Copula Model

Captures the probability that multiple frequency channels experience simultaneous extreme occupancy events. Unlike Kendall's Tau or Pearson correlation, copulas specifically model lower tail dependence (channels becoming busy together) and upper tail dependence (channels staying idle together). This is critical for assessing the risk of a secondary user losing all available backup channels at once during a primary user traffic surge.

The mathematical backbone of copula modeling. Sklar's theorem proves that any multivariate joint distribution can be decomposed into two independent parts:

• Marginal distributions: The individual occupancy behavior of each channel (e.g., busy/idle ratios)
• Copula function: The pure dependence structure linking the channels

This separation allows engineers to model channel marginals with high accuracy (e.g., using phase-type distributions) and then independently select the best copula family for the dependence.

Common parametric copula families used in spectrum mobility prediction:

• Clayton Copula: Strong lower tail dependence. Models scenarios where channels become busy together during interference bursts.
• Gumbel Copula: Strong upper tail dependence. Captures channels remaining idle together during extended spectrum availability windows.
• Frank Copula: Symmetric dependence with no tail emphasis. Suitable for channels with uniform correlation across all occupancy states.

Each family is defined by a generator function φ(t) and a single dependence parameter θ.

Linear measures like Kendall's Tau only detect monotonic relationships. Copulas reveal complex dependence patterns such as:

• Asymmetric dependence: Channel A's busy state strongly predicts Channel B's state, but not vice versa
• Tail asymmetry: Channels are highly correlated during congestion but independent during low utilization
• Non-monotonic relationships: Correlation that changes direction at different occupancy quantiles

This enables more accurate joint channel state prediction for proactive handoff target selection.

Fitting a copula to spectrum data involves two stages:

• Inference Functions for Margins (IFM): First estimate marginal parameters for each channel independently, then estimate the copula dependence parameter θ using maximum likelihood on the joint pseudo-observations
• Canonical Maximum Likelihood (CML): Uses empirical cumulative distribution functions to transform data to uniform margins without assuming parametric marginal models
• Bayesian estimation: Places priors on θ and uses MCMC sampling to obtain posterior distributions, quantifying uncertainty in the dependence structure

Extends static copulas to capture dynamic dependence evolution in spectrum environments:

• Dynamic Conditional Copula (DCC): Allows the dependence parameter θ_t to evolve via an autoregressive process, capturing how channel correlations shift during peak vs. off-peak hours
• Regime-switching copulas: Uses a hidden Markov model to switch between different copula families as the spectrum environment transitions between normal and congested states
• Rolling window estimation: Re-estimates copula parameters on sliding time windows to adapt to concept drift in primary user traffic patterns