11. Nonlinear and Higher-Order Signal Processing

Introduction and Context

Most introductory digital signal-processing methods are built on linearity, second-order statistics, and approximate Gaussian assumptions. These assumptions are powerful: linear time-invariant models are mathematically tractable, autocorrelation summarizes the dependence structure of many random processes, and the power spectral density provides an effective description of energy distribution over frequency. However, important engineering signals do not always satisfy these conditions. Speech production, biomedical activity, radar echoes, communication channels with nonlinear amplifiers, mechanical vibration, and biological sensors may contain asymmetry, impulsiveness, harmonic coupling, phase information, or dependence structures that second-order methods cannot reveal.

Nonlinear and higher-order signal processing extends the signal-processing toolbox to such situations. Nonlinear processing describes systems in which the output is not a linear combination of input samples. Higher-order statistical processing uses moments and cumulants of order greater than two, thereby capturing information discarded by mean and covariance analysis. Related tools such as Volterra models, nonlinear filters, robust estimators, bispectra, trispectra, and cyclostationary statistics support detection, identification, denoising, classification, synchronization, and fault diagnosis.

The central challenge is that additional descriptive power is accompanied by additional complexity. Higher-order estimates require more data, nonlinear models may be nonconvex or poorly identifiable, and an inappropriate estimator can amplify outliers rather than suppress them. Effective practice therefore depends not only on selecting an advanced method, but also on understanding assumptions, statistical reliability, computational cost, and the physical origin of the observed signal.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Explain why linear and second-order methods may be inadequate for nonlinear or non-Gaussian signals.

2. Formulate and interpret finite-order Volterra models.

3. Distinguish nonlinear filtering structures and select suitable estimation strategies.

4. Explain robust estimators and their use in impulsive or contaminated noise.

5. Define higher-order moments, cumulants, polyspectra, and their principal properties.

6. Interpret bispectral and trispectral information in practical signal-analysis problems.

7. Explain cyclostationarity and distinguish it from ordinary stationarity.

8. Design an appropriate nonlinear or higher-order processing workflow while recognizing limitations and risks.

Why Second-Order Processing Is Sometimes Insufficient

For a random process x(n), the mean and autocorrelation are defined as

μx = E{x(n)},

Rx(k) = E{x(n)x*(n−k)}.

For a wide-sense stationary process, the autocorrelation depends only on the lag k, and its Fourier transform is the power spectral density. These quantities describe average level, variance, correlation, and energy distribution. If x(n) is Gaussian, they completely determine its probability law. Consequently, no additional statistical information is required for many Gaussian linear systems.

The situation changes for non-Gaussian processes. Two processes may have the same mean and autocorrelation but different probability distributions, asymmetry, impulsiveness, or phase coupling. A power spectrum also cannot distinguish certain signals with different phase relationships. For example, a signal containing interacting components at frequencies f1 and f2 may generate a component at f1+f2 through a quadratic nonlinearity. The power spectrum can show energy at these frequencies, but it does not by itself establish whether the components are phase coupled. Higher-order spectra can provide that evidence.

Nonlinearity may be present in the source, the propagation medium, the sensor, or the measurement electronics. Saturation, dead zones, quantization, clipping, multiplicative noise, and nonlinear resonance are common mechanisms. A linear model may then produce biased residuals, fail to predict large excursions, or incorrectly attribute deterministic structure to random noise.

Volterra Models: A Systematic Representation of Nonlinearity

The Volterra series is a functional extension of the convolution representation of a linear system. For a causal, time-invariant discrete-time system, an Nth-order truncated model can be written as

y(n) = h0 + Σ h1(k1)x(n−k1)

+ ΣΣ h2(k1,k2)x(n−k1)x(n−k2)

+ ⋯

+ Σ⋯Σ hN(k1,…,kN) Π x(n−ki).

Here h0 is a constant term and hp is the pth-order Volterra kernel. The first-order kernel is equivalent to the impulse response of a linear filter. The second-order kernel describes interactions between pairs of delayed input samples, while higher-order kernels represent increasingly complex interactions.

For a finite-memory second-order model,

y(n) = h0 + Σk h1(k)x(n−k) + Σk1Σk2 h2(k1,k2)x(n−k1)x(n−k2).

If h2 is symmetric, h2(k1,k2)=h2(k2,k1), the number of independent coefficients can be reduced. This symmetry is useful in estimation because otherwise the same product appears repeatedly in different orders. Volterra models can approximate a broad class of fading-memory nonlinear systems, but approximation quality depends on memory depth and order.

