PaperlayerView exact revision on arXiv

Source-linked AI summary

Massive MIMO Channel Estimation with an Untrained Deep Neural Network

Eren Balevi, Akash Doshi, Jeffrey G. Andrews

arXiv:1908.00144v1eess.SP

TL;DR

The paper addresses the practical complexity and training-overhead challenges of channel estimation. It proposes a DNN followed by a simple LS-type estimator, reporting performance better than complex MMSE and approaching Genie-Aided MMSE while using few parameters.

  • Problem

    SVD-based processing is impractically complex for large channel matrices, while training overhead obstructs practical use of state-of-the-art DNNs.

  • Method

    The proposed deep channel estimator combines a DNN with a simple LS-type estimator.

  • Results

    The low-complexity estimator performs better than more complex MMSE and approaches Genie-Aided MMSE with perfectly known channel statistics.

  • Takeaways & Limitations

    The estimator efficiently exploits time-frequency correlations while using hundreds or thousands of weights rather than millions, with parameter growth below the square root of the antenna count.

Abstract

from arXiv · show

This paper proposes a deep learning-based channel estimation method for multi-cell interference-limited massive MIMO systems, in which base stations equipped with a large number of antennas serve multiple single-antenna users. The proposed estimator employs a specially designed deep neural network (DNN) to first denoise the received signal, followed by a conventional least-squares (LS) estimation. We analytically prove that our LS-type deep channel estimator can approach minimum mean square error (MMSE) estimator performance for high-dimensional signals, while avoiding MMSE's requirement for complex channel inversions and knowledge of the channel covariance matrix. This analytical result, while asymptotic, is observed in simulations to be operational for just 64 antennas and 64 subcarriers per OFDM symbol. The proposed method also does not require any training and utilizes several orders of magnitude fewer parameters than conventional DNNs. The proposed deep channel estimator is also robust to pilot contamination and can even completely eliminate it under certain conditions.

I. INTRODUCTION

Massive MIMO channel estimation must remain accurate and scalable despite pilot contamination, noise, and costly large-matrix operations. The paper develops a deep-learning-based LS-type estimator aimed at addressing these constraints without channel statistics or training data.

  • Accurate CSI is central to precoding and coherent combining in massive MIMO systems serving many users across shared time-frequency resources.
  • Pilot contamination and noise make multicell massive MIMO channel estimation challenging, while matrix inversion and SVD become impractically complex for large channel matrices.
  • The method avoids training datasets and channel-statistics knowledge, while remaining applicable to Gaussian or non-Gaussian, LOS or NLOS, and limited- or rich-scattering channels.
  • The proposed estimator modifies a deep image prior architecture for massive MIMO with moderate parameters and uses only information directly obtained from the received signal.
  • MMSE estimation is more accurate than LS but requires channel-correlation estimates and matrix inversion, with complexity growing as the cube of the antenna count.
  • Prior covariance- or sparsity-based approaches require large covariance matrices or assume them available and are limited mainly to NLOS zero-mean Gaussian channels.
  • Normal LS estimation decreases average spectral efficiency by about 50% versus MMSE, motivating an improved LS-type estimator without covariance matrices or inversions.

B. Contributions

The paper proposes an untrained, low-complexity DNN channel estimator that denoises received signals before LS estimation and approaches MMSE performance asymptotically. Simulations indicate practical operation in moderate dimensions and robustness to pilot contamination.

  • Estimator design: The proposed estimator uses a specially designed DNN to denoise the received signal before a conventional LS-type operation.The architecture is adapted from deep image prior methods and optimized for channel estimation.
  • Complexity and training: The method reduces the parameter count from millions to hundreds or a few thousand and requires no training.Avoiding training also avoids complexity growth due to training and performance loss across differing channel realizations.
  • Asymptotic performance: The proposed estimator mathematically approaches and ultimately achieves MMSE performance as the product of antennas, subcarriers, and coherence time grows.Simulations support this behavior for a 64 × 64 × 64 signal block.
  • Pilot contamination: The estimator is designed to be robust to pilot contamination by learning priors from interference-free OFDM regions and patching them into contaminated areas.The setup uses randomly spread pilots allocated orthogonally over the time-frequency grid during one coherence interval.
  • Pilot contamination: Under conditions including 5% contaminated OFDM-grid area, fourfold weaker interference, and low noise, the estimator can completely remove interference.This remains possible even when the desired-user and interferer eigenspaces fully overlap.

