical result (Theorem 3.10) explicitly computes the singular component of the stopping-gain measure generated by the kink in $G$ . The measure decomposes as $\Gamma = \Gamma_{ac} dx + \Gamma_\Delta$ , where the diagonal component is:
$\Gamma_\Delta(dx) = -\frac{q(x)}{2\sqrt{1+\alpha^2}} \sigma_\Delta(dx)$
Here, $q(x) = n^\top a(x) n$ with $n=(1, -\alpha)$ , and $\sigma_\Delta$ is the one-dimensional surface measure on $\Delta$ . Since $q(x) \ge 0$ , the singular component $\Gamma_\Delta$ is non-positive. This implies that the standard pointwise condition " $\Gamma \ge 0$ on the stopping set" cannot be applied literally to the full measure if the stopping set intersects $\Delta$ with positive length; the absolutely continuous and singular parts must be treated separately.
Killed-Resolvent Representation:
The paper derives the correct potential representation for the value function (Theorem 3.13):
$V(x) = G(x) - \mathcal{R}^C_r \Gamma(x)$
where $\mathcal{R}^C_r$ is the resolvent killed at the first entry time $\tau_D$ into the stopping set $D$ . The authors demonstrate that the unrestricted reflected resolvent $\mathcal{R}^R_r$ is generally incorrect because the unrestricted process may re-enter the continuation region after stopping, accumulating irrelevant occupation time.
Conditional Geometry of the Stopping Set:
Under structural monotonicity assumptions on the stopping advantage $H = V - G$ , the paper establishes a conditional epigraph structure for the stopping set (Theorem 3.7). Specifically, if $H$ is vertically non-increasing, the stopping region takes the form $D = \{(x_1, x_2) : x_2 \ge b(x_1)\}$ .
Candidate-Boundary Verification:
The paper provides a rigorous framework for verifying a proposed boundary (Theorem 3.17). It highlights that solving a boundary-trace equation is insufficient. A valid candidate must satisfy global measure-superharmonicity, which imposes a strict compatibility condition: the stopping set cannot contain a positive-length portion of the diagonal $\Delta$ in the interior unless the singular measure vanishes there (which generally requires the diagonal to lie in the continuation region).

Significance and Claims
The paper claims to provide a rigorous structural framework for optimal stopping problems involving reflected diffusions and nonsmooth payoffs. Its primary significance lies in correcting three common formal steps often taken in such problems:

Regularity: It asserts that viscosity supersolution status does not infer the Sobolev or measure-theoretic regularity required for verification; these must be assumed explicitly.
Kink Contribution: It treats the payoff kink not as a negligible irregularity but as a source of a singular signed surface measure, which fundamentally alters the sign conditions on the stopping set.
Resolvent Type: It corrects the potential representation by insisting on a killed resolvent rather than an unrestricted one, ensuring the potential is accumulated only before the stopping time.

The authors maintain a modest stance regarding regularity, noting that the framework provides a "rigorous conditional structure" rather than unconditional free-boundary regularity. The results are illustrated using a constant-coefficient reflected Brownian motion example to show that the singular diagonal term is an intrinsic feature of the max-type payoff, not an artifact of variable coefficients.