Adaptive Quantile-Based CUSUM Methods

• The paper introduces an adaptive quantile-based nonparametric CUSUM that leverages robust U-quantile kernels and data-driven weight selection for effective change-point detection.
• It employs adaptive quantile estimation and dynamic weighting to enhance detection sensitivity across location, scale, and distributional shifts in nonlinear or heavy-tailed data streams.
• Empirical studies confirm its low false-alarm rates and computational efficiency, making it a practical tool for real-time monitoring in challenging data environments.

Adaptive quantile-based nonparametric CUSUM methods comprise a class of change-point detection procedures that exploit the distributional robustness of quantile-based statistics and the adaptability of data-driven weighting, all within a nonparametric framework. These approaches do not require parametric modeling assumptions and are tailored to efficiently detect location, scale, or general distributional changes in possibly dependent, heavy-tailed, or nonlinear data streams. Adaptivity is introduced through online or data-driven selection of test parameters (such as quantile level or CUSUM weights), often with quantile estimation or critical value computation performed adaptively for precision and computational efficiency.

1. Quantile-Based Nonparametric CUSUM: Kernels and U-Quantiles

The foundational construction is the quantile-based CUSUM, where the target statistic is a robust estimator such as a U-quantile. A U-quantile kernel $$h:\mathbb{R}\times\mathbb{R}\times\mathbb{R}\to\mathbb{R}$$ is specified, symmetric in its first two arguments and nondecreasing in the third. Given data $X_1,\dots,X_n$, the empirical U-distribution function is

$$U_n(t) = \frac{2}{n(n-1)} \sum_{1 \leq i < j \leq n} h(X_i, X_j, t).$$

The (population) $p$-quantile is defined as $q = U^{-1}(p)$ and the empirical as $\hat{q}_n = U_n^{-1}(p)$. The sequential process of empirical quantiles tracks their evolution over sub-samples,

$$\hat{q}_{\lfloor nt\rfloor} = U_{\lfloor nt\rfloor}^{-1}(p), \quad t \in [0,1],$$

which is centered and scaled as

$$\mathbb{U}_n(t) = \sqrt{n}(\hat{q}_{\lfloor nt\rfloor} - q).$$

Under minimal dependence assumptions (near-epoch dependence on mixing sequences and no finite-moment requirements), a functional central limit theorem holds: $$\left(\mathbb{U}_n(t)\right)_{t \in [0,1]} \Rightarrow \sigma W(t),$$ with $W$ a standard Brownian motion and

$X_1,\dots,X_n$0

Variance components are estimated via HAC-type autocovariance estimators, and kernel/bandwidth selection is supported with practical guidelines (Vogel et al., 2015).

2. Adaptive Quantile Selection and Data-Driven Weights

Change-point sensitivity in quantile-based CUSUMs depends on the choice of quantile $X_1,\dots,X_n$1 or the weight exponent $X_1,\dots,X_n$2 in weighted CUSUM statistics. If the location of the change is unknown, an adaptive scheme can maximize test power:

• Begin monitoring with an initial quantile level, such as $X_1,\dots,X_n$3 (median).
• Estimate, after a pilot sample $X_1,\dots,X_n$4, the signal-to-noise ratio for a range of $X_1,\dots,X_n$5: $X_1,\dots,X_n$6
• Select $X_1,\dots,X_n$7 maximizing this statistic, then use $X_1,\dots,X_n$8 for the remainder.

Alternatively, adaptive data-driven weights based on a preliminary estimate of the potential change-point can tune sensitivity to boundary or central changes. For the $X_1,\dots,X_n$9-weighted CUSUM,

$$U_n(t) = \frac{2}{n(n-1)} \sum_{1 \leq i < j \leq n} h(X_i, X_j, t).$$0

one estimates $$U_n(t) = \frac{2}{n(n-1)} \sum_{1 \leq i < j \leq n} h(X_i, X_j, t).$$1 by:

1. Using a maximally boundary-sensitive $$U_n(t) = \frac{2}{n(n-1)} \sum_{1 \leq i < j \leq n} h(X_i, X_j, t).$$2 to produce an initial estimate $$U_n(t) = \frac{2}{n(n-1)} \sum_{1 \leq i < j \leq n} h(X_i, X_j, t).$$3 of the change location.
2. Mapping $$U_n(t) = \frac{2}{n(n-1)} \sum_{1 \leq i < j \leq n} h(X_i, X_j, t).$$4 via a smooth aggregator $$U_n(t) = \frac{2}{n(n-1)} \sum_{1 \leq i < j \leq n} h(X_i, X_j, t).$$5, e.g., $$U_n(t) = \frac{2}{n(n-1)} \sum_{1 \leq i < j \leq n} h(X_i, X_j, t).$$6, to produce $$U_n(t) = \frac{2}{n(n-1)} \sum_{1 \leq i < j \leq n} h(X_i, X_j, t).$$7.
3. Using this $$U_n(t) = \frac{2}{n(n-1)} \sum_{1 \leq i < j \leq n} h(X_i, X_j, t).$$8 as the exponent for $$U_n(t) = \frac{2}{n(n-1)} \sum_{1 \leq i < j \leq n} h(X_i, X_j, t).$$9 in a final CUSUM, yielding a statistic adaptively sensitive to the (unknown) true change location.

