Memory Effect

The most common physical origin of memory effect, caused by the self-heating of the transistor junction during high-power transmission bursts.

• As the die temperature fluctuates with the signal envelope, the transistor's gain and phase response shift dynamically.
• The thermal time constant (typically microseconds to milliseconds) creates a low-frequency dispersion that modulates the instantaneous transfer function.
• This results in an asymmetric intermodulation distortion spectrum, where the upper and lower sidebands are no longer equal in magnitude.

Originates from bias network impedance and trapping effects in semiconductor materials, independent of thermal changes.

• Bias circuits with non-zero impedance at the modulation frequency cause the drain/collector voltage to sag under high current draw, modulating gain.
• In GaN and GaAs devices, surface traps and buffer traps capture and release charge with time constants ranging from nanoseconds to seconds.
• This creates a history-dependent charge state that directly influences the transistor's pinch-off voltage and transconductance.

Memory effect manifests as a frequency-dependent AM-AM and AM-PM distortion that varies with the modulation bandwidth.

• Static non-linearity can be fully characterized by a single-tone power sweep, but memory effect requires two-tone or modulated stimulus to reveal.
• The magnitude of memory effect increases with signal bandwidth, making wideband signals (e.g., 5G NR, WiFi 6) far more revealing of a device's unique history-dependent signature.
• This bandwidth sensitivity provides a tunable parameter for fingerprint extraction: wider bandwidths expose more device-specific temporal behavior.

The mathematical framework for capturing memory effect is the Volterra series, which extends the Taylor series to include convolutional kernels.

• A first-order Volterra kernel captures linear memory (frequency response), while higher-order kernels model non-linear interactions with delayed signal samples.
• The generalized memory polynomial is a pruned Volterra model widely used in digital pre-distortion (DPD) that directly parameterizes the memory depth and non-linearity order.
• The extracted kernel coefficients serve as a compact, interpretable feature vector for device fingerprinting.

Memory effects are categorized by their time scale relative to the modulation period.

• Short-term memory: Effects with time constants comparable to the RF carrier period or symbol duration, primarily caused by reactive matching network components and carrier trapping.
• Long-term memory: Effects spanning multiple symbol periods, dominated by thermal dynamics and bias circuit charging/discharging.
• Long-term memory creates hysteresis-like patterns in the AM-AM/AM-PM curves when plotted with a modulated signal, forming distinct loop shapes that are highly device-specific.

Memory effect breaks the symmetry of intermodulation distortion (IMD) products, creating a measurable signature.

• In a memoryless system, the upper and lower third-order IMD products (IM3) have equal amplitude and phase.
• With memory effect, IM3 asymmetry emerges: the sidebands differ in magnitude and phase, and this asymmetry varies with tone spacing.
• Measuring the asymmetry across a sweep of two-tone spacings produces a memory signature profile that is uniquely tied to the amplifier's thermal and electrical time constants.

The broader category of amplitude distortion that depends on signal history or frequency, of which memory effect is a primary component. Unlike static non-linearity, dynamic non-linearity introduces a history-dependent signature that is significantly harder to clone or model without exact physical replication. Key manifestations include:

• Slew-rate limiting in amplifier stages
• Dielectric absorption in capacitors
• Thermal memory from transistor self-heating
• Charge trapping in semiconductor interfaces

The physical mechanism driving electro-thermal memory effects in power amplifiers. As a transistor amplifies a signal, instantaneous power dissipation causes localized self-heating, altering its gain and phase characteristics. The rate of heating and cooling is governed by multiple thermal time constants—from nanosecond-scale junction heating to millisecond-scale package-level dynamics—creating a unique, multi-rate thermal signature that is extremely difficult to spoof.

A semiconductor phenomenon where carriers are captured and released by defects at the gate oxide interface in FETs, creating a slow, signal-dependent modulation of threshold voltage. This electrical memory effect introduces a history-dependent bias shift with time constants ranging from microseconds to seconds. In GaN HEMTs, this is the dominant memory mechanism and a rich source of device-specific fingerprinting features.

A mathematical framework for modeling non-linear dynamic systems with memory. Unlike simple polynomial models, Volterra kernels capture the interaction between non-linearity and time-dependent effects. Key aspects:

• First-order kernel: Linear impulse response
• Higher-order kernels: Non-linear interactions at multiple time lags
• Used to extract memory-specific features from amplifier behavioral models
• Provides a complete representation of a device's dynamic non-linear fingerprint

A linearization technique that applies an inverse model of the power amplifier's non-linear behavior—including memory effects—to the input signal. Modern DPD systems use memory polynomial models or neural networks to compensate for both static non-linearity and memory effect. The DPD coefficient vector itself becomes a unique descriptor of the amplifier's impairments, effectively serving as a compressed fingerprint of the device's dynamic behavior.

A mechanism where the envelope of the RF signal modulates the DC bias point of an amplifier through a non-ideal bias network. At low modulation frequencies, the bias supply cannot maintain a constant voltage, causing the amplifier's operating point to shift with signal amplitude. This creates a low-frequency memory effect with time constants determined by the bias circuit's inductance and capacitance, leaving a distinct, circuit-specific signature in the transmitted waveform.