Delay-Penalty Comparison for Sequential Testing and Quickest Detection in State-Dependent Diffusion Models

Technical Summary: Delay-Penalty Comparison for Sequential Testing and Quickest Detection in State-Dependent Diffusion Models

Problem Statement
The paper addresses two fundamental problems in sequential analysis: sequential hypothesis testing and Bayesian quickest detection (the disorder problem). In both settings, an observer monitors a diffusion process $X_t$ whose drift switches between two alternatives ($\mu_0$ and $\mu_1$) at an unobservable time or under an unknown hypothesis. The distinguishing feature of this work is the assumption of state-dependent signal-to-noise ratios, where the diffusion coefficient $\sigma(X_t)$ and the drift difference $\mu_1(X_t) - \mu_0(X_t)$ depend on the current observation.

In standard models with constant signal-to-noise ratios, the posterior probability (or likelihood ratio) forms a closed one-dimensional Markov state. However, when the signal-to-noise ratio is state-dependent, the posterior probability alone is insufficient to describe the system's evolution. The natural sufficient statistic becomes the augmented process $(\Pi_t, X_t)$, consisting of the posterior probability $\Pi_t$ and the observed diffusion $X_t$. This results in a multidimensional, degenerate diffusion process. The paper focuses on the comparative-statics of these problems: specifically, how the optimal stopping rules (alarm thresholds) change when the structure of the delay penalty (running cost) changes, rather than seeking explicit solutions for specific models.

Methodology
The authors employ a unified framework based on optimal stopping theory and filtering:

1. Filtering Reduction: The paper first establishes that for state-dependent diffusions, the pair $(\Pi_t, X_t)$ (or equivalently the odds process $(\Phi_t, X_t)$) constitutes a time-homogeneous Markov sufficient statistic. The authors derive the explicit degenerate generator $\mathcal{L}$ for this augmented process. The degeneracy arises because both coordinates are driven by a single innovation Brownian motion, resulting in a diffusion matrix with a zero determinant.
2. Optimal Stopping Formulation: Both sequential testing and quickest detection are reduced to minimizing a cost functional of the form:
   $$V(y) = \inf_{\tau} \mathbb{E}_y \left[ \int_0^\tau f(Y_s) ds + G(Y_\tau) \right]$$
   where $Y$ is the Markov state, $f$ is the running cost (delay or sampling cost), and $G$ is the terminal cost (error or false-alarm cost). The value function $V$ satisfies a variational inequality (or complementarity system) involving the generator $\mathcal{L}$.
3. Comparison Principle: The core methodological contribution is a delay-penalty comparison theorem. Instead of solving the variational inequality, the authors compare two problems with the same terminal cost $G$ and the same state process $Y$, but with different running costs $f_1$ and $f_2$. By establishing a pointwise ordering of the running costs, they derive an ordering of the value functions, continuation regions, and optimal stopping times.

Key Contributions and Results

• Unified Markov State Identification: The paper rigorously proves that for state-dependent diffusions, the posterior probability is not a closed Markov state. The closed sufficient statistic is the augmented process $(\Pi_t, X_t)$, and the authors provide the explicit generator for this process, highlighting its degenerate elliptic nature.
• Delay-Penalty Comparison Theorem: The primary theoretical result (Theorem 6) states that for a fixed terminal cost, if the running delay penalty $f_1$ is pointwise larger than $f_2$ ($f_1 \ge f_2$), then:
  1. The value function increases ($V_1 \ge V_2$).
  2. The continuation region shrinks ($C_1 \subseteq C_2$).
  3. The optimal stopping time occurs earlier ($\tau^*_1 \le \tau^*_2$) almost surely.
• Boundary Monotonicity: If the stopping region admits a one-sided representation in the posterior coordinate (i.e., stopping occurs when $\Pi_t \ge b(X_t)$), the comparison theorem implies a monotone ordering of the alarm boundaries: $b_1(x) \le b_2(x)$. This means a higher delay penalty lowers the threshold for raising an alarm.
• Extension to Nonlinear Penalties: The framework extends to nonlinear delay costs (e.g., convex penalties) by augmenting the state space with a "penalty-memory" statistic (the conditional expectation of the marginal penalty). The comparison principle holds as long as the marginal penalties are pointwise ordered.
• Numerical Illustration: The authors provide a worked example using the Shiryaev model with a constant signal-to-noise ratio (where the state reduces to one dimension). They solve the variational inequality numerically using a finite-difference method with projected iteration. The results confirm the theoretical prediction: as the linear delay cost rate $c$ increases, the optimal posterior threshold $p^*(c)$ decreases monotonically.

Significance and Claims
The paper positions itself as a structural rather than constructive contribution. It does not aim to provide new explicit formulas for free boundaries in complex state-dependent models, which are often intractable. Instead, it offers a comparison principle that clarifies the qualitative behavior of optimal rules under varying penalty structures.

The authors claim that this framework:

• Clarifies how state-dependent information and nonlinear delay costs jointly affect the geometry of sequential testing and quickest-detection rules.
• Provides a robust tool for practitioners to understand how alarm thresholds shift when penalty structures change (e.g., in financial regime switching, structural health monitoring, or epidemiological surveillance) without needing to re-solve the full optimal stopping problem.
• Applies broadly to linear and nonlinear delay costs, provided the underlying Markovian structure of the augmented state is maintained.

The paper explicitly notes that the one-sided boundary structure required for the specific boundary ordering result relies on conditions (such as monotone signal-to-noise ratios) established in prior literature (e.g., Gapeev, Shiryaev, Ernst, and Peskir), and does not attempt to prove the existence of such structures for arbitrary state-dependent diffusions.