Disturbance Score

• Disturbance Score is a family of quantitative measures that capture deviation, severity, margin, or back-action by comparing observed data to a baseline.
• It employs various formulations such as standardized deviation, confidence-weighted severity, and discrepancy scores across applications from SAR mapping to quantum measurements.
• Its design pattern couples a nominal model, a discrepancy functional, and an interpretation layer to enable calibrated, domain-specific diagnostics.

Searching arXiv for the cited disturbance-score literature and closely related papers. Disturbance Score denotes a family of quantitative constructs used to measure deviation from nominal behavior, disturbance severity, disturbance-induced risk, or measurement back-action. The term is not standardized across disciplines. In the cited literature it ranges from recurrence-based markers of storm-time magnetospheric organization, to confidence-weighted event severity in power systems, to per-pixel deviation from a learned synthetic-aperture-radar baseline, to quantum root-mean-square disturbance of an observable, to resilience margins under repeated ecological shocks, to merger-driven dynamical disturbance in galaxy clusters, and to coupling-based discrepancy between probability distributions. This suggests that “Disturbance Score” is best understood as a cross-domain label for formally distinct but structurally related diagnostics rather than as a single universal metric (Donner et al., 2018, Lee et al., 2018, Harel et al., 2012, Hardiman-Mostow et al., 15 Jan 2025, Ozawa, 2021, Meyer et al., 2018, Kong et al., 25 Sep 2025).

1. Scope and conceptual taxonomy

Across the literature, a Disturbance Score serves one of four main roles. First, it can be a deviation score, quantifying how far an observed state lies from a learned or assumed baseline. Second, it can be a severity score, combining estimated disturbance magnitude with confidence or duration. Third, it can be a margin score, expressing distance to a stability, resilience, or threshold boundary. Fourth, it can be a back-action score, quantifying how much one operation alters a later observable or state. These roles recur even when the underlying mathematics differs substantially.

A useful summary is given below.

Domain	Formal score form	Primary interpretation
Geomagnetosphere	standardized recurrence metrics from Dst and drivers	storm-time dynamical organization
Interconnected power systems	$DS = |\hat m| \cdot \max_b P(y=b\mid x)$	severity weighted by localization confidence
Distribution comparison	$\operatorname{PV}_\varepsilon(P,Q)$	minimal mass unmatched within tolerance $\varepsilon$
SAR disturbance mapping	$d(\mathfrak x)=\max_p |x_{T+1}^p-\mu_p|/\sigma_p$	per-pixel anomaly against baseline distribution
Quantum measurement	$\eta_O(B)$ or $\eta_O(B)^2$	root-mean-square observable disturbance
Flow-kick ecology	$S=|K|/K_c(\tau)$ or $S=\tau_c(|K|)/\tau$	proximity to resilience boundary
Galaxy clusters	$S_{\mathrm{dist}} \approx \hat S(\text{features})$	merger-driven dynamical disturbance

The term also appears in statistical inference in a different sense: in linear mixed-effects state-space models, the “disturbance score” is the score vector of the observed-data log-likelihood written in terms of smoothed disturbances rather than a scalar severity index. This usage emphasizes estimation geometry rather than physical disruption (Zhou et al., 2014).

2. Recurrent mathematical forms

One common form is the standardized deviation score. In OPERA Sentinel-1 RTC-S1 disturbance mapping, the learned baseline for each pixel and polarization is Gaussian in the logit-transformed $\gamma^0$ domain, and the disturbance score is the one-dimensional Mahalanobis distance

$\operatorname{PV}_\varepsilon(P,Q)$0

followed by binarization $\operatorname{PV}_\varepsilon(P,Q)$1. The paper reports that a global $\operatorname{PV}_\varepsilon(P,Q)$2 standard deviations gave $\operatorname{PV}_\varepsilon(P,Q)$3 consistently across the three study regions (Hardiman-Mostow et al., 15 Jan 2025).