A practical estimation method is to construct a lifted regression vector containing products of delayed samples. For example, a second-order model may use

z(n) = [1, x(n), …, x(n−M), x²(n), x(n)x(n−1), …]T.

The model becomes y(n)=θTz(n)+e(n), which permits least-squares or regularized least-squares estimation. The apparent simplicity hides a major problem: the number of terms grows rapidly with order and memory. Regularization, sparse selection, symmetry constraints, and orthogonal polynomial bases are therefore important.

Volterra models are useful for amplifier characterization, nonlinear echo cancellation, acoustic systems, biomedical modeling, and system identification. They are less attractive when the system has long memory and strong high-order behavior, or when the available data set is too small to support reliable estimation. In such cases, block-oriented models, neural networks, kernel methods, or state-space nonlinear models may be more economical.

Nonlinear Filtering Architectures

A nonlinear filter maps an input sequence, or a set of input sequences, to an output without satisfying the superposition principle. The distinction between a nonlinear filter and a linear adaptive filter is not merely computational. The filtering objective, error criterion, stability behavior, and interpretation of coefficients may all change.

A memoryless nonlinear filter applies a static transformation, y(n)=g(x(n)). Examples include clipping, logarithmic compression, square-law detection, and soft limiting. Such filters are simple but cannot model temporal interactions. A nonlinear filter with memory may be represented by a Volterra structure, a polynomial filter, a radial-basis-function model, a neural network, or a state-dependent recursion.

In a Wiener structure, a linear dynamic filter is followed by a static nonlinearity:

v(n)=Σk h(k)x(n−k), y(n)=g(v(n)).

In a Hammerstein structure, the order is reversed:

v(n)=g(x(n)), y(n)=Σk h(k)v(n−k).

Wiener–Hammerstein arrangements combine both forms and are useful when a physical system contains dynamic filtering before and after a nonlinear element. The choice of structure should follow the mechanism being modeled, not simply the availability of a convenient algorithm.

Nonlinear prediction and equalization are important communication applications. A nonlinear equalizer may compensate for amplifier distortion, optical-channel effects, or phase-amplitude coupling. In speech and audio processing, nonlinear filters can suppress transients while preserving perceptually important structure. In image and sensor processing, median, bilateral, and rank-order filters provide nonlinear alternatives to ordinary smoothing.

Robust Estimation and Non-Gaussian Disturbances

The ordinary least-squares criterion minimizes

J = Σn e²(n),

where e(n) is the prediction or modeling error. Squaring gives large errors disproportionate influence. This is appropriate for Gaussian disturbances, but it is vulnerable to impulsive noise, faulty observations, and heavy-tailed distributions.

Robust estimation replaces the quadratic loss by a function ρ(e). The estimate minimizes Σρ(e(n)). The absolute-deviation criterion uses ρ(e)=|e| and is less sensitive to outliers. The Huber loss is quadratic for small errors and approximately linear for large errors:

ρδ(e) = 1/2 e², for |e|≤δ;

ρδ(e) = δ|e|−1/2δ², for |e|>δ.

This preserves efficient behavior near the presumed model while limiting the effect of extreme observations. M-estimators, least-absolute-deviation methods, trimmed estimators, median-based filters, and maximum-correntropy criteria are additional choices. The latter uses a localized similarity measure and can be effective for heavy-tailed error distributions, although bandwidth selection strongly affects performance.

Robustness has several meanings. An estimator may be resistant to isolated outliers, stable under heavy-tailed noise, insensitive to small model errors, or reliable when a fraction of samples is corrupted. These properties are not identical. A method designed for impulsive noise may lose efficiency when the data are genuinely Gaussian. Robust parameter selection should therefore consider contamination level, desired bias, computational constraints, and the cost of false decisions.

Higher-Order Moments and Cumulants

The pth-order moment of a random variable involves E{x^p}, with appropriate conjugation for complex signals. Moments describe distributional shape, but raw moments can contain redundant information arising from lower-order interactions. Cumulants remove this redundancy. For a zero-mean process, the fourth-order cumulant can be expressed schematically as

cum{x1,x2,x3,x4} = E{x1x2x3x4}

− E{x1x2}E{x3x4}

− E{x1x3}E{x2x4}

− E{x1x4}E{x2x3}.

