---
title: Window-Limited GLR-CuSum Detection
url: https://www.emergentmind.com/topics/window-limited-generalized-likelihood-ratio-glr-cusum
type: topic
---

# Window-Limited GLR-CuSum Detection

The window-limited generalized likelihood-ratio cumulative sum (GLR-CuSum) procedure is a family of sequential change detection methods designed for quickest change detection (QCD) when the post-change distribution is unknown and inferred online using only a finite sliding window of recent observations. These procedures were introduced to reconcile statistical optimality, computational tractability, fast adaptation, and robustness in both parametric and nonparametric settings. Window-limited GLR-CuSum achieves first-order asymptotic optimality under broad model assumptions, and its recent generalizations extend to nonparametric, nuisance, and nonstationary post-change regimes [2311.15128][1902.03460][2206.06777][2110.01581][1912.11693]. 

## 1. Quickest Change Detection and the Need for Window Limitation

In the classical QCD setup, observations $X_1,X_2,...$ are independent; for an unknown change-point $\nu$, the pre-change regime ($n<\nu$) has distribution $f_0$ (known), and the post-change regime ($n\geq\nu$) has distribution $f_1$ (typically unknown or partially specified). The objective is to minimize the worst-case detection delay
\[
\WADD(T)=\sup_{\nu\ge1}\esssup\,\E_\nu\bigl[(T-\nu+1)^+\mid\F_{\nu-1}\bigr]
\]
subject to a false-alarm constraint
\[
\E_\infty[T]\ge\gamma
\]
for a (large) target $\gamma$. While CuSum and likelihood-ratio-based schemes achieve minimax optimality when $f_1$ is known, in practice $f_1$ is often unknown and must be estimated. Standard GLR-CuSum statistics scan back over the entire observed history at each time, causing both computational and statistical delays (e.g., large burn-in). Window-limited GLR-CuSum restricts estimation and maximization to the most recent $m$ samples, maintaining optimality while drastically improving efficiency and responsiveness [2311.15128][1902.03460][2110.01581][2206.06777].

## 2. Window-Limited GLR-CuSum: Definitions and Algorithm

The generic window-limited GLR-CuSum statistic at time $n$ takes the form:
\[
W(n) = \max_{k\in \mathcal{K}_n} \sup_{\theta\in\Theta} \sum_{i=k}^n \log\frac{f_{1,i,k}^\theta(X_i)}{f_0(X_i)}
\]
where $\mathcal{K}_n = \{\max(1,n-m+1),...,n\}$ is the set of candidate change-points inside an $m$-length sliding window, and $f_{1,i,k}^\theta$ denotes the post-change density parameterized by $\theta$ (possibly nonstationary, i.e., dependent on time since the putative change at $k$). For a fixed or suitably growing window size $m=m_b$, the stopping rule is
\[
T_b = \inf\{n: W(n)\geq b\}
\]
where threshold $b$ is chosen so that $\E_\infty[T_b]\ge\gamma$ for the desired average run length (ARL) constraint [2110.01581][2206.06777].

Nonparametric variants (NGLR-CuSum) replace parametric likelihoods with sliding-window density estimators (e.g., kernel density estimates computed on samples $\{X_k,\dots,X_n\}\setminus\{X_i\}$ for leave-one-out correction), and use 
\[
\widehat{Z}_i^{n,k} = \log\frac{\widehat{f}^{n,k}_{-i}(X_i)}{f_0(X_i)}
\]
with the statistic
\[
W(n) = \max_{(n-m_b)^+ < k < n} \sum_{i=k}^n \widehat{Z}_i^{n,k}
\]
and otherwise identical stopping rule [2311.15128].

## 3. Statistical Properties and Optimality

Window-limited GLR-CuSum schemes are proven to be first-order asymptotically optimal in Lorden's sense:
\[
\inf_{T:\E_\infty[T]\ge\gamma} \WADD(T) \sim \frac{\log\gamma}{I}
\]
where $I$ is a generalized Kullback–Leibler information number dependent on the post-change regime. For parametric and certain nonstationary models, the cumulative or instantaneous mean KL-divergence governs the detection-delay scaling [2110.01581][2206.06777]. For nonparametric density estimation, rates depend on estimator convergence:
\[
\E\left[ D(f_1\Vert\widehat{f}^w) \right] = O(w^{-\beta_1}), \quad \E\left[ (\log \frac{f_1(X)}{\widehat{f}^w(X)})^2 \right] = O(w^{-\beta_2})
\]
ensuring optimal WADD as $\alpha\to0$ if the window $m_b$ grows at least linearly in $b$ (threshold) and $\log m_b = o(b)$ [2311.15128].

False-alarm rates are controlled via renewal-theoretic calculations, motivating thresholds of the form $b\approx\log\gamma$ with higher-order corrections for nonparametric estimators (to account for "max-product" bounds) [2311.15128].

