A Measure-Valued Obstacle Problem for an Obliquely Reflected Diffusion with a Max-Type Payoff

Technical Summary: A Measure-Valued Obstacle Problem for an Obliquely Reflected Diffusion with a Max-Type Payoff

1. Problem Formulation

The paper investigates an optimal stopping problem for a two-dimensional obliquely reflected diffusion process $X = (X_1, X_2)$ confined to the non-negative quadrant $\mathbb{R}^2_+$. The process evolves according to the stochastic differential equation:
$$dX_t = \beta(X_t) dt + \sigma(X_t) dW_t + \rho_1 dL^1_t + \rho_2 dL^2_t,$$
where $L^1$ and $L^2$ are boundary local times enforcing reflection along fixed inward vectors $\rho_1 = (1, \theta_1)$ and $\rho_2 = (\theta_2, 1)$.

The objective is to maximize the expected discounted payoff:
$$V(x) = \sup_{\tau} \mathbb{E}_x \left[ e^{-r\tau} G(X_\tau) - \int_0^\tau e^{-rs} c(X_s) ds \right],$$
where the reward function is a nonsmooth "max-type" payoff $G(x) = x_1 \vee \alpha x_2$ (with $\alpha > 0$), $r > 0$ is the discount rate, and $c \geq 0$ is a running cost.

Formally, the value function $V$ is expected to satisfy a reflected obstacle problem involving the generator $L$ of the diffusion:
$$\max\{LV - rV - c, G - V\} = 0 \quad \text{in } \mathbb{R}^2_+, \quad \rho_i \cdot \nabla V = 0 \quad \text{on } F_i,$$
where $F_i$ are the coordinate faces. The primary analytical challenge arises from the simultaneous presence of:

1. Oblique reflection: The reflection vectors $\rho_i$ are not necessarily normal to the boundary.
2. Nonsmooth obstacle: The payoff $G$ has a "kink" along the diagonal $\Delta = \{x_1 = \alpha x_2\}$, causing its gradient to jump.
3. Killing: The process stops upon entering the stopping set $D = \{V = G\}$.

2. Methodology and Analytical Framework

The authors develop a measure-valued potential-theoretic formulation to handle the singularities introduced by the nonsmooth payoff and the oblique reflection.

• Distributional Generator of the Payoff: The paper analyzes the distributional second derivative of $G$. Because $G$ is affine on either side of $\Delta$ but has a gradient jump across it, the generator $LG$ is not a function but a signed measure containing a singular component supported on $\Delta$. Specifically, $LG$ includes a term proportional to the one-dimensional surface measure $\sigma_\Delta$.
• Reflected Itô–Tanaka Formula: The authors derive a generalized Itô–Tanaka formula for $G(X_t)$. This formula reveals two distinct singular mechanisms:
  1. A diagonal local time term ($d\ell_0(Y)$ where $Y = X_1 - \alpha X_2$) arising from the kink in $G$.
  2. Boundary local time terms arising from the oblique reflection. Since $\nabla G$ is not necessarily orthogonal to the reflection vectors $\rho_i$ on the faces $F_i$, the reflection contributes additional boundary terms proportional to $\rho_i \cdot \nabla G$.
• Total Signed Stopping Measure: These components are unified into a total signed measure $\Gamma_{tot}$:
  $$\Gamma_{tot} = \Gamma_{ac} dx + \Gamma_\Delta + \Gamma_\partial,$$
  where $\Gamma_{ac}$ is the absolutely continuous interior gain, $\Gamma_\Delta$ is the singular measure on the diagonal, and $\Gamma_\partial$ is the signed boundary Revuz measure generated by the reflection.
• Killed Resolvent Representation: The paper establishes that the value function in the continuation region $C = \{V > G\}$ is represented by a killed resolvent of $\Gamma_{tot}$:
  $$V(x) = G(x) - R^C_r \Gamma_{tot}(x), \quad x \in C.$$
  Crucially, the authors demonstrate that the unrestricted reflected resolvent (integrating over the entire infinite horizon) is generally incorrect. This is because the process is not absorbed upon hitting the stopping set $D$; it may re-enter the continuation region $C$, rendering the unrestricted potential ill-defined for the variational problem.

3. Key Contributions and Results

A. Measure-Valued Formulation
The paper provides a rigorous bridge identity linking the discounted payoff to the total signed stopping measure:
$$e^{-rt}G(X_t) - \int_0^t e^{-rs}c(X_s)ds = G(x) - \int_0^t e^{-rs} dA^{\Gamma_{tot}}_s + M^G_t.$$
This identity serves as the analytic foundation for the subsequent analysis, treating the singularities as measures rather than attempting to smooth them.

B. Failure of Unrestricted Resolvents
A significant theoretical result is the proof that $V \neq G - R^R_r (\Gamma_{tot} \mathbf{1}_C)$ in general. The authors show that the unrestricted resolvent fails because it counts "occupation" of the continuation region after the process has already stopped (and potentially re-entered), which lacks variational meaning. The correct object is the killed resolvent, which integrates only up to the first exit time from $C$.

C. Free Boundary Characterization
The free boundary is characterized not by a classical smooth-fit condition, but by a continuation-side trace condition for the killed potential:
$$\text{Tr}_{\partial C} R^C_r \Gamma_{tot} = 0.$$
This condition is necessary because the standard smooth fit ($C^1$ or $C^2$ matching) breaks down due to the diagonal singularity and the boundary local-time contributions.

D. Verification Theorem for Epigraph Candidates
Under a vertical monotonicity hypothesis (assuming $V - G$ is non-increasing in $x_2$), the stopping set $D$ is shown to have an epigraph form $\{x_2 \geq b(x_1)\}$. The paper proves a verification theorem stating that any admissible epigraph candidate $b$ satisfying the following conditions coincides with the true value function $V$:

1. Contact and Strict Continuation: $U_b = G$ on $D_b$ and $U_b > G$ on $C_b$.
2. Reflected Neumann Compatibility: $\rho_i \cdot \nabla U_b = 0$ on the boundaries.
3. Trace Condition: The killed potential vanishes at the boundary from the continuation side.
4. Measure-Superharmonicity: $(r - L)U_b + c \geq 0$ as a signed smooth measure.
5. Geometric Constraint: The free boundary must lie on or above the diagonal ($b(x_1) \geq x_1/\alpha$) to avoid the singular measure $\Gamma_\Delta$ entering the interior of the stopping set.

4. Significance and Claims

The paper explicitly claims to provide a rigorous verification framework rather than a solution to the free boundary itself. Its significance lies in:

• Unifying Singularities: Successfully combining the diagonal singular measure (from the max-payoff) with boundary measures (from oblique reflection) into a single measure-theoretic framework.
• Correcting Potential Theory: Correcting the common misconception that unrestricted resolvents apply to non-absorbing stopping problems, establishing the killed resolvent as the correct tool.
• Generalizing Boundary Conditions: Formulating the free boundary condition as a trace condition on the killed potential, which is robust enough to handle the lack of smooth fit caused by the kink and oblique reflection.

The authors note that deriving the monotonicity of $V-G$ from primitive model parameters and establishing the regularity of the free boundary remain open analytical questions. The work is positioned as a foundational step in the intersection of optimal stopping, reflected diffusions, and measure-valued potential theory.