Oscillator Phase Noise

The most direct characterization of phase noise is its single-sideband power spectral density, measured in dBc/Hz at a given offset frequency from the carrier. This metric quantifies the noise power in a 1 Hz bandwidth relative to the carrier power.

• Close-in phase noise (offsets < 10 kHz) is dominated by flicker noise in the oscillator's active devices and frequency multiplier chains.
• Far-out phase noise (offsets > 1 MHz) is typically limited by the thermal noise floor of the amplifier stages.
• The slope of the phase noise profile—often decaying at 30 dB/decade, 20 dB/decade, then flattening—is a distinctive signature of the oscillator's phase-locked loop (PLL) design and loop filter components.

Phase noise manifests in the time domain as random timing jitter—the deviation of signal zero-crossings from their ideal positions. For RF fingerprinting, specific jitter statistics are highly discriminative.

• Period jitter: The cycle-to-cycle variation in the clock period, directly impacting the sampling instant of digital-to-analog converters (DACs).
• Accumulated jitter: The long-term wander of the clock edge over many cycles, which creates a unique phase error trajectory.
• Deterministic jitter (DJ) from power supply ripple or crosstalk creates spurs in the phase noise spectrum at specific offset frequencies, acting as a strong identifying feature.

Leeson's equation provides a foundational model for predicting phase noise from oscillator circuit parameters, revealing which physical components contribute to the fingerprint.

• Phase noise scales inversely with the square of the loaded quality factor (Q) of the resonator. A higher-Q crystal or SAW resonator yields lower close-in phase noise.
• The flicker corner frequency of the sustaining amplifier's active device (e.g., a bipolar transistor or FET) directly shapes the 1/f³ region of the phase noise profile.
• Power supply pushing—the sensitivity of the oscillator frequency to supply voltage variations—converts power rail noise into phase modulation, creating a signature tied to the device's power distribution network.

In the transmitted signal, phase noise acts as a multiplicative distortion in the time domain, equivalent to a convolution of the ideal spectrum with the phase noise profile. This smearing is the direct observable exploited by deep learning classifiers.

• For an OFDM signal, phase noise causes common phase error (CPE) and inter-carrier interference (ICI). The pattern of ICI across subcarriers is a function of the specific phase noise mask.
• The error vector magnitude (EVM) floor set by phase noise is a coarse, integrated metric, but the fine structure of the constellation cloud—its shape and orientation—encodes the detailed phase noise signature.
• This distortion is independent of the data payload, making it a persistent, unmodulated artifact for neural network feature extraction.

While phase noise captures short-term fluctuations, Allan variance characterizes frequency stability over longer observation intervals (milliseconds to hours). The Allan deviation curve reveals distinct noise processes.

• The slope of the Allan deviation plot identifies white phase noise, flicker phase noise, white frequency noise, and random walk frequency noise regimes.
• The Allan variance at tau=1 second is a common specification for precision oscillators, but the entire curve shape is a unique fingerprint of the oscillator's aging and environmental sensitivity.
• For emitter identification, the transition points between noise regimes provide features that are robust to short-term channel variations.

Real oscillators exhibit discrete spurious tones (spurs) in addition to continuous phase noise. These spurs are deterministic and highly specific to the synthesizer architecture.

• Reference spurs appear at the phase detector comparison frequency and its harmonics, caused by charge pump leakage and mismatch in the PLL.
• Fractional-N spurs arise from the delta-sigma modulator's quantization noise folding into the loop bandwidth, creating a unique spur pattern tied to the fractional division ratio.
• Power supply spurs at 50/60 Hz mains frequency or switching regulator frequencies (e.g., 100 kHz to 2 MHz) reveal the device's internal power management design.

The mathematical characterization of non-ideal behaviors in RF components. Phase noise is a primary impairment modeled as a stochastic process, often using a power-law noise model to describe its spectral density. This model breaks noise into components like white phase noise, flicker phase noise, and random walk frequency noise, each dominating at different offsets from the carrier. Accurate modeling is the first step in generating synthetic training data for deep learning classifiers.

The unique, unintentional modulation signature imparted on a waveform by the physical hardware of a specific transmitter. Oscillator phase noise is a cornerstone of this RF DNA, contributing a jitter-like spreading of the constellation points that is distinct to each device. Unlike intentional modulation, this signature cannot be easily cloned, making it a robust identifier for physical layer authentication.

The process of uniquely identifying a radio transmitter by analyzing subtle hardware-specific imperfections. Phase noise is a highly discriminating feature for SEI because it originates from the transmitter's local oscillator—a component with manufacturing variances that are impossible to identically replicate. Deep learning models can extract these phase noise patterns from raw IQ samples to classify emitters.

A signal processing technique that exploits the periodic statistical properties of modulated signals. While phase noise is random, its interaction with the signal's cyclostationary features creates unique, measurable patterns in the cyclic autocorrelation function. These features are particularly resilient to stationary additive white Gaussian noise, making them a robust complement to raw phase noise analysis for emitter classification.

A transfer learning technique used to align feature distributions of RF fingerprints captured under different channel conditions. Phase noise signatures can be masked by multipath fading and Doppler shift. Domain-adversarial neural networks with a gradient reversal layer force the feature extractor to learn channel-invariant representations of the phase noise, ensuring the fingerprint remains reliable across diverse operational environments.

A technique to normalize variations in an RF fingerprint caused by temperature-dependent changes in analog components. Oscillator phase noise is highly sensitive to temperature, with a device's signature potentially drifting by several dB across its operating range. Compensation methods include: