PaperlayerView exact revision on arXiv

Source-linked AI summary

Comparative Analysis and Behavioral Modelling of Power Amplifiers in FR3

Cel Thys, Rodney Martinez Alonso, Ali H Alsarraf, Zhuangzhuang Cui, Dominique Schreurs, Sofie Pollin

arXiv:2610.19889v1eess.SP

TL;DR

PA nonlinearity is a major transmitter distortion source, and the 3GPP baseline has not been validated by FR3 measurements. The paper measures PA behavior in FR3, extracts behavioral models, and evaluates sensing effects; the GMP model improves test NMSE across operating points, while compression raises sensing sidelobes.

  • Problem

    The 3GPP PA baseline was derived from measurements at 2.1 GHz and circuit simulations at 28 GHz, leaving it unvalidated by measurement in FR3, where PA nonlinearity also degrades communication and sensing signals.

    Source: §I · p. 1

  • Method

    The authors characterize FR3 PAs, release complex-baseband measurements, extract memoryless polynomial and GMP behavioral models, and evaluate them in an OFDM ISAC ambiguity-function simulation.

    Source: §I · p. 1

  • Results

    The GMP model improves test NMSE by 1.3 dB to 4.6 dB at every operating point; the GaN PA is best modelled by an 11th-order polynomial and shows pronounced nonlinear phase distortion.

    Source: §IV · p. 6

  • Takeaways & Limitations

    The released measurements and model coefficients provide a reference for incorporating FR3 PA distortion into communication and sensing simulations.

    Source: §I · p. 1

Abstract

from arXiv · show

Power amplifier (PA) nonlinearity is a critical transmitter impairment in the upper mid-band (7 GHz to 24 GHz), designated as Frequency Range 3 (FR3) in emerging 6G standardisation. The 3GPP PA model baseline, TR 38.803 Annex A, was derived from 2.1 GHz and 28 GHz measurements and has not been revisited yet for FR3 standards work. This paper addresses that gap with characterisation measurements of multiple PA devices-under-test (DUTs), driven into gain compression in the 7 GHz and 15 GHz bands with 100 MHz and 200 MHz bandwidth. Memoryless polynomial and Generalized Memory Polynomial (GMP) behavioral models are extracted by least-squares regression and compared with 3GPP models spanning FR1 and FR2. The GaN PA DUT is modelled best at a higher nonlinear order and shows markedly stronger phase distortion than the wideband PA. The memory terms of the GMP model measurably improve the Normalized Mean Square Error (NMSE) at all operating points. An orthogonal frequency-division multiplexing (OFDM) ambiguity function analysis from the Integrated Sensing and Communication (ISAC) literature, using the extracted models, illustrates how PA compression raises the sidelobe floor of the sensing response. These results provide a reference for PA impairment modelling in FR3, directly applicable to ISAC system design. All model coefficients and measurement data are publicly available [1].

I. INTRODUCTION

PA nonlinearity degrades communication signals and raises sensing range sidelobes, while the existing 3GPP baseline has not been validated by FR3 measurements. The paper addresses the measurement gap with PA data, behavioral models, and an OFDM ISAC ambiguity-function evaluation.

I. INTRODUCTION

  • PA nonlinearity is a dominant transmitter distortion source that degrades in-band modulation quality, spreads power out of band, and raises sensing range sidelobes.The sidelobe effect impairs target detection in sensing.Source: §I · p. 1
  • The 3GPP TR 38.803 Annex A baseline derives from 2.1 GHz measurements and 28 GHz circuit simulations, and has not been validated by measurement in FR3.The paper notes that measurement-based behavioral models had not been published for the 7 GHz and 15 GHz bands.Source: §I · p. 1
  • The paper releases complex baseband measurements for a GaN PA at 8 GHz and a wideband PA at 7.225 GHz and 15.0 GHz, with two signal bandwidths.The data supports fitting models or running simulations.Source: §I · p. 1
  • The authors extract memoryless polynomial, MP, and GMP models, compare them with 3GPP baselines, and report the accuracy gain from memory terms.They also publish polynomial coefficients with the dataset.Source: §I · p. 1
  • An OFDM ISAC ambiguity-function simulation evaluates sensing degradation using the extracted PA models.The analysis quantifies the impact of PA nonlinearity on sensing.Source: §I · p. 1

II. METHODOLOGY

The methodology characterizes three PA devices using calibrated wideband waveform generation and acquisition, then captures operating points across FR3 frequencies and bandwidths. Stimulus choices and device configurations support comparisons across operating conditions.

