Adjacent Channel Leakage Ratio (ACLR)

ACLR is the primary pass/fail metric for wireless transmitter certification. Standards bodies like 3GPP and FCC mandate specific limits:

• 3GPP TS 38.104 for 5G NR: -45 dBc ACLR for wide area base stations
• LTE (3GPP TS 36.104): -44.2 dBc for adjacent channel, -55 dBc for alternate channel
• Wi-Fi (IEEE 802.11): Typically -25 to -30 dBc depending on spectral mask Failure to meet ACLR limits results in regulatory non-compliance and inability to deploy commercially.

ACLR degradation is caused by spectral regrowth—the widening of a signal's bandwidth as it passes through a nonlinear power amplifier. Key contributors:

• AM-AM distortion: Amplitude-dependent gain compression that creates intermodulation products
• AM-PM distortion: Amplitude-dependent phase shift that further spreads the spectrum
• Memory effects: Thermal and electrical memory in the PA that creates asymmetric spectral regrowth Digital predistortion (DPD) directly targets these nonlinearities to restore ACLR to compliant levels.

ACLR is measured using a spectrum analyzer with specific channel power integration bandwidths:

• In-channel power: Integrated over the assigned channel bandwidth (e.g., 20 MHz for LTE)
• Adjacent channel power: Integrated over an equal bandwidth offset by the channel spacing
• ACLR (dBc) = 10 × log₁₀(P_adjacent / P_in-channel) Modern vector signal analyzers automate this measurement with gated sweeps to capture only active transmission periods, avoiding bias from idle intervals.

ACLR and Error Vector Magnitude (EVM) are complementary distortion metrics that trade off against each other:

• ACLR measures out-of-band distortion: Spectral leakage affecting other users and systems
• EVM measures in-band distortion: Modulation accuracy affecting the intended receiver's bit error rate Aggressive DPD linearization improves ACLR but can degrade EVM if the predistorter over-compensates and introduces in-band artifacts. Joint optimization of both metrics is essential for robust transmitter design.

Digital predistortion systems are evaluated by their ACLR improvement capability:

• Typical improvement: 10–20 dB of ACLR enhancement over the uncorrected PA
• Residual ACLR: The final ACLR after DPD, typically targeting 3–5 dB margin below the regulatory limit
• Wideband DPD challenge: For 5G signals with 100+ MHz bandwidth, maintaining ACLR across multiple adjacent and alternate channels simultaneously requires high-order predistorter models ACLR improvement is the definitive key performance indicator (KPI) for any DPD system.

In Massive MIMO and carrier aggregation systems, ACLR requirements become more complex:

• Cross-channel leakage: Distortion from one transmit chain can leak into adjacent antenna elements' assigned channels
• Concurrent multi-band DPD: Must linearize multiple carriers simultaneously while suppressing inter-band intermodulation products
• Beamforming-aware ACLR: The effective ACLR in a specific spatial direction may differ from conducted measurements, requiring over-the-air (OTA) testing for 5G mmWave arrays These multi-dimensional challenges drive the need for advanced neural network-based DPD architectures.

A metric quantifying the deviation of a digitally modulated signal's constellation points from their ideal locations. While ACLR measures out-of-band distortion, EVM measures in-band distortion. Both are caused by the same PA nonlinearity, but they are distinct regulatory requirements. A predistorter optimized solely for ACLR may inadvertently degrade EVM, necessitating a joint optimization strategy.

The ratio of the instantaneous peak power to the average power of a signal. High-PAPR waveforms like OFDM force the PA to operate with significant back-off to avoid clipping, which generates severe spectral regrowth and degrades ACLR. Crest Factor Reduction (CFR) is a complementary technique to DPD that deliberately reduces PAPR before the signal reaches the PA, easing the linearization burden and improving achievable ACLR.

An alternative figure of merit for nonlinear distortion, measured by injecting a notched broadband noise signal into the PA and observing the depth of the notch at the output. NPR provides a measure of distortion floor across the entire band, unlike ACLR which focuses on specific adjacent channels. It is particularly useful for characterizing distortion in multi-carrier or cable applications where the signal spectrum is densely populated.

In DPD training, ACLR cannot be directly used as a differentiable cost function for gradient descent. Instead, algorithms minimize a time-domain error like Mean Squared Error (MSE) between the desired linear output and the PA output. The implicit assumption is that minimizing MSE will also minimize ACLR. Advanced techniques explore frequency-weighted cost functions that penalize errors in adjacent channels more heavily to directly target ACLR improvement.