Under suitable conditions, window-limited GLR-CuSum procedures match the minimax lower bounds for detection delay while guaranteeing ARL constraints—robust to misspecification and model complexity [2110.01581][2206.06777][2311.15128].

## 4. Window Size Selection and Practical Considerations

Asymptotic results require the window size $m$ to satisfy:
\[
m_\alpha/\;g_\theta^{-1}(|\log\alpha|)\to c>1
\]
where $g_\theta^{-1}$ is the inverse growth of the cumulative post-change KL divergence, capturing the typical delay scale. In i.i.d. or stationary settings, $m\gtrsim (1+\epsilon)b/I$ suffices; nonparametric tests employ $m\approx b^{1.1}$ or $(\log\alpha^{-1})^\kappa$ for $\kappa\in(1,2)$ [2311.15128][2110.01581]. Window sizes that grow too slowly result in suboptimal delay, while excessive growth incurs unnecessary computation. For practical tuning, one sets $b$ close to $\log(\alpha^{-1})$ and verifies estimator-specific conditions (e.g., the "max-product" moment, leave-one-out cross-validation) [2311.15128].

Efficient recursive implementations exploit monotone-queue data structures and partial sum representations to keep per-sample cost $O(1)$ for classical and $O(mJ)$ for composite parameter or nonparametric settings (where $J$ is the number of grid points for parameter discretization) [2110.01581][2206.06777][2311.15128].

## 5. Model Variants and Extensions

Window-limited GLR-CuSum accommodates a broad array of QCD problems, including:

- **Nonparametric uncertainty:** NGLR-CuSum achieves minimax rates without any prior knowledge of $f_1$, driven solely by the smoothness (e.g., Hölder class) and convergence properties of the density estimator [2311.15128].
- **Nonstationary post-change:** Allows $f_{1,i,k}^\theta$ to vary with time since change. Asymptotic delay is governed by a function $g_{k,\theta}(m)$ tracking cumulative expected post-change divergence; minimax-optimality holds under regularity and concentration-of-measure assumptions [2110.01581].
- **Nuisance changes:** The window-limited GLR form extends to discriminating critical changes from nuisance shifts by maximizing test statistics over all candidates for the nuisance-change point, retaining optimality provided identifiability via KL-separation [1902.03460].
- **Sampling constraints and switching costs:** The methodology extends to settings with restricted observation policies, capturing both observation sparsity and switching penalizations via randomized finite-window patterns and Markovian policy synthesis [1912.11693].

## 6. Numerical Performance and Empirical Validation

Extensive numerical experiments validate theory across several canonical and applied settings:

- **Gaussian mean shift (unknown mean):** NGLR-CuSum and parametric window-limited GLR-CuSum achieve detection delays indistinguishable from the oracle CuSum with known mean, as long as window sizes are chosen per the theoretical guidelines. For moderate $m$, ARL and delay scale as predicted by renewal theory and minimax optimality [2311.15128][2206.06777].
- **Exponential and decaying mean models:** In scenarios with nonstationary or composite post-change parameters, window-limited GLR-CuSum tracks optimal known-parameter detection performance for large enough $m$; the influence of window selection is especially pronounced for heavy-tailed or fast-varying post-change models [2110.01581].
- **Nuisance regimes:** In the presence of nuisance changes, window-limited GLR-CuSum outperforms moving-average and two-stage procedures in both theoretical and empirical comparisons [1902.03460].
- **Applied contexts (e.g., pandemic monitoring):** Real-world data such as COVID-19 incidence timeseries illustrate the practical effectiveness of window-limited GLR-CuSum in rapidly and reliably identifying change points corresponding to major trend shifts [2110.01581].

## 7. Relation to Classical and Contemporary QCD Schemes

The window-limited GLR-CuSum generalizes several classical QCD detectors:

| Method                    | Post-change knowledge    | Statistic           | Windowed? | Asymptotic optimality |
|---------------------------|-------------------------|---------------------|-----------|----------------------|
| CuSum                     | $f_1$ known             | LLR sum             | No        | Yes                  |
| Parametric GLR-CuSum      | $f_1$ in known family   | Sup over $\theta$   | Optional  | Yes, windowed or not |
| Window-limited GLR-CuSum  | $f_1$ in (rich) family  | Windowed sup        | Yes       | Yes                  |
| NGLR-CuSum                | $f_1$ unknown, smooth   | Nonparametric LR    | Yes       | Yes with estimator   |

These approaches constitute a unified toolkit for detection under uncertainty, offering trade-offs between computational load, adaptation lag, and distributional assumptions. For many realistic applications, windowing is both necessary and sufficient for minimax-optimal QCD [2311.15128][1902.03460][2206.06777][2110.01581][1912.11693].

Source: https://www.emergentmind.com/topics/window-limited-generalized-likelihood-ratio-glr-cusum