II. METHODOLOGY

  • The methodology covers measurement setup and DUTs, a released dataset and behavioral models, and metrics for model selection and comparison with the 3GPP reference.It also quantifies memory effects and bandwidth dependence and evaluates sensing impact.Source: §II · p. 1

A. Measurement Setup and DUTs

  • The setup uses a VXG to generate wideband waveforms and a calibrated PNA-X for acquisition, with the optional linear driver preceding the DUT where needed.The stimulus drive level refers to modulated-signal RMS power, and its roughly 7 dB PAPR spans instantaneous input powers from the small-signal region to the signal peak.Source: A. Measurement Setup and DUTs · pp. 1–2, Figure 1 · p. 2
  • The study measures a GaN PA at 8 GHz and a wideband PA at 7.225 GHz and 15 GHz, alongside a GaAs LNA comparison point at 15 GHz.The GaN PA is measured at 100 MHz across 11 drive levels; the TLPA covers 200 MHz at 7.225 GHz and 100 and 200 MHz at 15 GHz.Source: A. Measurement Setup and DUTs · p. 2
  • Characterization uses single-carrier 64-QAM at 100 MHz, while additional TLPA tests use 256-QAM at 200 MHz, both with RRC pulse shaping and 12.8× oversampling.The oversampling captures higher-order intermodulation products without risk of aliasing.Source: A. Measurement Setup and DUTs · p. 2

B. Public Dataset

The paper provides public complex-baseband captures and defines a behavioral-modeling approach that extends a memoryless polynomial with delayed terms and envelope cross terms. The MP and GMP structures capture different forms of memory in the PA response.

B. Public Dataset

  • The public dataset contains multiple captures per operating point with complex baseband input and output IQ sequences, time-aligned using cross-correlation and phase rotation.It includes the measured GaN, TLPA, and GaAs LNA operating points, plus four swept-input-power TLPA captures at 15 GHz/100 MHz.Source: B. Public Dataset · p. 2
  • The dataset supports FR3 PA distortion simulations for communication or sensing and has also supported digital predistortion research [14].Source: B. Public Dataset · p. 2

C. PA Behavioral Model

  • The memoryless model is an odd-order polynomial in complex baseband, with coefficients fit by least squares.Its odd-order terms model in-band intermodulation products of the complex envelope.Source: C. PA Behavioral Model · p. 2
  • The GMP structure extends the memoryless model to capture memory, and its coefficients are also fit by least squares.With B = 0 it reduces to MP; MP adds delayed memoryless basis functions, while GMP cross terms combine a past sample with the current envelope.Source: C. PA Behavioral Model · p. 2

D. Model Evaluation

Model quality is evaluated using NMSE, with model selection separated from testing. Normalized AM/AM and AM/PM curves enable comparisons across devices and 3GPP references.

D. Model Evaluation

  • Models are fitted and selected on independent data, and reported NMSE is computed on a separate test capture.The selected model has the fewest coefficients whose validation NMSE is within 0.1 dB of the lowest overall NMSE.Source: D. Model Evaluation · p. 3
  • AM/AM reports gain relative to each device’s small-signal gain, while AM/PM reports phase shift relative to its small-signal phase.For models, the predicted output replaces the measured output in these quantities.Source: D. Model Evaluation · p. 3
  • The 3GPP references comprise four memoryless models derived from 2.1 GHz measurements or 28 GHz circuit simulations, each with a GMP variant.The references include GaAs and GaN devices at 2.1 GHz and CMOS and GaN devices at 28 GHz.Source: D. Model Evaluation · p. 3
  • Curves use input back-off from each device’s P1dB, where every AM/AM curve reaches −1 dB at 0 dB IBO.The measured P1dB is the input power where gain first drops 1 dB below small-signal gain; for 3GPP models, the same criterion is applied to model gain.Source: D. Model Evaluation · p. 3
  • The TLPA and GaAs LNA are driven just beyond P1dB, reaching 1.2–2.1 dB compression.Source: D. Model Evaluation · p. 3

III. MEASUREMENT RESULTS AND ANALYSIS

The results section applies the measurement and modelling method to the measured PAs. It covers the drive-level sweep, 3GPP comparison, memory and bandwidth effects, and sensing impact.