A second form is the confidence-weighted severity score. In model-free disturbance localization and magnitude estimation for interconnected power systems, logistic regression yields bus probabilities and linear regression yields disturbance magnitude. The proposed combined score is

$\operatorname{PV}_\varepsilon(P,Q)$4

with the variant $\operatorname{PV}_\varepsilon(P,Q)$5 if a “no disturbance” class is included. A bus-resolved map,

$\operatorname{PV}_\varepsilon(P,Q)$6

supports top-$\operatorname{PV}_\varepsilon(P,Q)$7 inspection and response prioritization (Lee et al., 2018).

A third form is the tolerance-based discrepancy score between distributions. The perturbed variation at radius $\operatorname{PV}_\varepsilon(P,Q)$8 is

$\operatorname{PV}_\varepsilon(P,Q)$9

equivalently the minimal proportion of mass that cannot be matched within $\varepsilon$0. At $\varepsilon$1, it reduces to total variation, and for $\varepsilon$2 it relaxes exact equality into similarity up to admissible perturbation (Harel et al., 2012).

A fourth form is the boundary or safety-margin score. In flow-kick ecology, the resilience boundary $\varepsilon$3 in $\varepsilon$4-space defines critical disturbance magnitude and recovery time, leading to

$\varepsilon$5

In response-based frequency stability assessment, the step-disturbance margin is

$\varepsilon$6

where $\varepsilon$7 is the maximum tolerable disturbance power under steady-state and transient frequency-deviation constraints (Meyer et al., 2018, Chen et al., 26 Nov 2025).

A fifth form is the operator disturbance score in quantum measurement. The operator-based disturbance of observable $\varepsilon$8 caused by a measurement $\varepsilon$9 in state $d(\mathfrak x)=\max_p |x_{T+1}^p-\mu_p|/\sigma_p$0 is

$d(\mathfrak x)=\max_p |x_{T+1}^p-\mu_p|/\sigma_p$1

This score is state-dependent, satisfies the correspondence principle when a joint probability distribution exists, and can be operationally nonzero even when the system state and the single-time distribution of $d(\mathfrak x)=\max_p |x_{T+1}^p-\mu_p|/\sigma_p$2 remain unchanged (Ozawa, 2021).

A sixth form is the disturbance score as likelihood gradient. In linear mixed-effects state-space models, disturbance smoothing yields smoothed observation disturbances $d(\mathfrak x)=\max_p |x_{T+1}^p-\mu_p|/\sigma_p$3 and state disturbances $d(\mathfrak x)=\max_p |x_{T+1}^p-\mu_p|/\sigma_p$4, and the score vector $d(\mathfrak x)=\max_p |x_{T+1}^p-\mu_p|/\sigma_p$5 is expressed in terms of their smoothed second moments. Here “disturbance score” is a parameter-estimation object rather than a severity indicator (Zhou et al., 2014).

3. Earth, environmental, and astrophysical systems

In geomagnetospheric analysis, the disturbance storm time index $d(\mathfrak x)=\max_p |x_{T+1}^p-\mu_p|/\sigma_p$6 is “the average change of the horizontal component of the Earth’s magnetic field recorded at four mid-latitude magnetic observatories,” sampled hourly. Recurrence quantification analysis and recurrence network analysis applied to embedded $d(\mathfrak x)=\max_p |x_{T+1}^p-\mu_p|/\sigma_p$7, $d(\mathfrak x)=\max_p |x_{T+1}^p-\mu_p|/\sigma_p$8, $d(\mathfrak x)=\max_p |x_{T+1}^p-\mu_p|/\sigma_p$9, and $\eta_O(B)$0 show that storm periods have higher determinism, higher trapping time, and higher recurrence transitivity than quiescent periods. In the 2001 analysis, $\eta_O(B)$1 complexity measures reliably separated storm from non-storm intervals, with a commonly used storm threshold $\eta_O(B)$2. A proposed composite disturbance score therefore emphasized standardized $\eta_O(B)$3, $\eta_O(B)$4, and $\eta_O(B)$5 from $\eta_O(B)$6, with smaller modulating contributions from $\eta_O(B)$7 and $\eta_O(B)$8 (Donner et al., 2018, Donner et al., 2018).

In self-supervised SAR disturbance mapping, the disturbance score is learned from baseline image sequences rather than imposed from hand-crafted thresholds. A vision transformer predicts per-pixel $\eta_O(B)$9 and $\eta_O(B)^2$0 for VV and VH, and disturbance is declared when the observed post-event backscatter deviates strongly from that predictive distribution. The resulting score is local, probabilistic, and operationally scalable through overlapping-window inference on OPERA RTC-S1 burst tiles. The paper reports $\eta_O(B)^2$1 and $\eta_O(B)^2$2 for the 2024 Papua New Guinea landslide, $\eta_O(B)^2$3 and $\eta_O(B)^2$4 for the 2024 Chile wildfires, and $\eta_O(B)^2$5 and $\eta_O(B)^2$6 for the 2024 Bangladesh floods (Hardiman-Mostow et al., 15 Jan 2025).

In ecosystem resilience, disturbance is parameterized explicitly by magnitude and frequency. The flow-kick framework models continuous recovery under $\eta_O(B)^2$7 and discrete kicks $\eta_O(B)^2$8, with stroboscopic map $\eta_O(B)^2$9. The resilience boundary $S=|K|/K_c(\tau)$0 partitions disturbance regimes that stabilize within a basin from those that cause escape. The associated score is therefore not a measure of observed damage but of distance to resilience loss. The paper argues that distance-to-threshold resilience can overestimate resilience under repeated kicks because flow-kick thresholds can lie far inside the undisturbed basin (Meyer et al., 2018).

In forest structure, disturbance is inferred from the range over which a finite-size scaling law fits the tree-height distribution. With crown-shape exponent $S=|K|/K_c(\tau)$1 defined by $S=|K|/K_c(\tau)$2, the predicted interior height-density slope is $S=|K|/K_c(\tau)$3. The recovery index

$S=|K|/K_c(\tau)$4

measures the normalized width of the range $S=|K|/K_c(\tau)$5 successfully fitted by the model. Semi-natural stands had $S=|K|/K_c(\tau)$6 and $S=|K|/K_c(\tau)$7, whereas formerly managed stands had $S=|K|/K_c(\tau)$8 and $S=|K|/K_c(\tau)$9, so the fitted power-law range narrowed sharply with disturbance (Anfodillo et al., 2012).

In galaxy clusters, the disturbance score is supervised by a merger-history target rather than analytically closed. The physical target is

$S=\tau_c(|K|)/\tau$0

with $S=\tau_c(|K|)/\tau$1 for past- and future-merger scores and a symmetric $S=\tau_c(|K|)/\tau$2 window for the full score. XGBoost maps phase-space and morphology features such as $S=\tau_c(|K|)/\tau$3, GMM dispersions, $S=\tau_c(|K|)/\tau$4, $S=\tau_c(|K|)/\tau$5, and a mass-ratio proxy to a learned $S=\tau_c(|K|)/\tau$6. Phase-space features quantify merger-driven asymmetry, while blue galaxy fraction and X-ray/BCG offset are used as timing-sensitive adjuncts (Kong et al., 25 Sep 2025).

4. Power and infrastructure applications

In interconnected bulk power systems, disturbance scores are often explicitly operational. The localization-and-estimation framework based on synchronized generator frequency measurements assumes disturbance start time $S=\tau_c(|K|)/\tau$7 is known and uses $S=\tau_c(|K|)/\tau$8 samples. Features are built from smoothed deviations $S=\tau_c(|K|)/\tau$9, localization is performed with softmax logistic regression, and magnitude is estimated with bus-specific linear regression. The combined score $S_{\mathrm{dist}} \approx \hat S(\text{features})$0 is therefore directly interpretable in megawatts scaled by localization confidence (Lee et al., 2018).

A different power-quality tradition uses wavelet multiresolution analysis. For a signal $S_{\mathrm{dist}} \approx \hat S(\text{features})$1 sampled at $S_{\mathrm{dist}} \approx \hat S(\text{features})$2, the instantaneous transient disturbance index is

$S_{\mathrm{dist}} \approx \hat S(\text{features})$3

where $S_{\mathrm{dist}} \approx \hat S(\text{features})$4 and $S_{\mathrm{dist}} \approx \hat S(\text{features})$5 is the mean-square energy of the approximation band. The global disturbance ratio is

$S_{\mathrm{dist}} \approx \hat S(\text{features})$6

a duration-weighted scalar summary. With $S_{\mathrm{dist}} \approx \hat S(\text{features})$7, a discrete Meyer wavelet, and a $S_{\mathrm{dist}} \approx \hat S(\text{features})$8-cycle window, the method achieved a $S_{\mathrm{dist}} \approx \hat S(\text{features})$9 classification success rate across $\gamma^0$0 test signals and remained effective down to $\gamma^0$1 (Borrás et al., 2024).