III. PROBLEM STATEMENT

The problem statement contrasts low-complexity LS estimation with more accurate but costly MMSE estimation and identifies the difficulty of applying conventional neural networks to high-dimensional channel signals. The paper therefore motivates an untrained denoising DNN followed by LS estimation.

  • Classical estimators: MMSE estimation requires users’ autocorrelation matrices and a matrix inversion whose complexity scales as (MNf)^2.These requirements make MMSE unsuitable for systems with many antennas or subcarriers.
  • Classical estimators: LS estimation has very low complexity but provides much less accurate channel estimates than MMSE estimation.The resulting spectral-efficiency decrease is particularly considerable with MMSE and ZF combiners.
  • Design goal: A desirable estimator would combine MMSE-level performance with LS-level complexity.This trade-off is especially important because channel-estimation quality strongly affects massive-MIMO performance.
  • Deep-learning challenge: Limited pilot labels make the training requirements of neural channel estimators an impediment to real-time channel estimation.The paper addresses this by denoising the received signal with an untrained DNN before LS estimation.

IV. DEEP CHANNEL ESTIMATOR MODEL

The deep channel estimator adapts an untrained DIP-style DNN to denoise and inpaint received massive MIMO-OFDM signals before LS channel estimation. Its shared, low-parameter architecture is optimized separately for each OFDM grid and combines DNN noise reduction with LS complexity.

  • Motivation: Training overhead is the primary obstacle to applying state-of-the-art DNNs to practical channel estimation.The paper notes that conventional approaches require large datasets or substantial training for channel realizations.
  • DIP adaptation: The estimator modifies a DIP architecture’s input and output layers for channel estimation without requiring training on large datasets.The resulting model is termed a deep channel estimator.
  • Two-stage estimator: The estimator first generates a less noisy signal with a specially designed DNN, then multiplies it by the pilot sequence’s Hermitian transpose for LS estimation.Pilots are used only after DNN filtering, during the LS stage.
  • Two-stage estimator: Replacing the received signal with the DNN-generated signal combines LS’s low complexity with the DNN’s noise-reduction capability.The paper presents this as an LS-type estimator intended to approach MMSE performance.
  • Practical operation: The DNN parameters must be fitted periodically for each OFDM grid, with the period determined by channel coherence time.The paper characterizes the resulting complexity increase as reasonable because the network uses few parameters.
  • Architecture: The network generates the received-signal tensor from a random input through randomly initialized hidden layers optimized by gradient descent.Hidden layers use shared 1×1 convolutions, upsampling, ReLU activation, and batch normalization.
  • Architecture: The architecture shares spatial transformation parameters across time-frequency slots and upsamples neighboring grid elements to exploit their coupling.The time-frequency signal is upsampled by a factor of 2 via bilinear transformation.

V. THEORETICAL ANALYSIS

The analysis argues that the LS-type deep estimator suppresses noise increasingly well in high-dimensional massive MIMO-OFDM signals and can approach MMSE performance. It also predicts strong behavior under localized pilot contamination, subject to explicit conditions.

  • Asymptotic performance: Theorem 1 states that the proposed LS-type deep channel estimator achieves MMSE estimator performance asymptotically for high-dimensional signals.The proof combines noise suppression, overparameterization arguments, and asymptotic error analysis.
  • Noise suppression: The probability of satisfying the noise-suppression condition tends to 1 for the high-dimensional massive MIMO-OFDM signal model.The analysis attributes this behavior to the signal dimensionality.
  • Parameter scaling: The hidden-layer spatial dimension increases sublinearly with the number of antennas and, in the stated bound, grows at worst with the square root of channel rank.Increasing only the last hidden layer is sufficient to fit the received signal and yields the stated scaling behavior.
  • Asymptotic performance: The estimator can ultimately achieve zero estimation error by increasing the numbers of antennas and subcarriers without increasing transmission power.This conclusion is stated as a consequence of the theorem.
  • Pilot contamination: Under pilot contamination, the estimator can resist or completely eliminate interference when it occupies a limited region of the OFDM grid.The paper associates this behavior with inpainting from interference-free regions.
  • Pilot contamination: With sufficiently localized pilot contamination, the LS-type estimator can attain single-cell MMSE performance even in the multicell case.The stated condition is localization in time and frequency.
  • Scope: Comparisons with dictionary-learning methods and integration with such methods for enhanced interference mitigation are left for future work.This marks a scope boundary of the analysis and contribution.

VI. SIMULATIONS

The simulations compare the proposed deep channel estimator with conventional LS and MMSE estimators using LTE-EPA and Kronecker channel models. Performance is evaluated with normalized mean square error.

  • Comparative setup: The proposed estimator is compared with traditional LS and MMSE channel estimators under LTE-EPA and Kronecker channel models.The comparison uses the estimator expressions given in the paper’s earlier formulation.
  • Metric: Normalized mean square error (NMSE) is the performance metric for evaluating channel-estimation quality.The metric compares actual and estimated frequency-domain channel taps over all antennas.
  • Evaluation scope: The simulation section presents experimental details, simulation results, and a discussion of estimator complexity.

A. Experimental Details

The experiments implement the estimator with a compact DNN and evaluate it under LTE-EPA and Kronecker channels, including random pilots and localized pilot-contamination scenarios. Results indicate no significant performance difference between the two channel models and equal performance for block and randomly spread pilots.

  • Implementation: The deep channel estimator is implemented with 6 hidden layers and optimized using two Nvidia GeForce GTX 2070 GPUs.The number of parameters is controlled through the hidden-layer dimensions.
  • Channel models: Performance is evaluated with LTE-EPA and Kronecker channel models, with most results reported for LTE-EPA.The authors use LTE Toolbox realizations with an M × 64 × 64 antenna-subcarrier-symbol structure.
  • Channel models: The Kronecker model uses exponential spatial correlation with coefficient ρ = 0.5 at the base station.
  • Pilot arrangement: Users within a cell receive orthogonal pilots, while pilots may be non-orthogonal across neighboring cells.The estimator imposes no pilot-arrangement constraint because pilots are not used when fitting DNN parameters.
  • Pilot arrangement: Pilots are used only for LS estimation after the received signal has been filtered by the DNN.Random pilot allocation spreads each base station’s pilot tones throughout the OFDM grid.
  • Pilot contamination: Pilot contamination experiments include a single dominant out-of-cell interferer and contiguous interference over selected time-frequency resource elements.The dominant-interferer case uses random QPSK symbols and an LTE-EPA channel realization with fully overlapped covariance matrices.
  • Architectural constraints: The hidden-layer time and frequency dimensions are fixed by the received-signal matrix size and the number of hidden layers.They are therefore not tunable design parameters in the architecture.
  • Pilot arrangement: A block pilot arrangement matches the performance of randomly spread pilots when both use the same number of pilot tones within one coherence interval.

B. Results

The results progress from single-antenna communication to single-cell and multi-cell massive MIMO, first isolating parameter scaling, then evaluating orthogonal and contaminated-pilot settings.

  • The evaluation begins with single-antenna OFDM communication to quantify how the number of network parameters scales with antenna count.
  • The single-cell massive MIMO experiment assumes orthogonal pilots and no pilot contamination at the base station.
  • The multi-cell experiment introduces pilot contamination through non-orthogonal pilot sequences used by neighboring-cell users.

1) Single antenna OFDM Communication:

In single-antenna OFDM communication, the estimator’s architecture and stopping rule affect NMSE, with smaller k favored at lower SNR and k = 8 performing strongly against LS and MMSE.