For a Gaussian process, all cumulants above order two are zero. This makes higher-order cumulants useful for detecting non-Gaussianity and separating non-Gaussian signals from Gaussian noise. In practice, sample estimates of cumulants are biased or variable for short records, and preprocessing such as mean removal and appropriate normalization is essential.

Important properties include translation invariance for cumulants of order greater than one, sensitivity to non-Gaussian structure, and behavior under linear filtering. A linear filter transforms higher-order cumulants according to products of its frequency response. Thus, cumulants can support blind identification and source separation. Their sign and magnitude may also distinguish distributions with similar variance but different tail or symmetry characteristics.

Bispectrum and Polyspectral Analysis

The Fourier transform of a third-order cumulant is the bispectrum. For a stationary process, one common form is

B(f1,f2) = E{X(f1)X(f2)X*(f1+f2)},

where X(f) denotes a Fourier transform or spectral representation. The bispectrum measures quadratic phase coupling: a nonzero value indicates statistical interaction among components at f1, f2, and f1+f2, subject to estimation assumptions.

The fourth-order counterpart is the trispectrum, which examines interactions among four frequency components. More generally, polyspectra are Fourier transforms of higher-order cumulant functions. Because of symmetries, only a principal region of the frequency plane or higher-dimensional domain may need to be analyzed.

The squared bicoherence is a normalized form of the bispectrum. A value near one indicates strong, consistently phase-related coupling under the adopted estimator, whereas a value near zero indicates weak coupling. Magnitude alone is not proof of physical causation: finite samples, leakage, harmonics, and nonstationarity can create apparent coupling. Segment averaging, window selection, confidence thresholds, and surrogate-data tests improve interpretation.

Applications include diagnosis of nonlinear vibration, detection of quadratic phase coupling in communications and radar, analysis of EEG and evoked potentials, identification of nonlinear systems, and characterization of harmonic generation. Bispectral analysis is particularly informative when the power spectrum contains overlapping components but the phase relationship carries discriminating information.

Cyclostationarity: Using Periodic Statistical Structure

A process is cyclostationary when one or more of its statistical properties vary periodically with time. A first-order cyclostationary mean satisfies μx(n+T)=μx(n), while a second-order cyclostationary correlation satisfies

R_x(n+T,k)=R_x(n,k)

for some period T. Communication signals often exhibit cyclostationarity because of symbol timing, carrier modulation, framing, coding, or periodic pilot structures. Rotating machinery may exhibit periodic statistics associated with shaft rotation.

The cyclic autocorrelation function is obtained by correlating the signal with a frequency-shifted version and averaging over time. Its Fourier transform with respect to lag yields the spectral correlation function (SCF), commonly denoted Sxα(f), where α is the cyclic frequency. Ordinary stationary spectra are associated with α=0; nonzero cyclic frequencies reveal periodic statistical structure.

Cyclostationary analysis is valuable in low signal-to-noise ratio detection because structured interference and the desired signal may have different cyclic frequencies. It supports spectrum sensing, modulation classification, synchronization, fault diagnosis, and separation of cochannel signals. The cost is substantial: reliable estimation needs adequate observation time, and errors in assumed cycle frequencies can obscure the structure. Nonstationary signals may also require time-varying or evolutionary spectral methods rather than a strictly cyclostationary model.

An Integrated Processing Workflow

An effective nonlinear or higher-order analysis begins with the physical question. The analyst should first identify whether the objective is prediction, denoising, system identification, detection, classification, or source separation. The data should then be inspected for saturation, missing samples, impulsive events, changing variance, periodicity, and dependence on operating conditions.

A sensible workflow is as follows:

1. Establish a baseline using mean, variance, autocorrelation, power spectrum, and a linear model.

2. Test for non-Gaussianity, outliers, nonlinear residual structure, and periodic statistical variation.

3. Select a model consistent with the suspected mechanism: robust filter, Volterra model, Wiener or Hammerstein structure, cumulant method, bispectral estimator, or cyclostationary detector.

4. Estimate parameters using regularization, segmentation, and validation data.

5. Compare against a simpler baseline using prediction error, detection probability, false-alarm rate, computational cost, and stability.

6. Verify that the apparent improvement is not caused by leakage, overfitting, parameter tuning on test data, or changing data distributions.