III. MEASUREMENT RESULTS AND ANALYSIS

  • The section applies the preceding method to measured PAs and first examines the GaN PA across its drive levels.Source: §III · p. 3
  • The analyses compare extracted models with 3GPP baselines and quantify memory effects and bandwidth dependence.Source: §III · p. 3
  • The section also evaluates how PA nonlinearity affects sensing.Source: §III · p. 3

A. GaN PA Drive-Level Sweep

The GaN PA’s compression and gain vary across its 11 drive levels, and its selected memoryless model requires high nonlinear order. The figure plots measured gain and phase against driver-input power.

A. GaN PA Drive-Level Sweep

  • Peak compression increases from 1.9 to 5.2 dB across the 11 drive levels, while output P1dB stays at 24.5–25.0 dBm.Small-signal gain drops by 1.4 dB with drive, so the levels do not lie on a single AM/AM curve.Source: A. GaN PA Drive-Level Sweep · p. 3
  • The GaN PA’s soft compression places output P1dB about 11 dB below saturated output, with gain slowly deteriorating between them.The passage notes that an efficient operating point would mean driving this DUT beyond 1 dB compression.Source: A. GaN PA Drive-Level Sweep · p. 3
  • The selected −20 dBm drive-level model uses nonlinear order 11, or six terms, at about 3 dB above input P1dB.This operating point has 3.3 dB peak compression and is selected as a tradeoff between available power and nonlinear distortion.Source: A. GaN PA Drive-Level Sweep · p. 3
  • Fig. 2 plots measured GaN PA gain and phase shift against driver-input power for 11 drive levels at 8 GHz and 100 MHz.Source: Figure 2 · p. 3

B. Comparison with 3GPP Baselines

The measured PA models and 3GPP baselines compress similarly through P1dB, but their behavior diverges beyond it. Phase shifts vary more across devices than gain, with the GaN PA showing strong phase distortion.

B. Comparison with 3GPP Baselines

  • Beyond P1dB, the 28 GHz references compress fastest, while the TLPA and GaN PA follow the 2.1 GHz GaN baseline to about 4 dB beyond P1dB.The measured curves extend 1.9–5.1 dB beyond P1dB for the TLPA and 9.7 dB for the GaN PA; at that GaN range edge, gain reaches −3.5 dB.Source: B. Comparison with 3GPP Baselines · p. 3, Figure 3 · p. 4
  • The GaN PA phase shift reaches +5.2° at P1dB and +16.1° at its measured-range edge, exceeding every 3GPP reference at a larger IBO.Source: B. Comparison with 3GPP Baselines · p. 3
  • Table I overviews extracted memoryless-polynomial and GMP models by polynomial order and number of complex coefficients, alongside four 3GPP references.Fitted coefficients are referenced in Section III-F and [1].Source: Table I · p. 4
  • The TLPA operating points stay within 1° of phase shift, while the same-technology GaAs LNA shows about −4° rather than the 2.1 GHz GaAs baseline’s +14.3°.Source: B. Comparison with 3GPP Baselines · p. 4

C. Extracted PA Models and Memory Effects

The selected polynomial order varies by PA, with the strongly compressed GaN requiring the highest order. Memory terms improve test NMSE, and at equal order the MP captures most of that gain with fewer coefficients than GMP.

C. Extracted PA Models and Memory Effects

  • The strongly compressed GaN PA’s test NMSE improves from −29.4 dB at 3rd order to −39.4 dB at 11th order.The TLPA at 7.225 GHz needs 5th order, while the TLPA at 15 GHz and GaAs LNA do not improve beyond 3rd order.Source: C. Extracted PA Models and Memory… · p. 4
  • GMP improves test NMSE over memoryless models at every operating point, by 1.4 dB for GaN and up to 4.6 dB for TLPA at 15 GHz.For GaN, the gain is 1.3–1.8 dB across all 11 drive levels.Source: C. Extracted PA Models and Memory… · p. 4
  • At equal nonlinear order, MP memory terms improve NMSE by 4.1 dB for TLPA at 15 GHz/200 MHz and 1.0 dB for GaN; GMP cross terms add at most 0.4 dB.MP uses 14–42 coefficients versus 14–64 for GMP; the table compares memoryless, MP, and GMP model coefficient counts and test NMSE.Source: C. Extracted PA Models and Memory… · p. 4, Table II · p. 4

D. Impact of Modulation and Bandwidth

For TLPA at 15 GHz, the compression trend is similar across the measured bandwidths, although the wider-bandwidth measurement has greater spread. In the OFDM ambiguity-function evaluation, PA compression raises the sensing sidelobe floor, while TLPA linear memory adds a close-in shoulder.

