Memory Effect

The primary physical mechanism behind memory effect is self-heating in the transistor junction. As input power varies, the junction temperature changes with a characteristic thermal time constant (typically microseconds to milliseconds). This temperature variation modulates the transistor's gain and phase response, creating a history-dependent transfer function. Key aspects:

• Thermal impedance of the die-attach and package creates a low-pass thermal filter
• Junction temperature lags behind instantaneous power dissipation
• Each amplifier's unique thermal path (die size, solder voids, package defects) produces a distinct thermal signature
• The effect is most pronounced with signals having high peak-to-average power ratios like OFDM

Beyond thermal effects, electrical memory arises from energy storage elements in the amplifier circuit:

• Bias network impedance: The DC bias supply cannot respond instantaneously to envelope variations, causing supply voltage modulation that depends on recent signal history
• Gate/base charge trapping: Semiconductor traps in GaN and LDMOS devices capture and release charge with time constants from nanoseconds to seconds
• Matching network energy storage: Reactive components in input/output matching networks store energy, creating a frequency-dependent memory of past signal states
• Envelope frequency dependence: The amplifier's response to envelope frequencies (typically DC to hundreds of MHz) reveals these memory mechanisms

Memory effect is mathematically captured using Volterra series expansions, which generalize the Taylor series to include time-dependent kernels:

• The first-order kernel represents linear memory (frequency response)
• Higher-order kernels capture non-linear interactions between current and past inputs
• The memory polynomial model is a simplified Volterra variant widely used for digital pre-distortion
• Each amplifier's unique Volterra kernel coefficients constitute a high-dimensional fingerprint
• Kernel extraction requires wideband stimulus signals that exercise the amplifier's dynamic range and envelope bandwidth

Memory effect manifests most clearly in the envelope domain—the amplitude and phase modulation of the RF carrier. Key observable signatures include:

• AM-AM hysteresis: The gain compression curve splits into different paths for increasing vs. decreasing envelope amplitude
• AM-PM hysteresis: Phase shift depends on whether the envelope is rising or falling
• Asymmetric intermodulation products: Upper and lower IM3 sidebands have unequal amplitudes, unlike memoryless non-linearity
• Envelope frequency-dependent gain: Gain variation as a function of envelope modulation frequency reveals the thermal and electrical time constants
• These hysteresis patterns are highly individual and persist across temperature and aging

Memory effect provides a rich, multi-dimensional feature space for RF fingerprinting because it captures the dynamic, history-dependent behavior unique to each amplifier:

• Transient response analysis: The amplifier's response to sudden envelope changes reveals its thermal and electrical time constants
• Multi-tone testing: Two-tone and multi-tone stimuli with varying tone spacing probe different memory depths
• Wideband modulation analysis: Modern communication signals (LTE, 5G, WiFi) inherently exercise memory mechanisms
• Deep learning extraction: Neural networks can learn to isolate memory-induced distortion from channel effects
• Memory-based fingerprints are resistant to cloning because they depend on physical construction, not just static non-linearity

Understanding the difference between memoryless and memory-based distortion is critical for fingerprinting:

• Memoryless (static) non-linearity: Output depends only on instantaneous input—described by AM-AM/AM-PM curves without hysteresis. Caused by instantaneous transistor transfer characteristics
• Memory effect (dynamic) non-linearity: Output depends on current AND past inputs—produces hysteresis, asymmetric spectra, and envelope-frequency-dependent behavior
• A purely memoryless model cannot reproduce the asymmetric intermodulation products observed in real amplifiers
• The ratio of memory to memoryless distortion varies between amplifier classes (Class A shows less memory than Class AB)
• GaN devices typically exhibit stronger memory effects than LDMOS due to more pronounced charge trapping

The overarching category of distortion that includes memory effect. When a PA operates near saturation, its output is no longer a linear function of its input. This generates harmonic distortion and intermodulation products. Memory effect is the history-dependent component of this non-linearity, distinguishing it from static or instantaneous non-linear behavior. Understanding the full non-linear transfer function is essential for isolating the memory component.

These are the two fundamental mechanisms of PA non-linearity. AM-AM distortion describes the non-linear relationship between input amplitude and output amplitude (gain compression). AM-PM distortion describes the unintended phase shift that varies with input amplitude. Memory effect manifests as a dynamic variation in these AM-AM and AM-PM curves, meaning the distortion at any instant depends on the signal's prior amplitude levels, not just the current one.

The primary mitigation technique for PA non-linearity, including memory effect. A DPD system applies an inverse model of the PA's distortion to the input signal. To correct memory effect, the DPD model must incorporate memory polynomials or Volterra series that account for past signal states. The accuracy of the DPD model directly determines how effectively the unique, history-dependent distortion signature of a specific amplifier can be suppressed.

A primary physical cause of memory effect. Transistors in a PA heat up and cool down at rates defined by their thermal mass and heat-sinking. These thermal time constants (often in the microsecond to millisecond range) cause the junction temperature—and thus the transistor's gain and phase characteristics—to depend on the recent signal envelope history. A high-power burst will cause a slow thermal transient that distorts subsequent low-power symbols.

A powerful mathematical framework for modeling systems with memory. Unlike a simple Taylor series that captures static non-linearity, a Volterra series uses multi-dimensional convolution kernels to represent how a system's output depends on past inputs. It is the theoretical foundation for many advanced DPD algorithms and provides a precise way to quantify the unique memory effect signature of an individual amplifier for RF fingerprinting purposes.

The ultimate goal of analyzing impairments like memory effect. The specific thermal time constants, bias circuit dynamics, and trapping effects in a given PA are a product of its unique process-voltage-temperature (PVT) variations from manufacturing. This means the precise memory effect pattern is an unclonable, physical-layer identifier that can distinguish two otherwise identical radios, forming the basis for physical layer authentication.