Cross-validation is especially important for flexible nonlinear models. For higher-order spectra, confidence intervals or surrogate-data comparisons should accompany visual displays. Reproducibility requires recording sampling rate, window, overlap, detrending, normalization, model order, regularization, and threshold choices.

Applied Case Study: Nonlinear Fault Detection in a Rotating Machine

Background

Consider an illustrative condition-monitoring system for a motor-driven gearbox. An accelerometer measures vibration at a high sampling rate. Under normal operation, the signal contains shaft-related periodic components, broadband measurement noise, and weak harmonics. A developing gear defect introduces repeated impacts and nonlinear contact dynamics.

Situation and Problem

A conventional power-spectrum monitor detects increased energy near a gear-mesh frequency, but load changes also alter that energy. False alarms occur because the power spectrum cannot distinguish load-related amplitude variation from phase-coupled components produced by a defect.

Analysis

The data are divided into records aligned with approximately constant operating conditions. A linear baseline estimates the power spectrum and an autoregressive prediction error. Examination of residuals shows heavy tails, indicating that occasional impacts are not well represented by Gaussian noise. The bispectrum is then estimated using averaged, windowed segments. A repeatable peak appears near frequency pairs whose sum equals the gear-mesh component. This suggests quadratic phase coupling rather than a simple independent increase in spectral power.

Because impact amplitudes are extreme, ordinary least-squares prediction is also unstable. A Huber-loss adaptive predictor is trained on normal-operation data. Its robust residual is used as a complementary indicator. Cyclostationary analysis is applied at shaft-related cyclic frequencies, separating periodic defect structure from broadband disturbances.

Approach and Solution

The monitoring design combines three features: robust prediction-error statistics, normalized bicoherence in a physically meaningful frequency region, and cyclic spectral energy at the shaft period. Thresholds are selected from normal-operation records and tested on separate records containing known maintenance events. A regularized second-order Volterra predictor is evaluated, but its additional complexity provides little improvement over the robust linear predictor for early detection; it is retained only for cases with confirmed amplitude-dependent distortion.

Outcome

The combined indicator is more stable across load conditions than the power-spectrum threshold alone. The bispectral feature contributes discrimination when harmonic energy is ambiguous, while the robust residual remains effective when isolated impacts occur. The system does not automatically identify the exact failed component; it flags a condition requiring engineering inspection.

Learning Points

The case illustrates four principles. First, higher-order features should be linked to a physical mechanism. Second, robust estimation and higher-order analysis solve different problems and can be complementary. Third, a more complicated model is not automatically a better monitoring model. Finally, validation must include operating conditions not used for threshold selection.

Challenges, Limitations, and Common Errors

Higher-order methods suffer from a severe data requirement. Variance of sample moment and polyspectral estimates can be large, especially at high order. Longer records improve statistical stability but may violate stationarity assumptions. There is therefore a trade-off between resolution, stationarity, and estimator variance.

Model-order selection is another difficulty. Excessively high Volterra order leads to parameter explosion and overfitting, while an order that is too low misses relevant interactions. Regularization helps but does not replace physical reasoning. Poorly scaled polynomial features can also create numerical ill-conditioning.

Interpretation can be misleading. A bispectral peak may result from deterministic harmonics, window leakage, or finite-record effects rather than nonlinear coupling. A nonzero higher-order cumulant can arise from non-Gaussian measurement noise rather than the target phenomenon. Cyclostationary peaks may be generated by instrumentation or data-acquisition clocks.

Common mistakes include applying nonlinear processing before removing sensor saturation, comparing unnormalized spectra, ignoring complex-signal conventions, using the same record for model selection and performance claims, and reporting an improvement without a false-alarm analysis. The analyst should also avoid assuming that Gaussian noise is always harmless: nonlinear transformations can convert Gaussian noise into statistically dependent, non-Gaussian components.

Professional, Ethical, and Governance Considerations

Signal-processing decisions may affect safety, maintenance, medical diagnosis, communication access, or industrial production. A nonlinear detector should therefore provide uncertainty information and an interpretable rationale when feasible. Automated alarms should not be represented as definitive diagnoses unless independently validated for that purpose.

Data governance includes secure handling of sensor records, documentation of calibration, protection of personally identifiable biomedical or acoustic data, and retention of preprocessing metadata. Models trained under one operating regime may fail under another; deployment should include drift monitoring and a procedure for human review. In safety-related applications, traceability, fail-safe behavior, and independently verified thresholds are more important than marginal gains in average accuracy.

Current Developments and Best Practices

Recent work increasingly combines classical higher-order statistics with learned representations. Neural networks can approximate nonlinear mappings and extract features from raw data, while cumulants, cyclic spectra, and bispectral descriptors provide physically meaningful inputs or validation tools. Physics-informed and hybrid models are especially attractive when data are limited but system structure is known.

Computational advances support streaming estimation, GPU-based spectral analysis, sparse Volterra kernels, and online robust adaptation. Nevertheless, increased computational capacity does not remove identifiability or data-quality constraints. Best practice remains to:

• begin with a transparent linear and second-order baseline;
• state stationarity, ergodicity, and noise assumptions explicitly;
• use robust preprocessing without destroying the phenomenon of interest;
• constrain model order and regularize high-dimensional estimates;
• use segmentation, averaging, normalization, and confidence assessment for polyspectra;
• validate across operating conditions and compare with simpler alternatives;
• preserve reproducible parameter and calibration records; and
• combine automatic indicators with domain expertise in consequential decisions.

Chapter Summary

Nonlinear and higher-order signal processing is required when linear models, Gaussian assumptions, and power spectra fail to describe important signal behavior. Volterra series provide a systematic finite-order representation of nonlinear dynamic systems, while Wiener, Hammerstein, and related block structures encode particular arrangements of dynamics and static nonlinearities. Robust estimators reduce sensitivity to outliers and impulsive disturbances by replacing the quadratic loss with more suitable criteria.

Higher-order moments and cumulants reveal non-Gaussian structure, and their Fourier transforms produce polyspectra such as the bispectrum and trispectrum. Bispectral analysis is especially useful for examining quadratic phase coupling. Cyclostationary analysis exploits periodic variation in statistical properties and is important in communications, machinery, and detection. Effective application requires careful model selection, sufficient data, validation, confidence assessment, and awareness of leakage, overfitting, nonstationarity, and physical confounding. The strongest professional practice combines advanced methods with transparent baselines and domain-informed interpretation.

Key Terms

Volterra series: A functional expansion representing a nonlinear dynamic system through kernels of increasing order.

Volterra kernel: The coefficient function associated with one order of a Volterra expansion.

Nonlinear filter: A filter whose output does not satisfy linear superposition.

Wiener model: A cascade of a linear dynamic block followed by a static nonlinearity.

Hammerstein model: A cascade of a static nonlinearity followed by a linear dynamic block.

Robust estimator: An estimator designed to limit the influence of outliers or model contamination.

M-estimator: An estimator obtained by minimizing a chosen loss function applied to residuals.

Huber loss: A loss that is quadratic for small residuals and linear for large residuals.

Cumulant: A statistical quantity that removes lower-order interaction contributions from moments.

Higher-order statistic: A statistic involving moments or cumulants above second order.

Bispectrum: The Fourier transform of a third-order cumulant or a third-order spectral representation.

Bicoherence: A normalized bispectral measure of quadratic phase coupling.

Polyspectrum: A frequency-domain representation of a higher-order cumulant.

Cyclostationarity: Periodic variation of one or more statistical properties of a process.

Cyclic frequency: A frequency at which periodic statistical structure appears.

Spectral correlation function: A frequency-domain representation of cyclic correlation.

Quadratic phase coupling: Consistent statistical phase interaction among frequency components and their sum.

Impulsive noise: Noise characterized by occasional, unusually large-amplitude disturbances.

Regularization: A constraint or penalty used to improve conditioning and reduce overfitting.

Non-Gaussianity: Deviation of a signal’s probability structure from the Gaussian model.

Review Questions

Short Answer Questions

1. Why can two processes with the same power spectrum still have different higher-order behavior?

2. What is the role of the second-order Volterra kernel?

3. Distinguish a Wiener model from a Hammerstein model.

4. Why is ordinary least squares sensitive to impulsive noise?

5. What information does bicoherence provide beyond the power spectrum?

Long Answer Questions

1. Derive and explain a finite-memory second-order Volterra model, including the effects of kernel symmetry and parameter growth.

2. Discuss robust estimation methods for non-Gaussian disturbances, comparing least squares, least absolute deviations, and Huber estimation.

3. Explain higher-order cumulants and polyspectra, and describe how the bispectrum can identify quadratic phase coupling.

4. Explain cyclostationarity, cyclic autocorrelation, and the spectral correlation function, with applications in communication systems.