  • k = 8 yields the lowest NMSE among the tested architectures, while k = 16 and k = 32 perform worse and k = 64 nearly matches k = 8 at 0 dB SNR.
  • At 20 dB SNR, the different architectures have very similar performance.
  • Larger noise levels favor smaller k values, while early stopping provides an alternative when noise is significantly larger.
  • The k = 8 estimator outperforms LS and MMSE and approaches Genie Aided MMSE without statistical information beyond the received signal.
  • Under SIR = 6 dB interference, k = 8 outperforms k = 64 below 10 dB SNR, while their performance becomes similar afterward.
  • With k = 8, the deep estimator beats MMSE up to 10 dB and therefore provides better interference mitigation in this experiment.

2) Single-cell Massive MIMO:

In single-cell massive MIMO, the estimator is evaluated with varying pilot counts, SNRs, antenna correlation, and architecture sizes under orthogonal pilots.

  • Increasing pilots from Np = 1 to Np = 4 improves NMSE, but additional pilots provide no further benefit.
  • The LTE-EPA channel model’s high temporal correlation allows accurate representation with very few time-domain pilots, and the experiments use Np = 1.
  • With M = 64, smaller k values perform better at lower SNR, while all architectures converge toward similar NMSE at higher SNR.
  • k = 16 has the best performance among the tested architectures in the single-cell massive MIMO setting.
  • Changing from LTE-EPA to a Kronecker model with spatial correlation coefficient ρ = 0.5 leaves deep-estimator performance almost unaffected.

3) Multi-cell Massive MIMO:

The multi-cell experiments test the estimator under random and contiguous pilot contamination, showing improved NMSE relative to LS and favorable comparisons with MMSE at low SNR.

  • For the multi-cell experiments, k = 16 is selected because it outperforms the other tested architectures under 5% grid contamination at SIR = 6 dB.
  • With up to 10% pilot contamination, the deep estimator outperforms MMSE up to 7 dB SNR.
  • The NMSE curve flattens as interference increases because the method does not patch regions corrupted beyond a certain interference limit.
  • For contiguous interference blocks covering approximately 3% of the time-frequency grid, performance remains better than LS across SNRs and better than MMSE up to 6 dB when SIR exceeds 10 dB.

C. Complexity

The estimator trades parameter count against optimization time: smaller architectures use fewer parameters but may require more epochs, while parameter growth with antennas remains sub-linear. It combines a denoising DNN with LS estimation and achieves strong performance relative to LS and MMSE-based references.

  • Complexity trade-off: The estimator’s complexity involves a trade-off between the number of parameters and the epochs required to reach minimum NMSE.Lower complexity can require more training epochs, whereas overparameterized architectures can converge faster.
  • Complexity trade-off: For one antenna, k = 8 uses 496 parameters and 2000 epochs, while k = 64 uses 25,472 parameters and 250 epochs with slightly higher NMSE.The comparison illustrates the parameter–convergence-time trade-off directly.
  • Architecture comparison: For 64 antennas, k = 16 has 3776 parameters and requires 1970 epochs to attain its lowest NMSE, whereas k = 128 has 116,224 parameters and requires 1000 epochs.The k = 128 architecture reaches its lowest achievable NMSE faster, but that NMSE is much higher than for k = 16.
  • Scaling: The architecture has 496 parameters for a decoder and 3776 parameters for massive MIMO, indicating sub-linear computational-complexity growth with antenna count.The paper highlights that parameter growth is less than quadratic in the number of antennas.
  • Estimator design: The proposed estimator uses a DNN followed by LS estimation and requires hundreds or a few thousand parameters rather than millions in conventional DNNs.The DNN denoises the received signal before LS channel estimation.
  • Estimation performance: The deep estimator is reported to outperform LS and MMSE estimators, approach Genie-Aided MMSE, and exploit time-frequency correlations efficiently.Its analysis supports strong performance, while the reported comparison addresses pilot-contamination limitations of LS and MMSE.
Loading 1908.00144v1…