D. Impact of Modulation and Bandwidth

  • At 15 GHz, TLPA’s compression trend is the same for 100 MHz and 200 MHz signals, but the spread widens at 200 MHz.The test NMSE of the 3rd-order polynomial rises from −31.8 to −28.1 dB; bandwidth and modulation order change together, so the gap reflects their joint effect.Source: D. Impact of Modulation and Bandwidth · pp. 4–5, Figure 6 · p. 5
  • The ambiguity-function analysis uses a 16-PSK CP-OFDM signal with 2048 subcarriers over 100 MHz and 128 symbols, scaled and clipped within the fitted input range.The RMS level is about 3 dB below that of the measured signal because OFDM has higher PAPR.Source: D. Impact of Modulation and Bandwidth · p. 5
  • In the OFDM ambiguity-function evaluation, GaN compression raises the sidelobe floor from −88 dB under linear amplification to −53 dB; TLPA reaches −65 dB with the memoryless model.The zero-Doppler cut compares linear amplification with memoryless and GMP models for GaN at highest drive and TLPA at 15 GHz/200 MHz.Source: D. Impact of Modulation and Bandwidth · p. 5, Figure 7 · p. 5
  • For TLPA, GMP adds a shoulder within about 15 delay bins of the mainlobe, while the farther-out floor matches the memoryless model.The shoulder is attributed to linear memory; its 0.7 dB and 6° in-band ripple over 200 MHz is deterministic and can be equalised.Source: D. Impact of Modulation and Bandwidth · p. 5

F. Model Coefficients

The paper provides memoryless polynomial coefficients for system-level simulation, with input normalization and operating-point scaling requirements.

F. Model Coefficients

  • The memoryless polynomial coefficients apply to input normalized to unit peak amplitude, |x(n)| ≤1.The coefficients are listed for the measured operating points summarized in Table I.Source: F. Model Coefficients · pp. 5–6
  • For a different waveform, scale its envelope to the intended operating point, specified by desired back-off from P1dB, before applying the coefficients.Source: F. Model Coefficients · p. 6
  • The MP and GMP coefficients are published with the measurement data.Source: F. Model Coefficients · p. 6

IV. CONCLUSION

The study presents measurement-based behavioral models for FR3 PAs and reports that the GaN and wideband devices differ in required polynomial order and phase distortion. Memory terms improve NMSE, while the ambiguity-function evaluation shows compression-associated sidelobe increases; the conclusions are bounded by measurements of single units of three devices.

IV. CONCLUSION

  • The measurements cover a GaN PA at 8 GHz, driven up to 5.2 dB compression over 11 levels, and a wideband PA at 7.225 and 15 GHz with 100 MHz and 200 MHz bandwidth.The paper describes these as the first publicly released FR3 PA dataset.Source: §IV · p. 6
  • A 3rd- to 5th-order memoryless polynomial suffices for the wideband PA, while the GaN PA is best modelled by an 11th-order polynomial.The GaN PA exhibits soft compression and pronounced nonlinear phase distortion, whereas TLPA remains within 1°.Source: §IV · p. 6
  • The extracted models follow the 2.1 GHz GaN baseline of TR 38.803 up to about 4 dB beyond P1dB.Source: §IV · p. 6
  • GMP improves test NMSE at every operating point by 1.3 dB to 4.6 dB; at equal order, MP provides most of the gain with fewer coefficients.Source: §IV · p. 6
  • In the OFDM ISAC ambiguity-function evaluation, compressed GaN raises the sidelobe floor by 35 dB, with the same floor from memoryless and GMP models.GMP additionally shows close-in sidelobes from linear, equalisable memory; memoryless coefficients suffice to predict the floor for link- and system-level simulations.Source: §IV · p. 6
  • The findings describe single units of three devices measured with single-carrier 64- and 256-QAM signals.Further work will investigate GaN PA memory effects on sensing and add measurement data from further FR3 bands.Source: §IV · p. 6
Loading 2610.19889v1…