Rule-based disturbance analytics for transmission-system digital fault recorders use a broader collection of event-specific thresholds on RMS quantities, derivatives, zero crossings, harmonic ratios, and phase angles. Fourteen event types are covered, including CT saturation, capacitor bank switching, ferroresonance, lightning, harmonic resonance, and incipient CVT failure. The paper does not define a single universal scalar severity score, but it does define explicit confidence logic for several events and reports approximately $\gamma^0$2 average accuracy on $\gamma^0$3 signal files. Continuous nominal data are summarized through cyclic histograms, with a reported memory reduction by a factor of $\gamma^0$4 (Boyd et al., 2023).

Frequency-stability assessment under high-renewable disturbances uses disturbance power inferred from generator electrical responses,

$\gamma^0$5

followed by classification into short-term, step, second-level slope, and minute-level slope events. For step disturbances, the explicit disturbance score is the safety-margin index

$\gamma^0$6

where $\gamma^0$7 is the minimum of steady-state and transient tolerable powers. For slope disturbances, the natural output is the over-limit time $\gamma^0$8 obtained from analytical frequency-response models. This gives a physically different but mathematically parallel margin concept: negative $\gamma^0$9 or short $\operatorname{PV}_\varepsilon(P,Q)$00 indicate high disturbance severity (Chen et al., 26 Nov 2025).

5. Quantum, statistical, and information-theoretic meanings

In quantum measurement theory, disturbance is defined at the observable level rather than at the level of state change alone. The operator-based disturbance measure

$\operatorname{PV}_\varepsilon(P,Q)$01

quantifies the change in $\operatorname{PV}_\varepsilon(P,Q)$02 under the induced channel $\operatorname{PV}_\varepsilon(P,Q)$03. A central result is that a system can incur operationally detectable disturbance without state change: in a qubit example, the system remains in $\operatorname{PV}_\varepsilon(P,Q)$04, the single-time distribution of $\operatorname{PV}_\varepsilon(P,Q)$05 is unchanged, yet $\operatorname{PV}_\varepsilon(P,Q)$06 because the disturbance appears in time-like correlations (Ozawa, 2021).

The disturbance-evaluation circuit turns this quantity into an experimentally accessible second-order decoherence coefficient. With a weak ancilla probe and weak interaction $\operatorname{PV}_\varepsilon(P,Q)$07, the decoherence metric $\operatorname{PV}_\varepsilon(P,Q)$08 satisfies

$\operatorname{PV}_\varepsilon(P,Q)$09

The paper compares this method to the three-state method and weak measurement method, both in simulation and on an IBM quantum computer, and reports that disturbance evaluation circuit accuracy is comparable to or better than the three-state method while remaining more compatible with error mitigation (Emori et al., 2024).

A related but logically distinct development is the disturbance-enhanced uncertainty relation. There, disturbance is any statistical distance $\operatorname{PV}_\varepsilon(P,Q)$10 between the outcome distribution of a measurement $\operatorname{PV}_\varepsilon(P,Q)$11 performed directly on $\operatorname{PV}_\varepsilon(P,Q)$12 and after a prior nonselective measurement $\operatorname{PV}_\varepsilon(P,Q)$13. Examples include total variation, Kullback–Leibler divergence, Rényi divergence, Tsallis relative entropy, and Euclidean distance. The paper proves lower bounds from disturbance to uncertainty, including

$\operatorname{PV}_\varepsilon(P,Q)$14

and analogous relations for Rényi and Tsallis entropies, thereby turning disturbance into an operational lower bound on preparation uncertainty and on relative-entropy coherence (Sun et al., 2022).

