Generalized Likelihood Ratio Test (GLRT)

The GLRT operates by a two-step procedure. First, it computes the Maximum Likelihood Estimate (MLE) of the unknown parameters under each hypothesis. Second, it plugs these estimates back into the likelihood function as if they were the true values. The test statistic is the ratio of the maximized likelihoods:

• H₀ (Null): Maximize likelihood over unknown parameters assuming H₀ is true.
• H₁ (Alternative): Maximize likelihood over unknown parameters assuming H₁ is true.
• Decision Rule: Reject H₀ if the ratio of maximized likelihoods exceeds a threshold.

This avoids the computationally intensive integration required by the Average Likelihood Ratio Test (ALRT).

While the GLRT is not uniformly most powerful for finite samples, it possesses strong asymptotic properties. As the number of independent observations grows large, the GLRT performance converges to optimality. The distribution of the log-GLRT statistic under the null hypothesis asymptotically follows a chi-squared (χ²) distribution:

• Degrees of Freedom: Equal to the difference in the number of unknown parameters estimated under H₁ versus H₀.
• Threshold Setting: This asymptotic distribution allows for analytical threshold setting to achieve a desired Constant False Alarm Rate (CFAR) without Monte Carlo simulations.
• Consistency: The GLRT is a consistent estimator, meaning the probability of error approaches zero as the sample size approaches infinity.

A primary strength of the GLRT in modulation classification is its ability to handle nuisance parameters—unknown variables like carrier phase offset, timing error, or channel gain that are not of direct interest but affect the likelihood function. The GLRT framework:

• Joint Maximization: Estimates both the modulation-dependent parameters and the nuisance parameters simultaneously via MLE.
• Concentrated Likelihood: Often, the likelihood can be concentrated by analytically maximizing over nuisance parameters first, reducing the dimensionality of the numerical search.
• Practical Example: In classifying between BPSK and QPSK, the GLRT jointly estimates the unknown carrier phase and the modulation type, rather than requiring a separate phase-locked loop.

The GLRT and Average Likelihood Ratio Test (ALRT) represent a fundamental trade-off in composite hypothesis testing:

• ALRT Approach: Treats unknown parameters as random variables with known prior distributions. Integrates the likelihood over the parameter space. Computationally intensive but Bayes-optimal for the assumed prior.
• GLRT Approach: Treats unknown parameters as deterministic but unknown. Maximizes the likelihood. Computationally simpler but lacks a principled way to incorporate prior knowledge.
• Selection Heuristic: Use ALRT when reliable prior distributions exist and computational budget permits. Use GLRT when parameters are truly deterministic, priors are unknown, or real-time constraints demand a simpler estimator.

In Automatic Modulation Classification (AMC) for cognitive radio, the receiver often has no prior knowledge of the transmitter's parameters. The GLRT is a natural fit for this blind scenario:

• Unknown Signal Parameters: Carrier frequency offset, symbol timing, phase jitter, and channel coefficients are all unknown and must be estimated.
• Composite Hypotheses: Each modulation candidate (e.g., 16-QAM, 64-QAM) forms a composite hypothesis because the likelihood depends on these unknown continuous parameters.
• Practical Implementation: A bank of GLRT processors runs in parallel, each computing the maximized likelihood for a candidate modulation. The classifier selects the modulation with the highest maximized likelihood value.
• Complexity Note: The primary computational burden is the multi-dimensional optimization required to compute the MLEs for each hypothesis.

The GLRT framework is deeply connected to model selection criteria like the Akaike Information Criterion (AIC) and Minimum Description Length (MDL). These criteria extend the GLRT by adding penalty terms for model complexity:

• GLRT Core: The log-likelihood ratio forms the data-fit term in both AIC and MDL.
• AIC Penalty: Adds a penalty of 2k, where k is the number of free parameters, to prevent overfitting.
• MDL Penalty: Adds a penalty of k*log(N)/2, where N is the number of observations, favoring simpler models more strongly as data increases.
• Unified View: The GLRT, AIC, and MDL can all be seen as threshold tests on the log-likelihood ratio, differing only in how the threshold is determined—fixed by desired false alarm rate (GLRT), or derived from information-theoretic principles (AIC/MDL).

The statistical engine inside the GLRT. Before comparing hypotheses, the GLRT first computes the Maximum Likelihood Estimate for all unknown deterministic parameters—such as carrier phase, timing offset, or channel coefficients—under each modulation candidate. This involves finding the parameter values that maximize the probability of observing the received signal. The resulting estimates are then plugged directly into the likelihood function, treating them as if they were the true values. This 'plug-in' approach is what makes the GLRT sub-optimal compared to the Average Likelihood Ratio Test (ALRT), as it does not account for estimation uncertainty.

The formal statistical framework that necessitates the GLRT. A hypothesis is composite when it contains one or more unknown parameters. In modulation classification, the hypothesis 'the signal is QPSK' is composite because the carrier phase, frequency offset, and symbol timing are unknown. Simple hypothesis tests like the basic Likelihood Ratio Test cannot handle this uncertainty. The GLRT addresses this by reducing a composite hypothesis to a simple one through parameter estimation, enabling a computable decision statistic.

The process of estimating unknown variables that are not of primary interest but must be resolved to evaluate the likelihood function. In a modulation classifier, the modulation type is the parameter of interest, while carrier phase, frequency offset, and channel gain are nuisance parameters. The GLRT treats all unknown parameters identically, estimating them via MLE regardless of their role. This is a key distinction from the Hybrid Likelihood Ratio Test (HLRT), which may average over some nuisance parameters while estimating others.

The computational backbone of GLRT implementation. The test statistic is formed by comparing the maximized log-likelihoods under each hypothesis. Working in the log domain transforms products of probability densities into sums, providing numerical stability and analytical tractability. The GLRT decision rule is: choose the modulation hypothesis that yields the largest maximized log-likelihood. This avoids the underflow issues common when multiplying many small probability values in the linear domain.

The Bayesian counterpart to the GLRT. Instead of estimating unknown parameters, the ALRT treats them as random variables with known prior distributions and integrates the likelihood function over these parameters. This marginalization accounts for estimation uncertainty, making the ALRT theoretically optimal when the priors are correct. However, the required multi-dimensional integration is often computationally intractable, motivating the use of the simpler GLRT as a practical alternative.

A fundamental performance benchmark for the parameter estimation step within the GLRT. The CRLB provides a lower bound on the variance of any unbiased estimator, expressed as the inverse of the Fisher Information Matrix (FIM). For a GLRT-based classifier, the accuracy of the MLEs directly impacts classification performance. The CRLB allows engineers to analytically determine the minimum achievable estimation error for nuisance parameters like carrier phase, given the signal-to-noise ratio and observation length.