Consistency and asymptotic validity of this adaptive method are established: under $p$0 the adaptive weight converges to unity, so the null distribution is unchanged; under $p$1 the plug-in weight converges to the optimal fixed-$p$2 case for the true change location (Schwaar, 2020).

3. Algorithmic Construction and Implementation

A practical algorithm for adaptive quantile-based nonparametric CUSUM includes:

• Selection of a robust U-quantile kernel (e.g., indicator of pairwise means/differences).
• For each time step or sub-sample, update empirical quantiles (e.g., by fast $p$3 algorithms).
• Use kernel density estimates and HAC smoothing for variance estimation, with theoretically motivated choices (Epanechnikov/bandwidth $p$4, Bartlett/Parzen kernels for autocovariance).
• After a pilot interval, update quantile level or CUSUM weighting exponent adaptively.
• Screen the maximal standardized deviation process,

$p$5

and reject stationarity if $p$6 exceeds critical value (e.g., $p$7 for level $p$8; critical values from the Kolmogorov distribution or as quantified by weighted Brownian bridge quantiles).

In sequential, real-time monitoring schemes, quantile partitioning and likelihood calculations can be updated online. Self-starting schemes provide unbiased cell probabilities for quantile bins under the null, maintaining exact false-alarm rates without large phase-I samples (Li, 2017).

4. Asymptotics, Critical Values, and Quantile Computation

Limiting distributions of these CUSUM-type statistics (both weighted and unweighted) are typically functionals of Brownian bridges,

$p$9

or for weighted statistics,

$q = U^{-1}(p)$0

with corresponding quantiles required for critical value calibration. Adaptive quantile computation for these suprema employs strong-approximation algorithms (adaptive time discretization with priority queues based on score functions) to efficiently and accurately estimate quantiles of weighted Brownian bridges. Algorithms achieve error $q = U^{-1}(p)$1 for all $q = U^{-1}(p)$2, providing high-precision critical values for complex weighting functions in CUSUMs (Franke et al., 2020).

This adaptive Monte Carlo framework enables practical application of theoretically justified critical values even in settings with nonstandard weights or extreme significance levels (e.g., $q = U^{-1}(p)$3). Computational rates are $q = U^{-1}(p)$4 for the adaptive method compared to $q = U^{-1}(p)$5 for naive uniform grids.

5. Extensions: Scale/Shape Sensitivity and Nonlinear/Heavy-Tailed Models

Modern adaptive quantile CUSUMs incorporate mechanisms for detecting a wide range of distributional changes. Location, scale, and general distributional changes are handled by constructing CUSUMs based on:

• Left-to-right quantile partitioning for location shifts;
• Center-outward quantile groupings for scale changes;
• Multinomial or likelihood-based weighting following observed change-type.

Nonparametric control charts can combine component CUSUMs with adaptive post-alarm diagnostics to differentiate between location and scale shifts. Posterior means via Dirichlet priors are incorporated into likelihood calculations, and self-starting quantile estimators ensure proper in-control calibration without tuning parameter dependence (Li, 2017).

In nonlinear quantile models, the quantile-CUSUM is based on subgradients of the quantile loss and appropriately normalized via a boundary-exponent parameter. The limiting distribution involves multidimensional Brownian paths and boundary-corrected scaling, and simulation is used to calibrate critical values for real-time monitoring (Ciuperca, 2016).

6. Robustness, Performance, and Empirical Evidence

Adaptive quantile-based nonparametric CUSUM methods are robust to heavy tails, outliers, and mild temporal dependence. No moment assumptions on inputs are required. Empirical studies confirm (for instance, using the Hodges-Lehmann estimator):

• Good robustness and efficiency, particularly for changes in central location.
• Superiority over classical LS-based CUSUM under heavy-tailed or nonlinear error distributions.
• Nominal in-control false-alarm rates and high sensitivity for a broad range of change scenarios, including small-scale and boundary point changes.
• Minimal computational cost in online applications due to fast updating algorithms (Vogel et al., 2015, Li, 2017).

Performance tables and simulation studies routinely find adaptive quantile-based CUSUMs at or near the top compared to distribution-free and rank-based competitors, with detection delay and ARL (average run length) metrics supporting their practical deployment (Li, 2017, Schwaar, 2020).

7. Limitations and Open Research Problems

Current limitations include incomplete theoretical strong-approximation bounds for certain weighted Brownian bridge functionals relevant in quantile-based adaptive CUSUMs, particularly for general Gaussian processes. Heuristic strategies for bias and Monte Carlo variance control in adaptive quantile algorithms are empirically effective but lack fully rigorously end-to-end guarantees.

Extension to high-dimensional or autocorrelated time series, further study of shape/adaptive partitioning strategies, and the integration of additional diagnostics for more complex change types remain areas for further research (Franke et al., 2020).

---

Key research underpinning adaptive quantile-based nonparametric CUSUM includes (Vogel et al., 2015, Franke et al., 2020, Li, 2017, Schwaar, 2020), and (Ciuperca, 2016). These works provide detailed theoretical justification, practical algorithms, and empirical studies supporting the effectiveness and flexibility of adaptive quantile CUSUM methodologies.