Granger Causality

The foundational axiom of Granger causality: a cause must occur before its effect. In spectrum mobility, this means past occupancy states of a candidate channel must contain statistically significant information about the future state of a target channel that is not available from the target channel's own history. The test is strictly directional in time, comparing the forecast error variance of a restricted model (using only the target channel's past) against an unrestricted model (adding the candidate channel's past).

The quantitative metric at the heart of the test. Granger causality is established if the Mean Squared Error (MSE) of a prediction model is significantly reduced by incorporating past observations of another time series. Key operational details include:

• Restricted Model: Y_t = α + Σ β_i * Y_{t-i} + ε_t
• Unrestricted Model: Y_t = α + Σ β_i * Y_{t-i} + Σ γ_j * X_{t-j} + η_t
• Causality Condition: Var(η_t) < Var(ε_t) with statistical significance. This directly quantifies the predictive utility of one channel's spectral history for another.

The validity of the standard F-test or Wald test used in Granger causality analysis depends critically on the covariance stationarity of the input time series. A stationary process has a constant mean, variance, and autocorrelation structure over time. In spectrum occupancy data, non-stationarity often arises from diurnal usage patterns. Violations require preprocessing steps such as:

• Differencing: Applying ΔY_t = Y_t - Y_{t-1} to remove trends.
• Detrending: Subtracting a rolling mean or fitted polynomial. Failure to ensure stationarity leads to spurious regression results, falsely indicating a causal link.

The choice of p (the number of past time steps included in the model) is a critical hyperparameter. An insufficient lag length omits relevant historical information, while an excessive lag length introduces multicollinearity and reduces the power of the statistical test. Selection is typically guided by information criteria:

• Akaike Information Criterion (AIC): Balances model fit with complexity.
• Bayesian Information Criterion (BIC): Imposes a stricter penalty for additional parameters, favoring more parsimonious models. The optimal lag is the one that minimizes the chosen criterion.

A crucial limitation: standard Granger causality tests are designed to detect linear predictive relationships within a vector autoregressive (VAR) framework. It is a test of incremental predictive power, not a test of true mechanistic causation. A finding of non-causality does not rule out a non-linear causal link that a linear model cannot capture. For spectrum applications, this motivates the use of non-linear extensions like kernel Granger causality or neural network-based tests to uncover complex, non-linear dependencies between frequency channels.

In a multi-channel spectrum environment, a spurious causal link can be detected between channels X and Y if both are driven by a common confounding variable Z (e.g., a primary user's activity pattern). Conditional Granger causality extends the test by including the past of Z in both the restricted and unrestricted models. This tests if X provides unique predictive information about Y beyond the information already provided by Z, isolating direct channel-to-channel influences from network-wide correlations.

A model-free, information-theoretic alternative to Granger Causality that quantifies the directed flow of information between two time series. It measures the reduction in uncertainty about a future channel state given the past states of another channel.

• Key Advantage: Captures non-linear causal interactions that linear Granger tests miss.
• Application: Used to detect non-linear coupling between spectrum occupancy patterns in heterogeneous networks.
• Calculation: Based on conditional mutual information, often estimated using k-nearest neighbor statistics.

A causality detection method designed for non-linear dynamic systems where Granger Causality fails. CCM tests if the historical states of one variable can be reliably reconstructed from the time-delayed embedding of another.

• Core Principle: If X causes Y, then information about X is encoded in the time series of Y.
• Spectrum Use Case: Identifies causal links in chaotic or complex primary user traffic patterns.
• Indicator: Causality is confirmed if cross-map prediction skill increases with the length of the time series.

A multivariate statistical model that captures the linear interdependencies among multiple time series. Each channel's occupancy is modeled as a linear function of its own past values and the past values of all other channels.

• Granger Implementation: The standard framework for performing a Granger Causality test on stationary data.
• Key Output: An F-statistic and p-value for each directed pair, testing if lagged coefficients are jointly significant.
• Stability Requirement: Requires the VAR model to be stable, meaning all eigenvalues of the companion matrix lie inside the unit circle.

A deep learning approach that uses attention-based convolutional neural networks to discover causal relationships in multivariate time series without pre-specifying a lag order.

• Mechanism: Applies dilated convolutions to capture long-range dependencies and an attention mechanism to weight the importance of past time steps.
• Causal Validation: The Attention Interpretation of the network's learned weights reveals the causal time delay and strength between channels.
• Advantage: Simultaneously predicts future channel states and identifies the underlying causal graph in a single end-to-end model.

A two-stage causal discovery algorithm that combines conditional independence testing with a modified PC-algorithm to handle high-dimensional, auto-correlated time series data.

• Stage 1: Estimates a set of relevant conditions for each variable using a variant of the PC algorithm.
• Stage 2: Calculates the momentary conditional independence (MCI) test to remove spurious correlations from autocorrelation.
• Spectrum Application: Robustly identifies the true causal drivers of spectrum occupancy by controlling for both lagged and contemporaneous confounding effects.

A dynamic neural network architecture where the prediction of a target channel's state depends on its own past states and the past states of an exogenous input channel.

• Causal Interpretation: If adding the exogenous channel's history significantly reduces prediction error compared to a purely autoregressive model, a non-linear causal link is inferred.
• Architecture: Typically implemented with a feedforward network with tapped delay lines or a recurrent network.
• Training: Uses series-parallel mode for stability, switching to parallel mode for multi-step-ahead prediction of spectrum availability windows.