Source-linked AI summary
Mixed-ADC Massive MIMO Detectors: Performance Analysis and Design Optimization
Ti-Cao Zhang, Chao-Kai Wen, Shi Jin, Tao Jiang
TL;DR
Massive MIMO’s many high-resolution ADCs raise hardware cost and power, while low-resolution ADCs complicate parameter estimation. The paper develops Bayesian detectors and asymptotic performance analyses for mixed-ADC receivers. Its analysis finds that PQN distortion is negligible at low SNR but may cause 1-bit detection performance loss at high SNR.
Problem
High-resolution ADCs at every massive-MIMO antenna increase hardware cost and power, while pure low-resolution architectures complicate parameter estimation.
Source: §I · p. 1Method
The paper develops a Bayesian detector framework for mixed-ADC uplink MIMO and uses a PQN likelihood model to reduce computational complexity.
Source: §I · pp. 1–2Results
PQN distortion can be ignored at low SNR but may cause 1-bit detection performance loss at high SNR.
Source: §VI · p. 11Takeaways & Limitations
The asymptotic performance expressions can be evaluated quickly and efficiently for receiver design optimization.
Source: §VI · p. 11
Abstract
from arXiv · showhide
Using a very low-resolution analog-to-digital convertor (ADC) unit at each antenna can remarkably reduce the hardware cost and power consumption of a massive multiple-input multiple-output (MIMO) system. However, such a pure low-resolution ADC architecture also complicates parameter estimation problems such as time/frequency synchronization and channel estimation. A mixed-ADC architecture, where most of the antennas are equipped with low-precision ADCs while a few antennas have full-precision ADCs, can solve these issues and actualize the potential of the pure low-resolution ADC architecture. In this paper, we present a unified framework to develop a family of detectors over the massive MIMO uplink system with the mixed-ADC receiver architecture by exploiting probabilistic Bayesian inference. As a basic setup, an optimal detector is developed to provide a minimum mean-squared-error (MMSE) estimate on data symbols. Considering the highly nonlinear steps involved in the quantization process, we also investigate the potential for complexity reduction on the optimal detector by postulating the common \emph{pseudo-quantization noise} (PQN) model. In particular, we provide asymptotic performance expressions including the MSE and bit error rate for the optimal and suboptimal MIMO detectors. The asymptotic performance expressions can be evaluated quickly and efficiently; thus, they are useful in system design optimization. We show that in the low signal-to-noise ratio (SNR) regime, the distortion caused by the PQN model can be ignored, whereas in the high-SNR regime, such distortion may cause 1-bit detection performance loss. The performance gap resulting from the PQN model can be narrowed by a small fraction of high-precision ADCs in the mixed-ADC architecture.
I. INTRODUCTION
The paper studies mixed-ADC massive MIMO detection, where a few full-resolution ADCs accompany mostly low-resolution ADCs. It develops Bayesian detectors and performance analyses for this architecture.
I. INTRODUCTION
- High-resolution ADCs across massive antenna arrays raise hardware cost and power, while low-resolution quantization complicates parameter estimation and receiver design.The paper motivates using a few full-resolution ADCs alongside mostly low-resolution units.Source: §I · p. 1
- A few full-resolution chains can support round-robin channel estimation without significant pilot overhead and assist frontend design, including AGC and synchronization.The proposed architecture equips most antennas with low-resolution ADCs and a few with full-resolution ADCs.Source: §I · pp. 1–2
- The paper addresses mixed-ADC detection with a unified Bayesian framework, where true priors and likelihoods yield MMSE estimates and mismatched measures yield simpler detectors such as LMMSE and MRC..Source: §I · p. 2
- The PQN model treats quantization noise as additive and independent, reducing computational cost while degrading performance only slightly.The model is introduced as a lower-complexity alternative to the exact nonlinear quantization likelihood.Source: §I · p. 2
- GAMP adapts the Bayes detector to mixed ADCs; a central-limit approximation yields state-evolution expressions for predicting MSE and BER and evaluating performance efficiently.Simulations are used to verify the analysis.Source: §I · p. 2
II. SYSTEM DESCRIPTION
- The system model is an uplink with K single-antenna users and an N-antenna base station; the analysis assumes known channel realizations and allows different ADC resolutions across antennas.The received signal is quantized separately in its real and imaginary components.Source: §II · pp. 2–3
III. MIMO DETECTORS
The detector framework forms a posterior from an input measure and a receiver-signal likelihood, then estimates transmitted symbols from that posterior. Choosing mismatched measures can reduce computational complexity.
