Kullback-Leibler (KL) Divergence

KL divergence is fundamentally directional: D_KL(P || Q) ≠ D_KL(Q || P). In likelihood-based classifiers, this asymmetry matters because the divergence from the true channel output distribution to a hypothesized modulation model penalizes different errors than the reverse direction. The forward KL, D_KL(P_true || Q_model), is mean-seeking and averages over all modes of P, while the reverse KL is mode-seeking and fits a single peak.

A foundational property guaranteeing that D_KL(P || Q) ≥ 0 for all distributions P and Q, with equality if and only if P = Q almost everywhere. This non-negativity makes KL divergence a valid measure of discrepancy, though it is not a true metric. For modulation classifiers, this ensures that the correct constellation model always yields the minimum divergence value when compared against the empirical distribution of received IQ samples.

Minimizing the KL divergence D_KL(P_data || P_model) is mathematically equivalent to maximizing the expected log-likelihood of the data under the model. In constellation classification, selecting the modulation format whose theoretical IQ distribution minimizes KL divergence from the observed sample distribution is equivalent to a maximum likelihood decision rule, providing the theoretical foundation for likelihood-based modulation classifiers.

For independent random variables, the KL divergence between joint distributions factorizes: D_KL(P(x,y) || Q(x,y)) = D_KL(P(x) || Q(x)) + D_KL(P(y) || Q(y)). This property is exploited in MIMO modulation recognition, where the total divergence across independent spatial streams can be decomposed into per-stream contributions, enabling efficient per-antenna classification without exponential complexity in the number of antennas.

KL divergence is invariant under invertible transformations of the random variable. If Y = f(X) where f is bijective, then D_KL(P_X || Q_X) = D_KL(P_Y || Q_Y). For signal processing, this means the divergence between distributions is preserved under operations like IQ rotation, scaling, or non-linear warping, making it a robust measure for comparing constellation geometries after channel equalization.

The KL divergence between the joint distribution P(X,Y) and the product of marginals P(X)P(Y) defines the mutual information I(X;Y). In modulation classification, this links KL divergence to the channel capacity and the maximum rate at which modulation format information can be reliably extracted from noisy observations, providing a theoretical upper bound on classifier performance.

A measure of the average number of bits needed to identify an event drawn from a true distribution P when using a coding scheme optimized for an estimated distribution Q. It is the sum of the Shannon entropy of P and the KL divergence from Q to P. In deep learning, minimizing cross-entropy loss is equivalent to minimizing the KL divergence between the empirical data distribution and the model's predicted distribution, making it the standard objective function for training modulation classifiers.

The fundamental measure of the average uncertainty or information content inherent in a random variable's possible outcomes. For a discrete distribution P, it is calculated as H(P) = -Σ P(x) log P(x). In constellation classification, the entropy of the received IQ sample distribution represents the total information before classification; the goal of a likelihood-based classifier is to reduce this uncertainty by assigning samples to specific modulation states.

A symmetric measure of the amount of information obtained about one random variable by observing another. It quantifies the reduction in uncertainty of the transmitted symbol given the received signal. Maximizing mutual information is the objective of geometric shaping and probabilistic shaping techniques. KL divergence can be used to compute mutual information as the divergence between the joint distribution P(X,Y) and the product of marginals P(X)P(Y).

The logarithm of the ratio of two conditional probabilities: the probability of observing a signal under one modulation hypothesis versus another. In binary hypothesis testing for modulation classification, the LLR is compared against a threshold to make a decision. The expected value of the LLR under the true hypothesis is exactly the KL divergence, linking this practical detection statistic directly to the information-theoretic measure of discriminability between constellation types.

A symmetrized and smoothed version of KL divergence, defined as the average of the KL divergence from each distribution to their mixture. Unlike KL divergence, it is always finite and symmetric, making it useful as a distance metric for comparing constellation models where neither distribution is a true reference. It is bounded between 0 and log(2), providing a normalized similarity score for template matching in blind modulation recognition.

An iterative optimization method for finding maximum likelihood estimates in models with latent variables, such as Gaussian Mixture Models (GMMs) used for blind constellation clustering. The E-step computes the posterior probability of each IQ sample belonging to each constellation point, while the M-step updates the centroid locations. The algorithm implicitly minimizes the KL divergence between the empirical data distribution and the GMM's fitted distribution at each iteration.