[Paper Review] Quickest Change-Point Detection: A Bird's Eye View

Aleksey S. Polunchenko, Grigory Sokolov+1 more|arXiv (Cornell University)|Oct 11, 2013

Advanced Statistical Process Monitoring38 references18 citations

TL;DR

This paper provides a comprehensive overview of sequential change-point detection in discrete-time settings with known pre- and post-change distributions, analyzing four major formulations: Bayesian, generalized Bayesian, minimax, and multi-cyclic. It establishes that the Shiryaev–Roberts–r procedure with an optimal head start $ r^* $ achieving $ C(r^*) = C_{ ext{infty}} $ is third-order asymptotically optimal, significantly improving delay detection over standard SR and SR–r procedures.

ABSTRACT

We provide a bird's eye view onto the area of sequential change-point detection. We focus on the discrete-time case with known pre- and post-change data distributions and offer a summary of the forefront asymptotic results established in each of the four major formulations of the underlying optimization problem: Bayesian, generalized Bayesian, minimax, and multi-cyclic.

Motivation & Objective

• To provide a unified, bird’s-eye view of the four major formulations of sequential change-point detection: Bayesian, generalized Bayesian, minimax, and multi-cyclic.
• To clarify the asymptotic performance of key detection procedures under each formulation, particularly focusing on detection delay and false alarm control.
• To establish conditions under which the Shiryaev–Roberts–r procedure achieves third-order asymptotic optimality in the minimax setting.
• To identify the optimal head start $ r^* $ that equalizes detection delays under both pre- and post-change regimes, enhancing robustness and performance.
• To quantify the improvement in detection delay over standard SR and SR–r procedures via explicit asymptotic expansions involving Kullback–Leibler divergence and overshoot constants.

Proposed method

• Formalizes the change-point model with a random change-point $ \nu \in \{0, \infty\} $, where observations follow $ \mathbb{P}_\infty $ before $ \nu $ and $ \mathbb{P}_0 $ after.
• Defines detection procedures as stopping times $ T $ adapted to the observation filtration $ \mathcal{F}_n $, with performance measured via average run length to false alarm $ \mathbb{E}_\infty[T] $ and expected detection delay $ \text{SADD}(T) $.
• Introduces the likelihood ratio process $ \Lambda_n = \frac{d\mathbb{P}_0^{(n)}}{d\mathbb{P}_\infty^{(n)}} $, and uses it to define the Shiryaev–Roberts (SR) and SR–r statistics $ R_n^{(r)} = \sum_{k=1}^n \Lambda_k + r $.
• Applies large-sample asymptotic theory: derives the overshoot $ \kappa_a = S_{\tau_a} - a $, and defines limiting constants $ \varkappa = \lim_{a\to\infty} \mathbb{E}_0[\kappa_a] $, $ \zeta = \lim_{a\to\infty} \mathbb{E}_0[e^{-\kappa_a}] $, and $ C_{\infty} = \mathbb{E}[\log(1 + R_\infty + \tilde{V}_\infty)] $.
• Uses the Kullback–Leibler divergence $ I = \mathbb{E}_0[\log \Lambda_1] $ as the key information-theoretic quantity to characterize detection performance.
• Derives asymptotic expansions for $ \text{SADD}(T) $ and $ \mathbb{E}_\infty[T] $, showing that the SR–r procedure with $ A = \gamma \zeta $ and $ r = r^* $ satisfying $ C(r^*) = C_{\infty} $ achieves third-order optimality.

Experimental results

Research questions

• RQ1 What is the asymptotic performance of the Shiryaev–Roberts–r procedure in terms of detection delay and false alarm rate?
• RQ2 How does the choice of initialization point $ r $ affect the balance between detection delay under the pre- and post-change regimes?
• RQ3 Can the SR–r procedure be made asymptotically optimal in the third-order sense, and if so, under what conditions?
• RQ4 What is the role of the overshoot constants $ \varkappa $ and $ \zeta $, and how do they influence the asymptotic detection delay?
• RQ5 How does the generalized SR–r procedure compare to the standard SR and SR–r procedures in terms of minimax optimality?

Key findings

• The SR–r procedure with $ A = \gamma \zeta $ and $ r = r^* $ satisfying $ C(r^*) = C_{\infty} $ achieves third-order asymptotic optimality, meaning its detection delay matches the theoretical lower bound up to $ o(1) $ as $ \gamma \to \infty $.
• The optimal head start $ r^* $ is a fixed constant independent of $ \gamma $, enabling effective and robust initialization without re-optimization for each false alarm constraint.
• The difference in expected detection delay between the standard SR procedure (starting from 0) and the optimized SR–r procedure is asymptotically $ (C(0) - C_{\infty})/I $, which can be substantial when the Kullback–Leibler divergence $ I $ is small.
• The asymptotic expansion for $ \text{SADD}(\mathcal{S}_A^r) $ is $ (1/I)[\log(\gamma \zeta) + \varkappa - C_{\infty}] + o(1) $, showing the dependence on $ \gamma $, $ \zeta $, and the overshoot constant $ \varkappa $.
• The expected false alarm time $ \mathbb{E}_\infty[\mathcal{S}_A^r] = \gamma(1 + o(1)) $ when $ A = \gamma \zeta $, confirming that the procedure satisfies the false alarm constraint asymptotically.
• For the SR–r procedure, $ \text{ADD}_0(\mathcal{S}_A^r) = (1/I)[\log A + \varkappa - C(r)] + o(1) $, and setting $ C(r^*) = C_{\infty} $ ensures that $ \text{ADD}_0 \approx \text{SADD} $, achieving equalizer-like behavior.