Additive White Gaussian Noise (AWGN)

The noise signal n(t) is summed directly with the transmitted signal s(t) at the receiver input, producing r(t) = s(t) + n(t). This linear superposition means the noise is independent of the signal's amplitude, phase, or modulation format. In practical receiver chains, this models the thermal agitation of electrons in the front-end low-noise amplifier (LNA), which is the dominant noise source before significant non-linear processing occurs.

The term 'white' indicates a flat, constant power spectral density across all frequencies of interest, mathematically expressed as N₀/2 watts per hertz. This implies the noise samples are uncorrelated in time. In practice, this holds true over the receiver's bandwidth because thermal noise has a flat spectrum up to approximately 1000 GHz, far exceeding the bandwidth of any practical communication system.

The instantaneous amplitude of the noise follows a zero-mean Gaussian (normal) probability density function. This arises from the Central Limit Theorem: thermal noise results from the aggregate random motion of countless independent electrons. Key implications for classification:

• The in-phase (I) and quadrature (Q) components are independent and identically distributed Gaussians
• The envelope magnitude follows a Rayleigh distribution
• The phase is uniformly distributed over [0, 2π]

Classifier performance is universally characterized by plotting accuracy against E_b/N₀ (energy per bit to noise power spectral density ratio) or E_s/N₀ (energy per symbol). Typical evaluation ranges:

• High SNR (> 20 dB): Nearly perfect classification for most schemes
• Moderate SNR (0–10 dB): The critical region where advanced deep learning models demonstrate superiority over traditional cumulant-based methods
• Low SNR (< -5 dB): Extreme regime where only robust cyclostationary or likelihood-based approaches maintain functionality

AWGN's closed-form probability density function makes it the ideal noise source for synthetic dataset generation. Training pipelines leverage this by:

• Generating clean modulated signals via software-defined radio simulations
• Adding calibrated AWGN at precise SNR levels to create labeled training pairs
• Applying data augmentation by randomizing the noise realization on each training epoch, effectively providing infinite unique training examples from a finite set of clean signals

While foundational, AWGN does not capture all real-world impairments. Engineers must extend the model for robust deployment:

• Multipath fading: Requires adding Rayleigh or Rician fading models
• Impulsive noise: Common in industrial and automotive environments, better modeled by Middleton Class A or Bernoulli-Gaussian distributions
• Phase noise and frequency offset: Introduced by local oscillator imperfections
• Adjacent channel interference: Structured interference from other transmitters, not captured by white noise assumptions

The fundamental metric quantifying signal power relative to noise power, typically expressed in decibels (dB). In AWGN channels, SNR directly determines the theoretical bit error rate and serves as the primary independent variable for evaluating classifier robustness.

• High SNR (>20 dB): Noise is negligible; classification is trivial
• Low SNR (<0 dB): Noise power exceeds signal power; classifier performance degrades rapidly
• SNR estimation is often a prerequisite step before modulation classification can proceed reliably

The probability density function that describes the amplitude distribution of thermal noise in electronic components. The Central Limit Theorem justifies this assumption, as AWGN arises from the sum of countless independent random electron movements.

• Zero mean: Noise has no DC bias component
• Variance σ²: Determines average noise power; equal to N0/2 for AWGN
• Probability density: Follows the bell curve f(x) = (1/σ√2π)exp(-x²/2σ²)
• Key property: Samples are statistically independent and identically distributed (i.i.d.)

The defining characteristic of white noise: a flat, constant power spectral density across all frequencies of interest. This means equal noise power per unit bandwidth, analogous to white light containing all visible wavelengths equally.

• Flat PSD: N0/2 watts per hertz for a two-sided spectrum
• Bandwidth-limited: Real systems only experience AWGN within the receiver's bandwidth
• Autocorrelation: A delta function at zero lag, confirming samples are uncorrelated
• Practical approximation: Valid when noise bandwidth far exceeds signal bandwidth

Claude Shannon's landmark theorem defines the maximum error-free data rate achievable over an AWGN channel: C = B log₂(1 + SNR). This theoretical limit establishes the fundamental trade-off between bandwidth, signal power, and noise.

• Bandwidth B: Wider channels support higher data rates
• Logarithmic relationship: Capacity grows slowly with SNR improvements
• Spectral efficiency: Measured in bits per second per hertz (bps/Hz)
• Practical systems: Modern coding schemes like LDPC and Turbo codes approach within 1 dB of the Shannon limit

AWGN is typically modeled in the complex baseband domain for modulation classification, where the noise becomes circularly symmetric complex Gaussian. This representation captures both I and Q channel impairments simultaneously.

• Complex noise sample: n = n_I + j·n_Q, where both components are independent real Gaussian
• Circular symmetry: Phase rotation does not change statistical properties
• Variance per dimension: σ² = N0/2 for each of I and Q
• Practical benefit: Enables direct simulation of noisy IQ samples for training deep learning classifiers

The additive nature of AWGN means the received signal is simply the transmitted signal plus noise: r(t) = s(t) + n(t). This linear superposition distinguishes AWGN from multiplicative impairments like fading, making it the simplest channel model for initial classifier development.

• No signal-dependent noise: Noise power is independent of signal amplitude
• Mathematical tractability: Enables closed-form optimal receiver design
• Combined impairments: Real channels experience both AWGN and multiplicative fading
• Benchmarking role: AWGN performance establishes the upper bound for classifier accuracy before introducing more complex channel effects