III. MIMO DETECTORS
- Bayesian inference forms the posterior of transmitted symbols from the receiver-signal likelihood and input measure, which may be mismatched to the true distributions.The likelihood is marginalized over the input distribution.Source: §III · p. 3
- The Bayes estimate for each symbol is its mean under the marginal posterior distribution.The estimate is framed as an estimation problem; for constellation modulation, a maximum-likelihood decision maps it to a constellation point.Source: §III · p. 3
- Using mismatched prior or likelihood measures can reduce computational complexity, and the framework includes detectors such as LMMSE and MRC.With true prior and likelihood, the Bayes detector achieves the best estimate in the MSE sense.Source: §III · p. 3
A. De-Quantization (DQ)-optimal Detector
The DQ-optimal detector uses the true quantized-output likelihood and input prior to obtain the posterior mean estimate. GAMP makes this computation iterative, though nonlinear integrations and high-dimensional marginals remain computationally demanding.
A. De-Quantization (DQ)-optimal Detector
- Using the true quantized-output likelihood with the true input prior gives the DQ-optimal detector, whose posterior mean achieves the best estimate in MSE.The quantized likelihood depends on antenna-specific integration intervals in the mixed-ADC architecture.Source: A. De-Quantization (DQ)-optimal Detector · p. 4
- GAMP approximates belief propagation and iteratively solves for marginal posterior probabilities, making it applicable to the DQ-optimal detector and other detectors in the framework.The exact marginal posterior involves a high-dimensional integral, while belief propagation remains too complex for practical use.Source: A. De-Quantization (DQ)-optimal Detector · p. 4
- GAMP’s first steps estimate Hx and its variance, then form residual terms; later steps produce an observation of x under an AWGN model and estimate posterior means and variances.The estimate of Hx includes an Onsager correction.Source: A. De-Quantization (DQ)-optimal Detector · p. 4
- The DQ-optimal detector’s nonlinear integrations make calculations difficult, and differing ADC resolutions add complexity, motivating mismatched measures to simplify computation.The nonlinear functions for the detector are described as highly nonlinear.Source: A. De-Quantization (DQ)-optimal Detector · pp. 4–5
B. Pseudo-De-Quantization (PDQ)-optimal Detector
The PDQ-optimal detector replaces the exact quantized-output likelihood with a pseudo-quantization-noise model. This simplifies the detector’s nonlinear calculations into linear steps, with antenna-specific noise variance adjustments for different ADC resolutions.
B. Pseudo-De-Quantization (PDQ)-optimal Detector
- The PQN model treats quantization error as additive noise, generally assuming it is independent of the input and modeled as zero-mean complex Gaussian noise.Its variance is denoted σ^2_q.Source: B. Pseudo-De-Quantization… · p. 5
- Using the PQN likelihood changes the nonlinear calculations into simple linear steps, defining the lower-complexity PDQ-optimal detector.The method is obtained by using the mismatched likelihood in the GAMP algorithm.Source: B. Pseudo-De-Quantization… · p. 5
- The PDQ-optimal detector is AMP with an adjusted γ for each antenna to reflect its ADC resolution.The detector arises from GAMP by postulating a mismatched output likelihood.Source: B. Pseudo-De-Quantization… · p. 5
C. Linear Detector
The linear detector uses a Gaussian input assumption within the PQN model to simplify Bayesian estimation, while the broader detector family is analyzed through a unified GAMP-based framework.
C. Linear Detector
- With a Gaussian postulated input, the PQN-model Bayes detector becomes linear; its regularization factor yields zero-forcing at 0 and MRC at infinity.The linear estimator can also be implemented through Algorithm 1 by substituting the Gaussian input prior into its posterior steps.Source: C. Linear Detector · p. 5
- Algorithm 1’s linear-estimator implementation converges in 8–10 iterations and has lower complexity than matrix inversion, using hardware-suitable matrix-vector multiplications.The stated complexity orders are O(NKtmax) for the iterative implementation and O(NK2) for matrix inversion.Source: C. Linear Detector · p. 5
IV. PERFORMANCE ANALYSIS
- The unified framework spans detectors from the optimal but complex estimator to the simple MRC detector, with choices of postulated prior and likelihood defining their variants.The passage identifies these as special cases of Bayesian inference.Source: §IV · p. 5
- Because the detector family is realized by GAMP, its performance can be analyzed using state-evolution equations.The analysis section derives state evolution for the generic Bayes detector and particularizes it to three detectors.Source: §IV · p. 5
A. State Evolution Analysis
The analysis derives state evolution for GAMP in the large-system regime, reducing detector behavior to a scalar equivalent channel. This supports performance prediction across postulated models and quantization processes.
A. State Evolution Analysis
- The large-system analysis characterizes each iteration’s input to the posterior step using the central-limit theorem and evolving parameters that do not depend on user index.The regime takes K and N to infinity while their ratios are held fixed.Source: A. State Evolution Analysis · pp. 5–6
- The scalar equivalent channel decouples the K posterior operations into uncoupled scalar operators, enabling computation of metrics such as MSE, BER, and mutual information.The scalar channel uses postulated prior and likelihood functions, which may be mismatched to the true measures.Source: A. State Evolution Analysis · p. 6
- The analytical framework accommodates arbitrary postulated priors and likelihoods, configurable quantization including nonuniform quantization, and mixed-ADC architectures.For the non-mixed case, the analysis recovers the asymptotic BER expression in.Source: A. State Evolution Analysis · pp. 6–7
- The state-evolution parameters can be computed iteratively with scalar operations, allowing theoretical performance to be predicted quickly instead of using time-consuming Monte Carlo simulations.For QPSK, the analysis also provides a theoretical BER expression.Source: A. State Evolution Analysis · pp. 6–7
B. Explicit Expressions
For QPSK, the paper specializes the state-evolution expressions to three detectors and compares analytical BER predictions with simulations. The results show that quantization precision and step size shape detector performance.
B. Explicit Expressions
- The QPSK state-evolution equations are specialized for the DQ-optimal, PDQ-optimal, and linear detectors, with further simplifications for the latter two.The specialized expressions are listed in Table II.Source: B. Explicit Expressions · p. 7, Table II · p. 8
- At BER=10^-3, the 3-bit PDQ-optimal detector incurs 0.58 dB loss versus the optimal unquantized MIMO detector.At 1 bit, the PDQ-optimal and linear estimators have error floors; at 3 bits, the gap between DQ-optimal and PDQ-optimal is extremely small.Source: Figure 3 · p. 7, B. Explicit Expressions · p. 7
- Analytical BER predictions closely match simulations, and the authors report that extensive BER and MSE checks also covered mixed-ADC cases.The plotted comparison uses N=200, K=50, and ∆=0.5; other examination plots are omitted for space.Source: B. Explicit Expressions · p. 7, Figure 3 · p. 7
- The expected performance ordering between the linear and PDQ-optimal detectors does not hold straightforwardly at 1-bit quantization, and PDQ-optimal performance is sensitive to step size ∆.The passage highlights parameter setting as important for these detectors.Source: B. Explicit Expressions · pp. 7–8
V. DISCUSSIONS
The discussion examines how detector performance depends on quantization-step size and how to choose the PQN model’s regularization variance. It also identifies when a Gaussian-input step-size rule is useful.
A. Quantization Step Size
- At a step size of 2.06, the 1-bit PDQ-optimal detector achieves its best BER; the DQ-optimal detector is step-size invariant, while PDQ-optimal detection is more sensitive than linear detection.At the step size 0.5 used in Figure 3, the PDQ-optimal detector’s BER is 0.280.Source: A. Quantization Step Size · p. 8, Figure 4 · p. 9
- For all bit resolutions, PDQ-optimal and linear detectors have remarkably close step sizes because both model quantization as additive error, suggesting prior mismatch may not significantly change optimized step size.For Gaussian inputs, normalized step sizes are very close to those in for input bit rates κ from 2 to 7. This rule does not apply to 1-bit quantization or the DQ-optimal detector.Source: A. Quantization Step Size · p. 8
B. Regularization Factor
- For the PDQ-optimal detector, MSE is insensitive to sufficiently small σ2 q (for example, below 1) at λ = 4, but σ2 q selection matters more at λ = 1.The rule-of-thumb variance is described as generally effective, though not precise.Source: B. Regularization Factor · p. 9
C. Mixed Architecture
The section evaluates mixed-ADC detectors after discussing parameter choices, comparing them with pure-ADC detectors for QPSK and Gaussian inputs. It finds that a small share of full-precision ADCs can narrow performance gaps associated with the PQN model, while the conclusion notes high-SNR limits to that approximation.
1) QPSK input:
- With QPSK, 5% full-precision ADCs eliminate the error floor for the mixed 1-bit PDQ-optimal detector, and 10% yields performance similar to the pure 1-bit DQ-optimal detector.For the mixed 2-bit architecture, an 80% 2-bit quantization mixed PDQ-optimal detector achieves performance similar to the pure 2-bit DQ-optimal detector.Source: 1) QPSK input · p. 9, Figure 6 · p. 10
2) Gaussian input:
- For Gaussian inputs at SNR = 0 dB, 1-bit quantization degrades MSE by approximately 3 dB versus the unquantized case; at high SNR, 1-bit resolution generally incurs significant losses.For λ = 16, the 1-bit PDQ-optimal detector generally loses 1 bit relative to the DQ-optimal detector; MSE generally improves 3–6 dB per bit-rate increase, nearer 6 dB at higher SNR.Source: 2) Gaussian input · p. 10, Figure 7 · p. 11
- For Gaussian inputs with λ > 10, the pure 1-bit PDQ-optimal detector incurs approximately 6 dB loss versus the pure 1-bit DQ-optimal detector; 5% and 20% full-precision ADCs reduce the gaps to 3 dB and 1 dB, respectively.Figure 8 presents these comparisons for the mixed architecture at SNR = 20 dB.Source: 2) Gaussian input · p. 10, Figure 8 · p. 11
VI. CONCLUSION
- The conclusion states that the PQN simplification may cause 1-bit performance degradation at high SNR but has no significant effect at low SNR, and that mixed architecture can narrow the resulting gap.The state-evolution expressions are quickly evaluated for system-design optimization.Source: §VI · p. 11
APPENDIX A PROOF OF PROPOSITION 1
The appendix derives state-evolution equations for Algorithm 1 in the large-system limit. It uses asymptotic distributions and relationships among detector variables to characterize the updates.
APPENDIX A PROOF OF PROPOSITION 1
- The derivation takes the large-system limit N, K → ∞ and restricts its presentation to a corresponding real-valued system, although Proposition 1 covers the complex-valued case.Iteration indices are omitted for brevity.Source: Appendix A · p. 11
- The state-evolution updates are characterized through asymptotic behavior of the detector variances and means, with parameters determined by the stated intermediate quantities.For the complex-valued case, the derivation adjusts the quantization interval and scalar multiplication.Source: Appendix A · pp. 11–13
- The derivation uses the central-limit theorem to treat the measurement variable as Gaussian conditioned on the detector estimate, then obtains a joint distribution for quantized signals and that variable.The joint distribution is used to calculate expectations needed for the asymptotic behavior of the updates.Source: Appendix A · pp. 11–12