Minimum Description Length (MDL)

MDL decomposes the total description length into two components: L(H) + L(D|H). The first term encodes the model hypothesis (e.g., modulation type and its parameters), while the second encodes the observed data given that model. The classifier selects the hypothesis minimizing this sum, inherently trading off model complexity against data fit. For modulation recognition, L(H) penalizes schemes with more parameters (like higher-order QAM), while L(D|H) rewards hypotheses that predict the received IQ samples with high probability.

The MDL principle is grounded in the concept of stochastic complexity, which measures the shortest possible encoding of data relative to a model class. Unlike AIC or BIC, which rely on asymptotic approximations, modern MDL uses normalized maximum likelihood (NML) distributions to compute the exact minimax optimal code length. For modulation classification, this provides a rigorous, non-Bayesian foundation for comparing hypotheses without requiring prior probabilities on signal types.

MDL inherently guards against overfitting by imposing an explicit complexity penalty. A complex modulation model (e.g., 256-QAM with many constellation points) may fit the received samples perfectly but incurs a high L(H) cost. A simpler model (e.g., QPSK) may fit less precisely but requires fewer bits to describe. The MDL criterion selects the model where the marginal improvement in data fit no longer justifies the additional model complexity, yielding parsimonious classification decisions.

MDL is philosophically rooted in Kolmogorov complexity—the length of the shortest computer program that outputs the data. While Kolmogorov complexity is non-computable, MDL provides a practical approximation by restricting the set of allowable descriptions to a specific model class. In modulation classification, this means the 'best' modulation scheme is the one that yields the most compressible representation of the received signal, aligning information theory with statistical inference.

MDL differs from related criteria in its penalty structure:

• AIC (Akaike Information Criterion): Minimizes expected prediction error; penalty = 2k (number of parameters). Tends to overfit with large samples.
• BIC (Bayesian Information Criterion): Approximates the Bayes factor; penalty = k·ln(n). Consistent but assumes a true model exists.
• MDL: Minimizes description length directly; penalty derived from coding theory. Provides the strongest theoretical guarantee for model selection without assuming a 'true' model exists, making it robust for real-world signal environments.

In likelihood-based AMC, MDL serves as a model order selection tool for scenarios where the modulation pool size is unknown or nested. Practical applications include:

• Selecting between PSK vs. QAM families when constellation cardinality is uncertain
• Determining the modulation order (e.g., QPSK vs. 8-PSK vs. 16-PSK) within a nested class
• Joint modulation and channel parameter estimation where the number of unknown parameters varies per hypothesis MDL-based classifiers are particularly valuable in spectrum monitoring where the receiver has no prior knowledge of the transmitter's configuration.