In quantum direction estimation with antiparallel spin-coherent pairs, the disturbance score is again different: disturbance is

$\operatorname{PV}_\varepsilon(P,Q)$15

where $\operatorname{PV}_\varepsilon(P,Q)$16 is the full-state operation fidelity of the postmeasurement state. The paper derives the optimal information-disturbance tradeoff $\operatorname{PV}_\varepsilon(P,Q)$17 through a covariant Choi-seed optimization. This use of “disturbance” belongs to the same back-action family as $\operatorname{PV}_\varepsilon(P,Q)$18, but it measures loss of global state fidelity rather than root-mean-square change of a designated observable (Sacchi, 16 Jun 2026).

The perturbed variation provides an information-theoretic notion of disturbance without physical dynamics. It measures how much probability mass must be displaced more than $\operatorname{PV}_\varepsilon(P,Q)$19 to align two distributions. In this sense, disturbance is a tolerated perturbation budget rather than a temporal event. The sample estimator is obtained from maximum bipartite matching on an $\operatorname{PV}_\varepsilon(P,Q)$20-threshold graph, and the paper gives finite-sample convergence bounds, bootstrap BCa confidence intervals, and explicit high-dimensional lower bounds that exhibit the curse of dimensionality (Harel et al., 2012).

6. Calibration, interpretation, and limitations

Across applications, a Disturbance Score is meaningful only relative to a baseline, threshold, or reference model. In geomagnetic recurrence analysis, fixed recurrence rate $\operatorname{PV}_\varepsilon(P,Q)$21, embedding delays, and window placement are chosen to ensure comparability across windows, but nonstationarity and variance changes remain consequential. In SAR, $\operatorname{PV}_\varepsilon(P,Q)$22 is selected by precision–recall analysis, and the learned $\operatorname{PV}_\varepsilon(P,Q)$23 absorbs some but not all variability due to phenology, soil moisture, or residual geometry. In power-system localization, confidence is only as reliable as softmax calibration, and the paper notes that probability calibration was not reported (Donner et al., 2018, Hardiman-Mostow et al., 15 Jan 2025, Lee et al., 2018).

Several disturbance scores are explicitly regime dependent. The flow-kick score depends on a specified basin of attraction and on the resilience boundary $\operatorname{PV}_\varepsilon(P,Q)$24, so it is a model-based quantity rather than a model-free severity indicator. The forest recovery index depends on the finite-size scaling fit range and on the assumption that the stand is close enough to saturation for the power-law regime to be meaningful. In galaxy clusters, phase-space features diagnose merger-driven asymmetry but “are not sensitive to whether the secondary progenitor is infalling or receding,” so disturbance magnitude and disturbance timing are partly decoupled; the paper therefore introduces blue fraction and X-ray/BCG offset as auxiliary timing tracers (Meyer et al., 2018, Anfodillo et al., 2012, Kong et al., 25 Sep 2025).

Statistical and quantum uses expose additional caveats. Perturbed variation can be zero for $\operatorname{PV}_\varepsilon(P,Q)$25 if the distributions can be coupled within $\operatorname{PV}_\varepsilon(P,Q)$26, so it is not a metric and it inherits strong sensitivity to the chosen tolerance scale. The operator-based quantum disturbance is state dependent and can be invisible in single-time marginals, which is precisely why sequential statistics or weak-probe circuits are required for experimental evaluation. Disturbance-enhanced uncertainty relations therefore convert measured disturbance into rigorous lower bounds, but those bounds depend on the selected divergence and on the measurement ordering (Harel et al., 2012, Ozawa, 2021, Sun et al., 2022, Emori et al., 2024).

Taken together, these constructions indicate that “Disturbance Score” functions less as a fixed formula than as a design pattern. A disturbance score typically couples three ingredients: a nominal model or baseline, a discrepancy or margin functional, and an interpretation layer tied to action or inference. The specific mathematics may be recurrence geometry, wavelet energy, regression confidence, quantum operator change, optimal transport with tolerance, or distance to a resilience boundary, but the shared purpose is to compress disturbance-relevant structure into a calibrated scalar or low-dimensional diagnostic that is technically interpretable within its native domain.