Additive White Gaussian Noise (AWGN)

The noise is added to the signal, not multiplied or convolved. This means the received signal r(t) is simply the transmitted signal s(t) plus the noise n(t): r(t) = s(t) + n(t). This linear superposition simplifies mathematical analysis and is a valid model because thermal noise in receiver front-ends sums with the incoming waveform at the antenna. The additive nature distinguishes AWGN from multiplicative impairments like fading, which scale the signal amplitude.

"White" indicates that the noise has a flat power spectral density across all frequencies, analogous to white light containing all visible wavelengths equally. In practice, the power spectral density is N₀/2 watts per hertz, constant from DC to approximately 1000 GHz before quantum effects dominate. This flatness means the noise is uncorrelated in time—any two samples separated by any non-zero interval are statistically independent. The autocorrelation function is a Dirac delta function: R(τ) = (N₀/2) · δ(τ).

The noise amplitude follows a Gaussian (normal) probability density function with zero mean and variance σ² = N₀B, where B is the measurement bandwidth. This arises from the Central Limit Theorem: thermal noise is the superposition of countless independent electron movements in resistive components, each contributing a tiny random voltage. Key properties:

• Zero mean: no DC bias; the noise is equally likely to be positive or negative
• 68-95-99.7 rule: 68.3% of samples fall within ±1σ, 95.5% within ±2σ, 99.7% within ±3σ
• The in-phase (I) and quadrature (Q) components are independent, identically distributed Gaussians

Each noise sample is statistically independent from every other sample. This memoryless property is a direct consequence of the flat spectral density: no frequency is favored, so there is no temporal structure or predictability. In discrete-time simulations, this means each sample can be drawn independently from a Gaussian random number generator without modeling any correlation. This contrasts with colored noise models, where spectral shaping introduces sample-to-sample correlation. Independence simplifies derivation of optimal receiver structures like the matched filter.

AWGN is quantified by the Signal-to-Noise Ratio (SNR), the ratio of signal power to noise power. Two common formulations:

• Eb/N₀: Energy per bit to noise power spectral density ratio, the fundamental metric for digital communications. A BPSK system requires Eb/N₀ ≈ 9.6 dB for a bit error rate of 10⁻⁵
• Es/N₀: Energy per symbol to noise density, related to Eb/N₀ by the bits-per-symbol count
• SNR: The ratio of signal power to noise power within the signal bandwidth In synthetic RF impairment generation, SNR is a critical simulation parameter swept across a range (e.g., -10 dB to +30 dB) to train robust fingerprinting models.

AWGN physically originates from Johnson-Nyquist noise in resistive electronic components. The random thermal agitation of charge carriers in a conductor at temperature T (Kelvin) produces a noise voltage with power P = kTB, where:

• k = 1.38 × 10⁻²³ J/K (Boltzmann's constant)
• T is absolute temperature (typically 290K for room-temperature models)
• B is the bandwidth in Hz At 290K, the noise floor is -174 dBm/Hz. This irreducible physical phenomenon sets the ultimate sensitivity limit for any receiver, making AWGN the universal reference impairment against which all other distortions are compared.

The critical simulation parameter defining the ratio of desired signal power to injected background noise power. In AWGN modeling, SNR is the primary variable controlling the difficulty of emitter identification tasks.

• Definition: SNR = P_signal / P_noise, typically expressed in decibels (dB)
• Training Impact: Models must be trained across a range of SNR values (e.g., -10 dB to +30 dB) to ensure robust performance in both noisy and clean channel conditions
• Feature Masking: At low SNR, subtle hardware impairments like I/Q imbalance become buried in the noise floor, making fingerprint extraction significantly harder
• AWGN Relationship: AWGN sets the theoretical lower bound on signal quality; all other impairments are additive to this baseline

A statistical model for simulating dense multipath environments with no dominant line-of-sight path. When combined with AWGN, Rayleigh fading creates a more realistic channel emulation for training robust fingerprinting models.

• Envelope Distribution: The received signal magnitude follows a Rayleigh probability density function
• Physical Cause: Numerous reflected, diffracted, and scattered signal components arrive at the receiver with random phases, causing constructive and destructive interference
• Deep Fades: Signal power can drop 20-30 dB below the mean, temporarily burying the fingerprint in noise
• AWGN Interaction: The fading multiplies the signal amplitude while AWGN adds to it; the combined effect requires models to learn channel-invariant features

A statistical channel model where a dominant line-of-sight (LOS) component coexists with scattered multipath components. The Rician K-factor quantifies the power ratio between the LOS and scattered paths.

• K-Factor: K = P_LOS / P_scattered; high K (e.g., K=10) means strong LOS, approaching AWGN-only conditions; low K (e.g., K=0) reduces to Rayleigh fading
• Application: Common in rural or open-area deployments, drone-to-ground links, and satellite communications where a clear path exists
• Fingerprinting Impact: A strong LOS component preserves more of the transmitter's unique hardware signature compared to Rayleigh fading, making identification easier
• AWGN Baseline: Even with a perfect LOS path, AWGN remains the fundamental sensitivity limit

A time-domain representation of a multipath channel's effect on a transmitted signal, used as a filter kernel to synthetically impose delay spread and fading on a clean waveform before AWGN addition.

• Structure: A set of complex-valued taps, each with a specific delay, amplitude, and phase shift
• Convolution: The transmitted signal is convolved with the CIR to produce the channel-distorted output; AWGN is then added to model the receiver's thermal noise floor
• Tapped Delay Line (TDL): The discrete-time implementation of a CIR, where each tap represents a resolvable multipath component
• Synthetic Generation: By varying CIR parameters (tap delays, Doppler spectra), engineers create diverse training datasets that force fingerprinting models to learn channel-agnostic features

A holistic metric quantifying the combined effect of all impairments—including AWGN—on modulation accuracy. EVM measures the Euclidean distance between ideal constellation points and actual received symbols.

• Definition: EVM = RMS(error vector magnitude) / RMS(reference signal magnitude), expressed as a percentage or in dB
• AWGN Contribution: Thermal noise directly increases the error vector spread around each constellation point, setting a fundamental EVM floor that no hardware improvement can overcome
• Fingerprinting Relevance: Different transmitters exhibit unique EVM patterns due to their specific hardware impairments; AWGN adds a common-mode degradation that must be statistically separated
• Synthetic Control: In simulation, EVM is deliberately degraded by injecting calibrated AWGN and hardware impairment models to create labeled training data across quality levels

The simulated change in a signal's carrier frequency caused by relative motion between transmitter and receiver. When combined with AWGN, Doppler effects create time-varying channel conditions that challenge fingerprinting models.

• Physical Origin: f_Doppler = (v/c) × f_carrier, where v is relative velocity and c is the speed of light
• Jakes Model: A widely-used statistical model generating a Doppler spectrum with the classic 'U-shaped' power spectral density for isotropic scattering
• Coherence Time: The time over which the channel remains approximately constant; inversely proportional to maximum Doppler shift
• Fingerprinting Challenge: Rapid Doppler variations can mask the slow-varying hardware impairments used for identification, requiring models to separate motion-induced changes from device-intrinsic signatures