Hybrid Likelihood Ratio Test (HLRT)

HLRT's core innovation is its partitioned handling of nuisance parameters. Unknowns modeled as random variables with known priors (e.g., phase jitter) are averaged out via integration. Deterministic unknowns (e.g., a fixed timing offset) are replaced by their maximum likelihood estimates (MLEs). This hybrid approach avoids the full computational burden of ALRT's multi-dimensional integration while maintaining higher accuracy than GLRT for parameters with well-characterized statistical behavior.

HLRT occupies a critical middle ground in the likelihood-based classifier hierarchy:

• Versus ALRT: HLRT dramatically reduces computational complexity by replacing high-dimensional integration over deterministic parameters with point estimation, at the cost of some optimality.
• Versus GLRT: HLRT improves classification accuracy by correctly modeling random parameters instead of treating them as unknown constants, which is especially beneficial at low SNR.
• Result: A tunable architecture where the designer decides which parameters justify the computational expense of full Bayesian treatment.

For a received signal r under hypothesis H_i, the HLRT statistic is constructed by:

1. Partitioning the unknown parameter vector Θ_i into random Θ_R and deterministic Θ_D components.
2. Averaging the likelihood function over the prior distribution p(Θ_R): L_avg(r | H_i, Θ_D) = ∫ p(r | H_i, Θ_R, Θ_D) p(Θ_R) dΘ_R.
3. Maximizing the averaged likelihood over Θ_D: Λ_HLRT(r) = max_{Θ_D} L_avg(r | H_i, Θ_D). This structure reduces to ALRT when Θ_D is empty and to GLRT when Θ_R is empty.

A canonical HLRT use case in automatic modulation classification involves unknown carrier phase and symbol timing:

• Phase (Random): Modeled as uniformly distributed over [0, 2π). The likelihood is averaged by integrating a modified Bessel function I_0(·), a tractable operation.
• Timing Offset (Deterministic): Estimated via an MLE by correlating the received signal with a reconstructed template at various delays. This specific HLRT configuration significantly outperforms a pure GLRT that treats both parameters as deterministic, particularly for higher-order QAM constellations.

HLRT is a direct solution to the composite hypothesis testing problem where the signal model contains both stochastic and non-stochastic unknowns. It avoids the philosophical commitment of a fully Bayesian framework (which requires priors for all parameters) while being more statistically principled than the purely frequentist GLRT. The test is closely related to Bayes Risk Minimization when misclassification costs are symmetric and to the Neyman-Pearson Criterion when a constant false alarm rate constraint is applied.

Efficient HLRT implementation requires careful numerical methods:

• Numerical Integration: Gaussian quadrature or trapezoidal rules for low-dimensional random parameter spaces. Monte Carlo integration for higher dimensions.
• MLE Optimization: Gradient ascent or grid search over the deterministic parameter space. The Expectation-Maximization (EM) algorithm is often used when the deterministic parameters depend on hidden variables.
• Log-Domain Processing: All operations are performed in the log-likelihood domain to prevent numerical underflow when multiplying many small probability values.

The Bayesian counterpart to HLRT that treats all unknown parameters as random variables with known prior distributions. ALRT forms its test statistic by integrating the likelihood function over the complete parameter space weighted by these priors. While theoretically optimal under perfect prior knowledge, ALRT suffers from computationally intractable multi-dimensional integrals for complex modulation sets. HLRT emerges as a practical compromise when priors are available for some parameters (e.g., phase) but not others (e.g., timing offset).

The frequentist alternative that treats all unknown parameters as deterministic and replaces them with their maximum likelihood estimates before forming the ratio. GLRT avoids integration entirely but suffers from performance degradation at low SNR due to estimation errors propagating into the decision statistic. HLRT bridges ALRT and GLRT by averaging over nuisance parameters with known distributions while estimating those without, achieving near-optimal performance with manageable complexity.

The core design principle behind HLRT involves partitioning unknown parameters into two distinct categories:

• Random nuisance parameters (e.g., carrier phase under a uniform prior): averaged out via integration
• Deterministic nuisance parameters (e.g., symbol timing, amplitude scaling): estimated via maximum likelihood This decomposition directly determines the computational complexity vs. classification accuracy trade-off and must be tailored to the specific channel model and modulation candidate set.

HLRT operates within the formal structure of composite hypothesis testing, where each modulation candidate represents a family of distributions parameterized by unknown variables. The test statistic is constructed as:

• Numerator: likelihood averaged over random parameters and maximized over deterministic parameters under the signal hypothesis
• Denominator: likelihood under the noise-only null hypothesis This structure enables rigorous performance analysis via asymptotic detection theory and derivation of analytical bounds.

For HLRT implementations where the expectation over random parameters lacks a closed-form solution, the Expectation-Maximization (EM) algorithm provides an iterative computational framework. The E-step computes the expected log-likelihood with respect to the random parameters' posterior distribution, while the M-step updates deterministic parameter estimates. This synergy allows HLRT to handle complex hierarchical models with latent variables that would otherwise be analytically intractable.

The theoretical performance of HLRT is characterized by:

• Asymptotic optimality: approaches the ALRT bound as the number of observations grows large
• Cramér-Rao Lower Bound (CRLB): provides a fundamental variance floor for the estimated deterministic parameters
• Deflection coefficient: quantifies the separability of modulation hypotheses in the test statistic space These metrics guide practical design choices, including the minimum observation interval required for reliable classification at a target SNR.