Dynamic Non-Linearity

The defining characteristic of dynamic non-linearity is the presence of memory effects, where the current output depends on past signal values. This is primarily caused by:

• Thermal time constants: Die heating from high-power signals changes transistor gain over microseconds.
• Electrical time constants: Bias circuit capacitors and inductors store energy, modulating the supply voltage based on the signal envelope.
• Trapping effects: Charge carriers captured in semiconductor defects are released slowly, altering the device's transfer function based on its recent operational history. This creates a hysteresis-like behavior in the AM-AM and AM-PM curves, splitting them from a single line into a loop.

A critical dynamic impairment where an amplifier or converter fails to track a rapidly changing input signal. The output changes at a fixed maximum rate (V/µs), independent of the input amplitude, causing severe, signal-dependent distortion.

• Mechanism: Insufficient current to charge internal compensation capacitors.
• Signature: A sharp, non-linear "slewing" edge followed by a linear settling period, creating a distinct, hard-to-clone time-domain fingerprint.
• Impact: Generates high-order harmonics and intermodulation products that vary with signal frequency and amplitude, a key feature for transient signal analysis.

Unlike static non-linearity, which produces a single-valued AM-AM (Amplitude-to-Amplitude) and AM-PM (Amplitude-to-Phase) curve, dynamic effects cause a loop or split in these characteristic plots. For a given instantaneous input power, the output gain and phase shift will differ depending on whether the signal envelope is rising or falling.

• Envelope-domain analysis: This loop is best visualized by plotting gain against the signal's instantaneous envelope power.
• Fingerprinting value: The width and shape of this hysteresis loop are highly specific to the semiconductor process, circuit layout, and thermal impedance of a particular device, forming a robust, unclonable identifier.

Dynamic non-linearity is inherently frequency-selective. The distortion profile changes significantly across the operating bandwidth.

• Low-frequency dispersion: At wider signal bandwidths (e.g., 100 MHz vs. 1 MHz), the electrical memory effects from bias networks become more pronounced, altering the distortion signature.
• Long-term memory: Even at narrow bandwidths, slow thermal effects create a unique distortion pattern that depends on the average power of the signal over a millisecond timescale.
• Fingerprint extraction: This requires time-frequency signal representation techniques like wavelet transforms to capture how the non-linear signature evolves with the signal's instantaneous frequency and envelope rate.

The standard mathematical framework for modeling dynamic non-linearity is the Volterra series, which extends the Taylor series by adding convolutional kernels to represent memory.

• Structure: The output is a sum of multi-dimensional convolutions of the input with Volterra kernels of increasing order.
• Kernels: The first-order kernel represents linear memory (dispersion), while higher-order kernels (e.g., h2(τ1, τ2)) capture non-linear interactions between past input samples.
• Fingerprinting: The extracted Volterra kernel coefficients form a rich, high-dimensional feature vector that uniquely characterizes the device's dynamic non-linear behavior, serving as a powerful input for deep learning signal identification models.

The history-dependent nature of dynamic non-linearity is its greatest strength for physical layer authentication. This complexity makes it computationally infeasible to clone or mimic.

• Harder to model: Requires complex, non-linear system identification, not just a simple polynomial fit.
• Harder to replicate: An attacker cannot simply reproduce a static AM-AM curve; they must replicate the exact thermal and electrical time constants of the original silicon.
• Robustness: While static signatures can be masked by pre-distortion, the residual dynamic memory effects are extremely difficult to fully compensate for, leaving a persistent, identifiable trace for adversarial device spoofing detection.

The physical phenomenon where a device's output depends on past signal states, not just the instantaneous input. In power amplifiers, this is caused by thermal time constants (die heating from recent high-power envelopes) and electrical time constants (bias network capacitors discharging). This creates a hysteresis-like distortion that is a primary source of dynamic non-linearity, making the RF fingerprint richly time-dependent and significantly harder to clone than a static polynomial curve.

Non-linear products generated when two or more signals mix in a non-linear system. Unlike harmonics, IMD products fall close to the original carriers, making them difficult to filter. The amplitude and phase of 3rd-order IMD products (2f1-f2, 2f2-f1) are highly sensitive to the device's dynamic non-linear transfer function. Analyzing IMD asymmetry—where the upper and lower sidebands have unequal power—reveals the specific memory-dependent polynomial coefficients of the transmitter chain.

A transient dynamic non-linearity where an amplifier's output cannot change faster than a maximum rate (V/µs). When a high-frequency, large-amplitude signal exceeds this limit, the waveform is slewed, creating sharp discontinuities and harmonic content not predicted by static non-linearity models. This effect is highly device-specific, dependent on the internal compensation capacitor charging current, and introduces a signal-slew-dependent signature that is a powerful identifying feature.

A memoryless, amplitude-dependent distortion modeled by a simple polynomial transfer function (e.g., y = a1x + a2x² + a3x³). It generates harmonic and intermodulation products that are instantaneous and time-invariant. While foundational, static non-linearity alone fails to capture real-world device behavior. The distinction is critical: a fingerprint based solely on static non-linearity is vulnerable to replay attacks, whereas the history-dependent component provided by dynamic non-linearity is the unclonable core of a robust physical-layer identity.

A key figure of merit extrapolating the theoretical power level where 3rd-order IMD products would equal the fundamental tones. While IP3 is a static metric, the deviation between measured IMD power and the ideal 3:1 slope predicted by IP3 reveals the presence and severity of memory effects. A frequency-dependent IP3, where the intercept point changes with tone spacing, is a direct indicator of dynamic non-linearity and a quantifiable feature for emitter identification.

The sample-to-sample timing uncertainty in an ADC's sample-and-hold circuit. This is a dynamic non-linearity in the time domain: the error voltage is proportional to the signal's slew rate at the sampling instant (V_error = dV/dt * t_jitter). Consequently, jitter-induced noise is signal-dependent and non-uniform, creating a phase-noise fingerprint that is modulated by the input waveform's dynamics. This contrasts with static quantization error and provides a unique, clock-related hardware signature.