Nonparametric Rank-Based CUSUM

• Nonparametric rank-based CUSUM is a sequential change detection method leveraging rank statistics to identify shifts without relying on specific distribution assumptions.
• It replaces traditional log-likelihood increments with robust rank or quantile-based scores, ensuring adaptability in heavy-tailed, high-dimensional, or unknown distribution settings.
• Applications include univariate, multivariate, time series, and anomaly detection, supported by adaptive algorithms such as the Mann-Whitney and GEM CUSUM approaches.

Nonparametric rank-based CUSUM procedures constitute a class of sequential change detection and monitoring methods that leverage rank statistics or nonparametric test statistics in place of the usual parametric log-likelihood ratios within the Cumulative Sum (CUSUM) framework. These approaches are motivated by the need for distribution-free, robust, and adaptive change detection procedures, especially under unknown or heavy-tailed distributions, high-dimensional regimes, or weak distributional assumptions. Both classical and modern research develops such methods for univariate and multivariate settings, process control, time series, panel/longitudinal data, and high-dimensional anomaly detection. Notable contributions include algorithms based on the Mann-Whitney statistic, signed sequential ranks, adaptive quantile binning, self-normalized rank statistics, and outlier detection via nonparametric entropy measures.

1. Classical and Modern Foundations of Rank-Based CUSUM

The CUSUM algorithm, originated by Page, is fundamentally optimal in quickest change-point detection under known parametric models, accumulating log-likelihood ratios. Rank-based and nonparametric extensions replace parametric elementary statistics with appropriately constructed rank-based, quantile-based, or permutation-invariant statistics. For example, Wang and Xiong (Wang et al., 2013) construct a CUSUM based on the classical Mann-Whitney statistic, showing that the resulting monitoring scheme is exactly distribution-free under any continuous in-control distribution. Lombard and van Zyl (Lombard et al., 2017) develop the signed sequential rank CUSUM, which is fully self-starting and exploits the independence property of signed ranks under symmetry.

Recent years have witnessed a proliferation of adaptive schemes, such as the nonparametric adaptive CUSUM of Li (Li, 2017), which achieves distribution-freeness and adapts to arbitrary distributional changes using online quantile-based binning and multinomial models, and the time series approaches using self-normalized rank statistics for long-range dependent sequences (Betken et al., 2020). For high-dimensional and structured data, the Geometric Entropy Minimization (GEM) CUSUM (Yilmaz, 2017) generalizes the CUSUM principle to distance-based outlier scores.

2. Construction of Rank-Based and Nonparametric CUSUM Procedures

The generic structure of a nonparametric, rank-based CUSUM replaces parametric increments by rank-based scores. Several canonical choices include:

• Sequential Mann-Whitney CUSUM: At each time, compute the Mann-Whitney statistic among current and reference observations, standardize, and update a one-sided or two-sided CUSUM recursion:

$$S_j = \max\{0, S_{j-1} + \mathrm{SMW}_j - k\}, \quad S_{\max}(m,n) = \max_j S_j$$

where $\mathrm{SMW}_j$ is the standardized Mann-Whitney statistic (Wang et al., 2013).

• Signed Sequential Rank CUSUM: At time $n$, evaluate the signed sequential rank $R_n = \mathrm{sign}(X_n) r_n^+$, transform via an odd score function (Wilcoxon or Van der Waerden), standardize, and compute Page’s recursion:

$S_n^+ = \max\{0, S_{n-1}^+ + \xi_n - k\}$

with reference value $k$ and control limit $h$ (Lombard et al., 2017).

• Adaptive Multinomial Quantile CUSUM: Partition the sample space at each $t$ via estimated quantiles, record cell frequencies (indicators $Y_{t,j}$), model as categorical, and accumulate online log-likelihood ratio or adaptive score increments in CUSUM fashion, exploiting the probability-integral transform to ensure nominal uniform behavior (Li, 2017).
• Rank-Score CUSUM for Time Series: Use a score function $a(i)$ applied to data ranks or empirical cdf values, and compute centered partial sums:

$\mathrm{SMW}_j$0

with self-normalization or subsampling to calibrate significance under serial dependence (Betken et al., 2020).

• GEM Outlier Score CUSUM: For high-dimensional nominal training data partitioned into two sets, construct for each new $\mathrm{SMW}_j$1 an outlier score via k-nearest neighbor distances. Accumulate $\mathrm{SMW}_j$2 in a CUSUM recursion:

$\mathrm{SMW}_j$3

using a pre-estimated drift value $\mathrm{SMW}_j$4, yielding a fully nonparametric, high-dimensional outlier CUSUM (Yilmaz, 2017).

3. Statistical Properties and Theoretical Guarantees

All forms of nonparametric, rank-based CUSUM inherit null distribution-freeness (or symmetry/finiteness of moments) under broad conditions:

• Distribution-Free or Asymptotically Pivotal Null Laws: The run-length, ARL, and size (type I error) properties are constant across all continuous in-control distributions, e.g., the Mann-Whitney and signed-rank statistics are permutation-invariant under the null (Wang et al., 2013, Lombard et al., 2017, Li, 2017). For self-normalized time series CUSUMs, the critical values can be calibrated via block resampling or subsampling, providing asymptotic validity (Betken et al., 2020).
• Local Power and Efficiency: For Gaussian margins, the asymptotic relative efficiency of rank-based CUSUMs with optimal scores (Van der Waerden) matches that of the optimal parametric CUSUM (ARE=1), while the Wilcoxon variant is recommended for heavy-tailed margins (Betken et al., 2020, Lombard et al., 2017).
• Detection Delay: The nonparametric online GEM-based CUSUM achieves near-optimal average detection delay under Lorden’s minimax criterion; as $\mathrm{SMW}_j$5, worst-case delay matches the parametric optimum (Yilmaz, 2017).
• Under Alternatives: All CUSUMs are designed to accumulate positive excursions of rank-based evidence, requiring “persistent” outlier evidence or systematic rank deviations to signal a change. This reduces susceptibility to isolated outliers.

4. Implementation and Computational Considerations

Efficient implementation is enabled by score-based and rank-only updating, obviating the need for density estimation, parametric fitting, or smoothing:

• Updates: Rank-based statistics require sorting or data structures to maintain order information. With standard sorting libraries, O($\mathrm{SMW}_j$6) or O($\mathrm{SMW}_j$7) per update is typical for univariate data (Wang et al., 2013, Lombard et al., 2017). For GEM-based CUSUM, kd-trees or efficient nearest neighbor searches enable tractable computation even for high-dimensional data, provided $\mathrm{SMW}_j$8 (the number of neighbors) remains modest (Yilmaz, 2017).
• Phase I Reference Data: Many schemes use a fixed set of historical in-control data (Mann-Whitney, panel-data, GEM), while “self-starting” variants, such as the signed sequential rank CUSUM and the adaptive quantile CUSUM, require no reference phase and can be initialized online (Lombard et al., 2017, Li, 2017).
• Control Limit Selection: For distribution-free procedures, ARL$\mathrm{SMW}_j$9 curves can be simulated under a benchmark (e.g., $n$0 or $n$1), and control limits interpolated. For time series, block subsampling or bootstrap is recommended for significance calibration (Betken et al., 2020, Pommeret et al., 2011).
• Handling of Ties: While theoretical guarantees assume continuous distributions, ties are handled by midrank or breaking at random as a pragmatic (and nearly invariant) solution (Lombard et al., 2017).
• Software and Pseudocode: Detailed pseudocode is available for all primary algorithms, including the GEM CUSUM, rank-score time series CUSUM, and both Mann-Whitney and signed-rank CUSUMs (Yilmaz, 2017, Wang et al., 2013, Lombard et al., 2017).

5. Empirical Performance and Comparative Analysis

Extensive empirical studies confirm the competitive or superior performance of nonparametric, rank-based CUSUMs in various settings:

• Robustness: All referenced methods are robust against misspecification of underlying distribution, heavy tails, or outlier contamination. For example, signed-rank CUSUMs are not affected by single large outliers and can correctly reflect true shifts present in industrial data (Lombard et al., 2017).
• Sensitivity to Small Shifts: The nonparametric rank-based CUSUM outperforms several alternative nonparametric control charts (e.g., Bakir–Reynolds, within-group signed-rank, sequential exceedance count) when the shift is small and the distribution is unknown (Wang et al., 2013). For large shifts, parametric CUSUMs may have a slight advantage if the model is correct.
• Delay/FAR Tradeoffs: GEM-CUSUM (ODIT) achieves performance close to the clairvoyant parametric CUSUM with known alternative, and in high-dimensional regimes and for simulated/real data, offers speed close to the parametric limit (Yilmaz, 2017). Adaptive quantile-based CUSUMs show uniform or better performance across diffuse scenarios and provide in-built diagnostics of shift type (Li, 2017).
• Panel Data and Long-Range Dependence: Panels with monotone link changes and long-memory sequences can be efficiently monitored using the empirical CUSUM/quantile approaches, which extend to multivariate/properly weighted settings (Pommeret et al., 2011, Betken et al., 2020).

6. Extensions, Limitations, and Application Domains

Rank-based CUSUMs have been extended or adapted for various generalizations:

• Panel and Multivariate Data: Extensions include stacked ranks for multi-sample tests, componentwise or depth-based ordering for multivariate data, and pairwise resampling for panel models. However, theoretical analysis becomes more intricate beyond $n$2 or homogenous sample sizes (Pommeret et al., 2011).
• High-Dimensional and Anomaly Detection: Procedures like GEM-CUSUM are tailored for high-dimensional input; the primary bottleneck is efficient nearest neighbor search, mitigated by kd-tree algorithms (Yilmaz, 2017).
• Distributional Change Types: Several procedures offer diagnostics for the type of distributional change—location/scale up or down—based on which CUSUM substatistic first crosses the threshold (Li, 2017).
• Serial Dependence: Bootstrap or block resampling is necessary when strong temporal dependence is present to maintain valid type I error rates (Betken et al., 2020, Pommeret et al., 2011).
• Limitations: While these tests are powerful for monotonic, persistent, or step changes, more subtle regimes (e.g., transient or oscillatory deviations) may not be as swiftly detected. The performance can also degrade if tuning parameters (window length for subsampling, $n$3 in GEM) are poorly chosen.

7. Summary Table of Key Nonparametric Rank-Based CUSUMs

Method	Statistic / Score Used	Optimal For / Key Features
Mann-Whitney CUSUM	Mann-Whitney two-sample rank	Small shift, location changes, d.f.-free (Wang et al., 2013)
Signed Seq. Rank CUSUM	Signed sequential rank score	Location/scale shifts, self-starting, symmetric distributions (Lombard et al., 2017)
Adaptive Quantile CUSUM	Quantile-binned multinomial	Arbitrary changes, fully adaptive, diagnostics (Li, 2017)
Rank-Score CUSUM (time series)	General rank-score function	Long-range dependence, self-normalization (Betken et al., 2020)
GEM-based CUSUM	k-NN-based outlier score	High-dimensional anomaly detection, near-parametric delay (Yilmaz, 2017)

Each method offers a distinct combination of adaptability, robustness, and computational tractability, making these tools central to modern sequential analysis and change detection under minimal distributional assumptions.