Power Amplifier Memory Effect

Caused by impedance variations at baseband and harmonic frequencies due to non-ideal bias networks and decoupling capacitors. These time-varying impedances modulate the PA's instantaneous transfer characteristic.

• Mechanism: Envelope-frequency-dependent load modulation
• Time Constant: Nanoseconds to microseconds
• Primary Source: Bias circuit inductance and capacitance
• Observable As: Asymmetric intermodulation distortion sidebands

Originate from self-heating dynamics within the transistor junction. As instantaneous power dissipation varies with the signal envelope, the junction temperature fluctuates, altering electron mobility and threshold voltage.

• Mechanism: Dynamic junction temperature variation
• Time Constant: Microseconds to milliseconds
• Primary Source: Thermal resistance between junction and heatsink
• Observable As: Slow gain compression and phase shift drift over a burst

Unlike static non-linearity, memory effects cause the AM/AM (gain) and AM/PM (phase) conversion curves to become multi-valued loops rather than single-valued functions. The output for a given instantaneous input amplitude depends on whether the envelope is rising or falling.

• Static Case: Single-valued transfer curve
• Memory Case: Hysteresis loop opening proportional to memory depth
• Fingerprinting Value: Loop shape and width are device-unique

The Volterra series is the canonical mathematical framework for modeling PA memory effects. It extends the Taylor series by adding convolutional kernels that capture the influence of past inputs on the current output.

• 1st Kernel: Linear memory (frequency response)
• 3rd Kernel: Dominant non-linear memory term
• Kernel Asymmetry: Reveals physical origin of memory (electrical vs. thermal)
• Fingerprint Extraction: Kernel coefficients serve as compact device signatures

The effective memory time constant is not fixed but varies with the instantaneous envelope power. High-power states drive faster thermal transients, while low-power states allow cooling, creating a signal-dependent memory profile.

• High PAPR Signals: Excite wider range of memory dynamics
• OFDM Waveforms: Particularly effective at revealing memory signatures
• Fingerprint Richness: Correlated with signal crest factor and bandwidth

A critical distinguishing feature: memory effects often produce asymmetric spectral regrowth around the carrier, where the upper and lower intermodulation sidebands exhibit different power levels and phase relationships.

• Symmetric Regrowth: Indicates memoryless non-linearity
• Asymmetric Regrowth: Definitive evidence of memory effects
• Asymmetry Pattern: Unique to each PA's bias network and thermal design
• Robust Feature: Survives channel fading and moderate noise

The foundational static non-linearity metrics that memory effect compounds. AM/AM conversion describes the amplitude-dependent gain compression, while AM/PM conversion captures the amplitude-dependent phase shift. In a memoryless system, these are simple curves. With memory effect, they become multi-valued, hysteretic loops where the current distortion depends on the signal envelope's history.

• Static AM/AM: A single-valued curve of output amplitude vs. input amplitude
• Dynamic AM/AM: A spread of output values for the same input, depending on prior states
• Hysteresis: The path-dependent widening of the conversion curves, a direct visual signature of memory

A rigorous mathematical framework for modeling non-linear dynamic systems with memory. Unlike a simple polynomial, the Volterra series captures the interaction between the current input and past inputs through multi-dimensional convolution kernels.

• 1st-order kernel: Standard linear impulse response
• 2nd-order kernel: Captures quadratic non-linear interaction between two time instants
• Truncation: Practical models limit the series to 3rd or 5th order and finite memory depth
• Pruning: Removing insignificant kernel coefficients to reduce computational complexity for digital pre-distortion

A simplified, widely adopted subset of the Volterra series that retains only the diagonal terms of the non-linear kernels. The memory polynomial captures both non-linearity and memory by summing polynomial functions of the current and delayed input samples.

• Structure: A double summation over polynomial order and memory depth
• Coefficients: Learned via least-squares estimation from input-output measurements
• Trade-off: Sacrifices the ability to model cross-terms between different delays for significant computational simplicity
• Application: The dominant behavioral model used in real-time digital pre-distortion systems

An enhanced behavioral model that extends the memory polynomial by including cross-terms between the signal and its lagging envelope. This captures the interaction between the instantaneous signal and the past envelope power, a key mechanism of the electrical memory effect caused by bias network impedance.

• Envelope memory terms: Products of the current signal with delayed envelope powers
• Lead/lag cross-terms: Capture asymmetric memory effects around the main tap
• Accuracy: Significantly outperforms the standard memory polynomial for wideband signals with strong memory
• Complexity: More coefficients than a memory polynomial, but far fewer than a full Volterra series

A slow memory mechanism caused by the self-heating of the transistor junction during high-power operation. As the input envelope fluctuates, the instantaneous power dissipation changes, causing the junction temperature to vary with a time constant of microseconds to milliseconds.

• Mechanism: Temperature-dependent electron mobility and threshold voltage shifts
• Time constant: Typically 1 µs to 1 ms, much slower than electrical memory
• Signature: Low-frequency gain and phase variations that track the average envelope power
• Mitigation: Thermal management through heat sinking and bias circuit compensation

A fast memory mechanism originating from the finite impedance of the bias and matching networks at the envelope frequency. The time-varying envelope current interacts with non-zero bias network impedance, causing a dynamic modulation of the transistor's drain or collector bias voltage.

• Mechanism: Voltage drop across bias inductor resistance at envelope frequencies
• Time constant: Typically 10 ns to 1 µs, comparable to the signal envelope variations
• Signature: Asymmetric intermodulation distortion sidebands in two-tone tests
• Mitigation: Active bias networks and envelope-tracking power supplies with very low output impedance