Power Amplifier Non-Linearity

The core mechanisms of non-linearity, describing how the amplifier's gain and phase shift vary as a function of the instantaneous input signal amplitude.

• AM/AM Distortion: Amplitude-dependent gain compression or expansion. As input power increases, the output saturates, creating a unique gain curve.
• AM/PM Distortion: Amplitude-dependent phase shift. The signal's phase is unintentionally modulated by its own amplitude envelope, a highly device-specific artifact.
• These curves are not static; they form a complex, hysteretic trajectory when memory effects are present, creating a multi-dimensional fingerprint.

Non-linear amplification causes spectral regrowth, where signal energy spills into adjacent frequency channels. This out-of-band interference is a direct, measurable consequence of the PA's unique distortion profile.

• The spectral shoulder shape—the specific power roll-off into adjacent channels—is highly repeatable for a given device.
• Asymmetrical regrowth between upper and lower sidebands is a classic indicator of memory effects and a strong distinguishing feature.
• This leakage pattern serves as a passive, non-cooperative fingerprint that can be captured without demodulating the signal.

Memory effects cause the PA's output to depend not just on the current input, but on previous signal states. This creates a dynamic, time-dependent distortion signature.

• Electrical Memory: Bias circuit impedance at the envelope frequency causes supply voltage modulation, leading to long-term memory.
• Thermal Memory: Die heating from high-power symbols changes transistor transconductance, creating a slow, signal-dependent drift.
• In an I/Q constellation, memory effects manifest as trajectory-dependent symbol deviation, where a symbol point's exact location depends on the preceding symbol sequence.

When a multi-tone or complex modulated signal passes through a non-linear PA, new frequency components are generated at the sum and difference of the original frequencies.

• Third-Order Intermodulation (IM3) products fall close to the original carriers and are difficult to filter, making them a persistent and measurable artifact.
• The IM3 phase asymmetry—the phase relationship between the upper and lower IM3 products—is a direct consequence of the PA's complex non-linear transfer function.
• These products create a unique, frequency-domain comb structure that acts as a robust fingerprint, even in the presence of noise.

A powerful mathematical framework for modeling non-linear dynamic systems with memory. The Volterra series decomposes the PA's behavior into a sum of multi-dimensional convolution integrals.

• The Volterra kernels (linear, quadratic, cubic, etc.) capture the PA's response to impulses, fully characterizing its non-linear dynamics.
• The coefficients of a pruned Volterra model serve as a compact, highly discriminative feature vector for a specific amplifier.
• Deep learning models like convolutional neural networks implicitly learn a high-dimensional, non-linear embedding that approximates these Volterra kernels from raw IQ data.

The residual distortion that remains after ideal linear equalization, representing the irreducible error caused by the PA's non-linearity. This distortion floor is a unique hardware signature.

• The EVM surface is not uniform; it varies as a function of instantaneous power and frequency, creating a multi-dimensional fingerprint.
• Phase noise and non-linearity interact, causing a modulation-dependent EVM degradation that is distinct from simple additive white Gaussian noise.
• This floor sets a fundamental limit on the PA's modulation accuracy and provides a stable, long-term identifier that is independent of the transmitted data payload.

The definitive mathematical framework for modeling a non-linear power amplifier with memory effects. Unlike simple polynomial models, the Volterra series captures the dynamic interactions between current and past input signals that generate unique intermodulation products.

• Represents the amplifier as a sum of multi-dimensional convolution integrals
• Captures both AM/AM and AM/PM distortion
• The kernel coefficients of a pruned Volterra model serve as a compact, highly discriminating feature vector for emitter identification

A hardware impairment in direct-conversion transceivers where the in-phase and quadrature signal paths have mismatched gain or phase. This creates an unwanted mirror image of the signal spectrum that is unique to each device.

• Gain imbalance: Amplitude mismatch between I and Q branches
• Phase imbalance: Deviation from the ideal 90-degree offset
• When combined with PA non-linearity, the resulting cross-talk between impairments creates a signature that is exceptionally difficult to clone

Short-term, random frequency fluctuations in a transmitter's local oscillator that manifest as spectral spreading of the carrier. This phase noise is convolved with the modulated signal and further distorted by the non-linear PA.

• Characterized by its single-sideband phase noise profile (dBc/Hz)
• The interaction between phase noise and PA non-linearity creates reciprocal mixing products
• These products form a unique, hardware-dependent spectral regrowth pattern that is highly persistent across temperature ranges

The primary technique for linearizing a power amplifier, which ironically requires an extremely accurate model of the very non-linearity that RF fingerprinting exploits. A neural network-based DPD system learns the inverse transfer function of the PA.

• Uses an indirect learning architecture to adapt in real-time
• The residual distortion after DPD correction still contains identifiable hardware signatures
• The DPD coefficient vector itself can be used as a behavioral fingerprint of the amplifier

A neural network training technique that forces the feature extractor to learn channel-invariant representations of the PA non-linearity fingerprint. A gradient reversal layer is inserted between the feature extractor and a domain classifier.

• The domain classifier tries to identify the channel conditions (e.g., indoor vs. outdoor)
• The feature extractor is trained to maximize the domain classifier's error
• This ensures the learned fingerprint is robust to varying propagation environments and receiver hardware

The gradual change in a transmitter's hardware fingerprint over time due to component degradation. Electromigration, hot carrier injection, and oxide breakdown in the PA transistors slowly alter the non-linear transfer function.

• The drift is typically slow and monotonic, allowing for incremental model updates
• A Siamese network can detect when a stored reference fingerprint has drifted beyond a threshold
• Adaptive models use exponential moving average updates to track the evolving signature without requiring full retraining