5. Develop a complete workflow for selecting and validating a nonlinear signal-processing method for an industrial monitoring problem.

Analytical / Application Questions

1. A vibration signal has a strong gear-mesh peak whose amplitude changes with load. Design an analysis using robust prediction, bicoherence, or cyclostationary features to reduce false alarms, and justify each choice.

2. A communication receiver experiences amplifier saturation and impulsive interference. Compare a Volterra equalizer with a robust nonlinear adaptive filter, considering model accuracy, data requirements, and computational cost.

3. Two signals have nearly identical power spectra, but only one is generated by a quadratic nonlinear oscillator. Explain an experiment using third-order cumulants or bispectral analysis to distinguish them and identify possible sources of false evidence.

22. Multiple Choice Questions with Answers

Question 1

Which property most directly motivates the use of higher-order cumulants for detecting non-Gaussian structure?

A. All cumulants of a Gaussian process above second order are zero.

B. Higher-order cumulants are independent of sampling rate.

C. Higher-order cumulants always have lower variance than autocorrelation estimates.

D. Cumulants eliminate all effects of filtering.

Correct Answer: A. All cumulants of a Gaussian process above second order are zero.

Question 2

In a second-order Volterra model, the second-order kernel primarily represents:

A. The system’s constant offset only.

B. Interactions between pairs of delayed input samples.

C. Only independent additive measurement noise.

D. The Fourier transform of the power spectrum.

Correct Answer: B. Interactions between pairs of delayed input samples.

Question 3

A Wiener model consists of:

A. A static nonlinearity followed by a linear dynamic block.

B. Two independent stochastic processes combined by addition.

C. A linear dynamic block followed by a static nonlinearity.

D. A higher-order spectrum followed by an adaptive filter.

Correct Answer: C. A linear dynamic block followed by a static nonlinearity.

Question 4

Why is Huber loss often useful for impulsive disturbances?

A. It gives increasingly large quadratic weight to extreme errors.

B. It treats all errors as exactly equal.

C. It is quadratic for small errors and limits the influence of large errors.

D. It removes the need for parameter estimation.

Correct Answer: C. It is quadratic for small errors and limits the influence of large errors.

Question 5

A nonzero bispectral value involving f1, f2, and f1+f2 most directly suggests:

A. Quadratic phase coupling among those components.

B. That the signal is necessarily white noise.

C. That the process is strictly Gaussian.

D. The absence of harmonic interactions.

Correct Answer: A. Quadratic phase coupling among those components.

Question 6

Which situation is most suitable for cyclostationary analysis?

A. A signal whose statistical properties contain periodic structure caused by symbol timing.

B. A completely memoryless signal with no temporal variation.

C. A deterministic constant with no modulation or periodicity.

D. A signal for which only the sample mean is available.

Correct Answer: A. A signal whose statistical properties contain periodic structure caused by symbol timing.

Question 7

A major practical disadvantage of increasing Volterra order and memory length is:

A. The model becomes unable to represent linear behavior.

B. The number of parameters grows rapidly, increasing overfitting risk.

C. The model automatically becomes independent of the input.

D. The model can only process Gaussian signals.

Correct Answer: B. The number of parameters grows rapidly, increasing overfitting risk.

Question 8

In the illustrative gearbox case, why was robust prediction combined with bispectral analysis?

A. Both methods measure exactly the same feature and are redundant.

B. Robust prediction handles impacts, while bispectral analysis examines phase coupling.

C. Bispectral analysis eliminates the need for validation data.

D. Robust prediction directly estimates the failed mechanical component.

Correct Answer: B. Robust prediction handles impacts, while bispectral analysis examines phase coupling.

Question 9

Which practice best improves the reliability of a finite-record bispectral estimate?

A. Avoiding segmentation and using a single unwindowed record.

B. Using averaging, appropriate windows, and confidence or surrogate-data assessment.

C. Increasing model order without increasing data length.

D. Interpreting every visible peak as proof of causation.

Correct Answer: B. Using averaging, appropriate windows, and confidence or surrogate-data assessment.

Question 10

Why should a nonlinear detector be compared with a simpler baseline?

A. To determine whether added complexity produces validated improvement rather than overfitting.

B. Because nonlinear methods cannot process real signals.

C. Because power spectra contain all possible signal information.

D. To guarantee that the detector will identify the exact physical fault.

Correct Answer: A. To determine whether added complexity produces validated improvement rather than overfitting.