Additive White Gaussian Noise (AWGN)

The noise process adds linearly to the transmitted signal. The received signal is mathematically expressed as r(t) = s(t) + n(t), where s(t) is the transmitted waveform and n(t) is the noise. This superposition principle implies the noise is independent of the signal's amplitude, phase, or frequency. Unlike multiplicative channel effects like fading, AWGN does not scale or convolve with the signal. This property simplifies the derivation of optimal receiver structures, such as the matched filter, and allows the likelihood function for a given hypothesis to be expressed as a multivariate Gaussian centered on the noise-free signal point.

"White" denotes a flat power spectral density (PSD) across all frequencies, theoretically from negative to positive infinity. In practice, the PSD is constant at N₀/2 Watts/Hz over the bandwidth of interest. This implies that noise samples separated by any non-zero time interval are completely uncorrelated. For a discrete-time model with sampling at the Nyquist rate, this results in independent noise samples. The autocorrelation function is a Dirac delta function, (N₀/2)δ(τ), which is the Fourier transform pair of the flat PSD. This property is critical for deriving the Cramér-Rao Lower Bound for parameter estimation in AWGN.

The noise voltage at any instant follows a zero-mean Gaussian (normal) probability density function. The PDF is given by p(n) = (1/√(2πσ²)) exp(-n²/(2σ²)), where the variance σ² = N₀/2 is the average noise power. This distribution arises from the Central Limit Theorem, as thermal noise is the aggregate effect of countless independent electron movements. The Gaussian assumption is mathematically convenient: it makes the log-likelihood function a simple squared Euclidean distance, and it is the maximum entropy distribution for a given variance, representing the worst-case noise scenario.

The AWGN channel is statistically memoryless. The noise component at any time instant is independent of the noise at any other instant. This means the channel does not introduce intersymbol interference (ISI) on its own. The conditional probability of receiving a sequence of symbols factors into the product of individual symbol probabilities. This property is fundamental to the derivation of the Maximum Likelihood Sequence Estimator (MLSE) and simplifies the analysis of symbol error rate (SER) for various digital modulation schemes like QPSK and 16-QAM.

AWGN serves as the theoretical benchmark for evaluating modulation classifiers. The performance of an Average Likelihood Ratio Test (ALRT) or Generalized Likelihood Ratio Test (GLRT) is first derived under pure AWGN to establish an upper bound on accuracy. The confusion matrix and probability of correct classification are analytically tractable in this environment. Real-world impairments like multipath fading, phase noise, and adjacent channel interference are then introduced as degradations from this ideal baseline, allowing engineers to quantify the robustness of their automatic modulation classification algorithms.

The defining performance parameter in an AWGN channel is the Signal-to-Noise Ratio. It is typically expressed as Eₛ/N₀ (energy per symbol to noise PSD) or E_b/N₀ (energy per bit to noise PSD). The SNR dictates the symbol error rate (SER) and the ultimate channel capacity. For a fixed modulation scheme, the probability of a classification error in a Maximum A Posteriori (MAP) classifier is a monotonically decreasing function of SNR. The relationship between SNR and the Kullback-Leibler divergence between modulation hypotheses quantifies the theoretical discriminability of signal types.

An equivalent lowpass signal model that captures both in-phase (I) and quadrature (Q) components as a single complex-valued signal. This representation is fundamental because it simplifies the mathematical analysis of modulated waveforms, allowing AWGN to be modeled as a circularly symmetric complex Gaussian process. It directly supports the formulation of likelihood functions for coherent and non-coherent classifiers.

A demodulation method that requires accurate recovery of the carrier phase and frequency at the receiver. In an AWGN channel, a coherent detector achieves optimal performance by exploiting full signal knowledge, using a matched filter and sampling at the correct symbol timing. The likelihood function for a coherent classifier is a multivariate Gaussian distribution centered on the hypothesized signal constellation point.

A demodulation approach that operates without carrier phase recovery, treating the phase as a random nuisance parameter. This sacrifices some performance for reduced complexity and robustness to phase noise. In an AWGN channel, the resulting likelihood function is derived by averaging the coherent likelihood over a uniform phase distribution, often leading to a Rician or modified Bessel function form for the test statistic.

The natural logarithm of the likelihood function, used to transform products of conditional probability densities into sums. This is a critical step for numerical stability in practical classifiers, as raw likelihoods for long sequences can underflow floating-point precision. For an AWGN channel, the log-likelihood simplifies to the Euclidean distance between the received signal and the hypothesized noiseless signal, scaled by the noise variance.

A function of the observed data that captures all information relevant to a parameter of interest. For modulation classification in AWGN, the output of a matched filter sampled at the symbol rate is a sufficient statistic for the transmitted symbol sequence. This enables dimensionality reduction without any loss of statistical power, forming the input to optimal classifiers like the Maximum Likelihood Sequence Estimator (MLSE).

The probability that a detected symbol differs from the transmitted symbol. In an AWGN channel, the SER for a given modulation scheme is a deterministic function of the Signal-to-Noise Ratio (SNR) and the constellation geometry. It serves as a fundamental performance metric for evaluating a modulation classifier's accuracy, with the theoretical SER under AWGN providing a lower bound on the classifier's achievable error probability.