Extreme Value Theory (EVT)

Traditional statistical models focus on central tendency (average channel holding times), which fails to capture the rare, catastrophic events that cause link failure. EVT specifically models the tail distribution of the data, providing precise probabilities for extreme channel busy periods that exceed a high threshold. This is critical because the mean time-to-failure is often irrelevant; what matters is the probability of a forced termination during an unusually long primary user transmission.

The Balkema-de Haan theorem states that for a sufficiently high threshold u, the distribution of excesses over that threshold converges to the Generalized Pareto Distribution (GPD). In spectrum analysis, this means the duration of a primary user's transmission beyond a typical busy period can be modeled with a GPD characterized by a shape parameter (ξ) and scale parameter (σ). A positive shape parameter (ξ > 0) indicates a heavy-tailed distribution, meaning extremely long busy periods are more likely than a standard exponential model would predict.

EVT offers two primary sampling strategies for spectrum data:

• Block Maxima: Divide the observation timeline into fixed blocks (e.g., 1-hour windows) and fit a Generalized Extreme Value (GEV) distribution to the maximum channel busy duration within each block. This is useful for daily or hourly worst-case analysis.
• Peaks-Over-Threshold (POT): Select all observations exceeding a high threshold u. This method uses data more efficiently by retaining all extreme events, not just block maxima, and fits a GPD to the excesses. POT is preferred for real-time spectrum mobility prediction due to its higher data utilization.

A core output of EVT is the return level, defined as the channel busy duration expected to be exceeded once every m observation periods. For a cognitive radio, the N-year return level (or N-hour return level) answers: 'What is the maximum primary user transmission length I should expect in the next N hours?' This directly informs the prediction horizon and the selection of a target channel for proactive handoff. A channel with a high return level for busy periods is a poor candidate for secondary transmission.

The choice of threshold u is a critical bias-variance trade-off in POT modeling. A threshold too low violates the asymptotic basis of the GPD, introducing bias; a threshold too high leaves too few exceedances, inflating variance. The Mean Residual Life (MRL) plot is the primary diagnostic tool: it plots the average excess over u against u. The threshold is selected at the point where the plot becomes approximately linear, indicating the region where the GPD is a valid model for the spectrum data.

EVT provides not just point estimates of extreme quantiles but also confidence intervals via profile likelihood or delta methods. For a spectrum mobility predictor, this means the system can output a predicted channel holding time with an associated uncertainty bound (e.g., 'the 99th percentile busy duration is 450ms ± 50ms'). This allows the cognitive radio to make risk-aware decisions, choosing a handoff strategy that accounts for the worst-case scenario within a specified confidence level, rather than relying on a single, potentially fragile prediction.

The cornerstone distribution of EVT used to model exceedances over a high threshold. In spectrum mobility, the GPD models the distribution of channel busy periods that exceed a critical latency threshold.

• Shape parameter (ξ): Determines the tail heaviness (heavy-tailed if ξ > 0).
• Scale parameter (σ): Controls the dispersion of the excesses.
• Application: Estimating the probability of a primary user occupying a channel for an abnormally long duration, triggering a forced termination.

A classical EVT method that divides a time series into non-overlapping blocks and fits the Generalized Extreme Value (GEV) distribution to the maximum observation in each block.

• Use case: Modeling the worst-case interference level observed in a frequency band over fixed monitoring intervals.
• GEV families: Weibull (finite endpoint), Gumbel (light tail), and Fréchet (heavy tail).
• Trade-off: Simpler than threshold methods but discards data by using only block maxima.

The modern EVT methodology that models all observations exceeding a high threshold u, rather than just block maxima. This is the preferred approach for spectrum occupancy due to efficient data usage.

• Threshold selection: A bias-variance trade-off; too low violates asymptotic theory, too high increases variance.
• Mean Residual Life Plot: A diagnostic tool for selecting the optimal threshold u.
• Declustering: Technique to ensure temporal independence of exceedances by separating clusters of threshold violations.

A key EVT output that estimates the value expected to be exceeded once every T time units (the return period). In spectrum mobility, this quantifies worst-case channel occupancy durations.

• N-year return level: The busy period duration expected to be exceeded once every N observation windows.
• Confidence intervals: Typically computed via profile likelihood or delta method to quantify estimation uncertainty.
• Application: Dimensioning spectrum handoff buffers to guarantee a target forced termination probability.

While univariate EVT models marginal tail behavior, copula models capture the joint tail dependence between multiple frequency channels. This is critical when primary user activity is correlated across bands.

• Tail dependence coefficient: Measures the probability of simultaneous extreme occupancy on two channels.
• Clayton copula: Captures lower tail dependence, useful for modeling joint idle periods.
• Gumbel copula: Captures upper tail dependence, modeling simultaneous busy extremes.
• Relevance: Prevents underestimating the risk of correlated spectrum unavailability.

A parameter (θ) that measures the tendency of extreme events to cluster in time. A value of θ < 1 indicates clustering, which is common in bursty primary user traffic patterns.

• Interpretation: 1/θ is the mean cluster size of extreme exceedances.
• Estimation methods: Runs estimator, blocks estimator, or intervals estimator.
• Impact: Failing to account for clustering leads to overly optimistic return level estimates and underestimation of handoff failure risk.