Adjacent Channel Error Power Ratio

ACEPR is defined as the ratio of the error signal power in the adjacent channel to the reference signal power in the main channel. Unlike ACLR, which measures the total output power, ACEPR measures the power of the modeling residual (the difference between the measured PA output and the model's predicted output) in the adjacent bands. This directly quantifies the model's failure to predict spectral regrowth.

While Adjacent Channel Power Ratio (ACLR) is a regulatory metric measuring total spectral leakage, ACEPR is a model validation metric. A model can achieve excellent in-band NMDS while having a poor ACEPR, indicating it fails to capture the nonlinear dynamics responsible for out-of-band emissions. ACEPR specifically answers: 'How much of the distortion is my model failing to explain?'

ACEPR is calculated by:

• Step 1: Compute the error signal e(n) = y_measured(n) - y_model(n)
• Step 2: Apply a bandpass filter to e(n) centered on the adjacent channel
• Step 3: Compute the power of the filtered error signal
• Step 4: Normalize by the total power of the reference signal in the main channel

This is typically expressed in dBc (decibels relative to carrier).

ACEPR is particularly sensitive to long-term memory effects in power amplifiers. Thermal and trapping phenomena create slow-varying distortion components that are difficult to model. A model with insufficient memory depth will exhibit a degraded ACEPR even if its in-band prediction accuracy is high. This makes ACEPR a critical metric for validating Volterra series and memory polynomial models.

In Digital Pre-Distortion (DPD) design, ACEPR serves as a gating criterion for model selection:

• A low ACEPR confirms the behavioral model accurately captures the PA's nonlinear dynamics
• A high ACEPR indicates the model will fail to generate effective predistortion for adjacent channel correction
• Engineers use ACEPR to compare neural network models against classical memory polynomial structures for out-of-band prediction fidelity

ACEPR provides a robust guard against overfitting. A model that memorizes training data noise may show excellent in-band NMDS but will fail to generalize to the adjacent channel dynamics. By evaluating prediction error specifically in the adjacent bands, ACEPR exposes models that have not learned the true underlying nonlinear transfer function. Cross-validation with ACEPR ensures the extracted model captures physical distortion mechanisms rather than measurement artifacts.

The foundational regulatory metric that ACEPR extends. ACPR measures the ratio of power leaked into an adjacent frequency channel to the power in the main channel. While ACPR quantifies the absolute spectral regrowth of a physical amplifier, ACEPR specifically measures the prediction error power in those adjacent channels, isolating model inaccuracy from hardware performance. ACPR is typically expressed in dBc and is critical for spectral mask compliance in standards like 3GPP and IEEE 802.11.

A time-domain metric that quantifies the average power of the error signal normalized by the power of the reference signal. NMSE provides an aggregate measure of model fidelity but does not distinguish between in-band and out-of-band errors. ACEPR complements NMSE by specifically isolating prediction accuracy in adjacent channels, where nonlinear distortion products are most critical. An excellent NMSE can coexist with poor ACEPR if the model fails to capture spectral regrowth dynamics.

An in-band signal quality metric measuring the magnitude of the vector difference between the ideal reference signal and the actual transmitted signal at symbol times. EVM focuses on constellation fidelity within the occupied bandwidth, while ACEPR addresses out-of-band prediction accuracy. Together, they provide a complete picture: EVM for modulation quality and ACEPR for spectral containment. Both are essential for validating digital predistortion performance in modern communication systems.

The physical phenomenon that ACEPR quantifies in the modeling domain. Spectral regrowth is the appearance of unwanted frequency components in adjacent channels caused by intermodulation distortion when a band-limited signal passes through a nonlinear power amplifier. Key characteristics include:

• Third-order intermodulation products are typically the dominant contributors
• Regrowth bandwidth is approximately 3-5x the signal bandwidth for strong nonlinearities
• Thermal memory effects cause asymmetric regrowth that simpler models fail to capture

A statistical method essential for ensuring ACEPR measurements reflect true model generalization rather than overfitting. The process involves:

• Partitioning measured data into training and independent test sets
• Computing ACEPR on unseen test signals that differ from training waveforms
• Validating performance across multiple power levels and signal types

Without proper cross-validation, a model may achieve excellent ACEPR on training data while failing catastrophically on new signals, a critical concern for deployed DPD systems.

A modeling failure where the extracted model memorizes training data noise instead of learning underlying amplifier dynamics. Overfitting manifests as:

• Excellent NMSE and ACEPR on training signals
• Dramatically degraded ACEPR on test signals with different statistics
• Excessive coefficient count relative to the system's true degrees of freedom

Regularization techniques like ridge regression and pruning are essential countermeasures that directly improve ACEPR generalization by constraining model complexity.