Blind Joint MIMO Channel Estimation and Decoding

Thomas R. Dean, Mary Wootters, Andrea J. GoldsmithIEEE Trans. Information Theory 2019 · DOI

TL;DR

This paper introduces a novel method for simultaneously estimating the wireless channel and decoding transmitted data in MIMO (Multiple-Input Multiple-Output) systems without prior knowledge of the channel. This "blind" approach leverages the geometric structure of common signal constellations and offers a practical solution for scenarios where channel state information is unavailable or unreliable, outperforming existing methods in certain conditions.

What Problem Does This Paper Address?

In modern wireless communication, accurately knowing the characteristics of the wireless channel (channel state information, or CSI) is crucial for efficient data transmission. However, obtaining this CSI often requires dedicated pilot signals, which consume valuable transmission bandwidth and reduce overall data rates. For instance, in cellular systems, up to 15% of transmission overhead can be dedicated to channel estimation. While some methods exist to communicate without CSI, they often incur significant performance penalties or are computationally complex. This paper tackles the challenge of performing joint channel estimation and data decoding in MIMO systems blindly, meaning without any pilot symbols or prior knowledge of the channel, aiming to reduce overhead and improve performance, especially in scenarios with imperfect CSI.

Methods

The core of this paper's contribution lies in formulating the blind MIMO decoding problem as a novel, non-convex optimization problem. The key idea is to fit a parallelepiped to the received signal samples. This parallelepiped's inverse is related to the unknown channel gain matrix. The method exploits the geometric structure of hypercubic constellations (like BPSK or M-PAM), where transmitted symbols form a regular n-dimensional polytope with mutually perpendicular sides.

The optimization problem involves maximizing the determinant of a matrix $U$ (related to the inverse of the channel matrix) subject to constraints that ensure the transformed received signals lie within a bounded region. This is framed as fitting a minimum volume parallelepiped to the observed data.

Two primary algorithms are proposed:

  1. Algorithm 1 (Gradient Descent): A standard gradient descent approach is used to solve the non-convex optimization problem.
  2. Algorithm 2 (Modified Gradient Descent): This algorithm is an enhancement of Algorithm 1, designed to address potential issues with local optima in the non-convex landscape by ensuring termination at a vertex of the feasible region.

The theoretical analysis focuses on deriving conditions under which these algorithms can reliably recover the channel and decode the data, particularly for smaller MIMO dimensions (n=2, 3, 4).

Datasets

Not discussed in the paper. The experiments are based on simulated data generated according to the described system model.

Key Insights and Ideas

  • Geometric Structure is Key: The paper demonstrates that the geometric properties of common signal constellations (hypercubes) are sufficient to enable blind channel estimation and decoding. This is a powerful insight, as it suggests that even without explicit training signals, the inherent structure of the data can be exploited.
  • Optimization as a Solution: Reframing the blind decoding problem as fitting a parallelepiped to received samples leads to a well-defined (though non-convex) optimization problem. This allows the application of powerful optimization techniques.
  • Blind Decoding Outperforms Imperfect CSI: Empirically, the proposed blind decoding method shows comparable performance to traditional methods with perfect CSI and significantly outperforms them when CSI is imperfect (e.g., with 1% or 10% estimation error). This highlights the robustness of the blind approach.
  • Theoretical Guarantees for Small Systems: The paper provides rigorous theoretical conditions for the correctness of the algorithm for small MIMO dimensions (n=2, 3, 4), linking the problem to the Hadamard Maximal Determinant problem.
  • Practical Feasibility: Despite the non-convex nature of the problem, empirical results show that gradient descent (Algorithm 1) works well in practice, requiring a surprisingly small number of samples, making it suitable for block-fading channels.

Key Results

  • Performance Loss: The blind decoding approach incurs a loss of less than 3dB compared to zero-forcing with perfect CSI.
  • Sample Efficiency: The method requires significantly fewer samples than previous blind decoding techniques, especially for hypercubic constellations. For example, for n=2, M=16, the proposed method needs 56 samples, while a prior method [10] requires 913.
  • Outperforms Imperfect CSI Methods: In simulations with imperfect CSI (1% and 10% error), the blind decoder consistently outperformed Zero-Forcing (ZF) and Maximum Likelihood (ML) decoders.
  • Theoretical Success Rates: For n=2, 3, and 4, theoretical probabilities of success are derived, closely matching empirical results. For n=4, the algorithm is guaranteed to succeed with probability 1 under specific conditions on the transmitted data.

Limitations and Future Work

  • Non-Convexity: While gradient descent works well empirically, the non-convex nature of the optimization problem means that theoretical guarantees are challenging to establish for larger MIMO dimensions.
  • AWGN and Higher Modulations: The theoretical analysis is primarily for the noiseless case (AWGN $\sigma=0$) and BPSK modulation (M=2). While empirical results show good performance for AWGN and higher M, analytical explanations are left as future work.
  • Complex and Rectangular Channels: Extensions to complex-valued channels and scenarios with different numbers of transmit and receive antennas (rectangular channels) are suggested as future research directions.
  • Algorithm Efficiency: Improving the runtime of the algorithm, possibly through intelligent selection of received samples, is identified as a topic for future research.
  • Open Theoretical Questions: The paper highlights open questions regarding the Hadamard Maximal Determinant problem for larger dimensions and the precise conditions for guaranteed convergence of blind decoding algorithms in more general MIMO systems.

Summary updated

BibTeX
@article{dean2019blind,
  author = {Thomas R. Dean and Mary Wootters and Andrea J. Goldsmith},
  title = {{Blind Joint MIMO Channel Estimation and Decoding}},
  journal = {IEEE Trans. Information Theory},
  year = {2019},
  doi = {10.1109/tit.2018.2878016},
}
Figure