Capacity of Compound MIMO Channels with Additive Uncertainty

Yin Sun, Can Emre Koksal, Ness B. ShroffarXiv 2012 · OpenReview

Introduction

In the evolution of wireless communications, Multiple-Input Multiple-Output (MIMO) technology has become a cornerstone for achieving high data rates. However, the theoretical promise of MIMO often assumes perfect Channel State Information (CSI) at the transmitter. In practical scenarios—ranging from high-mobility environments to systems with limited feedback—the transmitter only possesses an estimate of the channel, known as the nominal channel ($H_0$).

This paper addresses the fundamental problem of communicating reliably over a MIMO Gaussian channel where the actual channel matrix $H$ is unknown but resides within a bounded uncertainty region around $H_0$. This is modeled as a compound MIMO Gaussian channel. The objective is to determine the "max-min" capacity: the highest achievable rate that can be guaranteed for any possible channel realization within the uncertainty set.

The Challenge of Non-Convexity

Finding the max-min capacity is traditionally a daunting task. The mutual information objective function is generally non-convex with respect to the channel matrix $H$. Consequently, standard optimization techniques often fail to find global optima for the worst-case channel. For years, researchers have sought to understand whether the optimal transmit strategy for a nominal channel remains optimal under uncertainty, and whether the "max-min" capacity equals the "min-max" capacity—a property known as a zero duality gap.

Methodology: Revealing Hidden Convexity

The authors overcome these challenges by demonstrating that the problem possesses a hidden convexity. Their approach relies on three primary mathematical pillars:

  1. Diagonalization of the Nominal Channel: The authors prove that the optimal transmit covariance matrix $Q$ must share the same singular vectors as the nominal channel $H_0$. This simplifies the problem significantly, as it implies the transmitter should align its spatial layers with the estimated principal components of the channel.
  2. A New Matrix Determinant Inequality: A central contribution of the work is "Lemma 1," a technical proof showing that when minimizing a determinant-based objective under spectral norm constraints, the optimal error matrix must be diagonal.
  3. Unitary Invariance: By exploiting the properties of the spectral norm (which is unitarily invariant), the authors show that the orientation of the uncertainty does not disrupt the fundamental structure of the optimal solution.

Proving the Loyka-Charalambous Conjecture

A landmark result of this paper is the formal proof of the conjecture proposed by Loyka and Charalambous (2012). The conjecture posited that for MIMO channels with additive uncertainty, the order of the "max" and "min" operations can be swapped without changing the result ($C_{max-min} = C_{min-max}$).

By establishing a zero duality gap, the authors prove that the worst-case capacity is equivalent to the capacity of the worst-case channel. This result allows the complex max-min optimization to be reduced to a much simpler, one-dimensional power allocation problem over the singular values of the channel, which can be solved efficiently using water-filling-like algorithms.

Key Results and Structural Insights

The research provides a closed-form characterization of the optimal transmit covariance $Q^*$ and the worst-case channel $H^*$. The findings confirm that:

  • The optimal strategy is to perform independent power allocation across the eigenmodes of the nominal channel.
  • The worst-case channel realization is one where the uncertainty "subtracts" from the strongest singular values of the nominal channel.
  • The spectral norm is a unique boundary; the authors provide a counterexample using the nuclear norm, showing that the "diagonalization is optimal" rule does not necessarily hold for all types of uncertainty measures.

Conclusion and Future Directions

This work provides a definitive answer to a long-standing problem in robust information theory. While it solves the case for spectral norm uncertainty, it opens new questions regarding other unitarily invariant norms, such as the Frobenius norm. Furthermore, while the theoretical capacity is now established, the next step for the engineering community is to translate these robust covariance designs into practical, low-complexity algorithms for the next generation of wireless systems.

Summary updated

Abstract

This paper considers reliable communications over a multiple-input multiple-output (MIMO) Gaussian channel, where the channel matrix is within a bounded channel uncertainty region around a nominal channel matrix, i.e., an instance of the compound MIMO Gaussian channel. We study the optimal transmit covariance matrix design to achieve the capacity of compound MIMO Gaussian channels, where the channel uncertainty region is characterized by the spectral norm. This design problem is a challenging non-convex optimization problem. However, in this paper, we reveal that this problem has a hidden convexity property, which can be exploited to map the problem into a convex optimization problem. We first prove that the optimal transmit design is to diagonalize the nominal channel, and then show that the duality gap between the capacity of the compound MIMO Gaussian channel and the min-max channel capacity is zero, which proves the conjecture of Loyka and Charalambous (IEEE Trans. Inf. Theory, vol. 58, no. 4, pp. 2048-2063, 2012). The key tools for showing these results are a new matrix determinant inequality and some unitarily invariant properties.
BibTeX
@misc{sun2012capacity,
  author = {Yin Sun and Can Emre Koksal and Ness B. Shroff},
  title = {{Capacity of Compound MIMO Channels with Additive Uncertainty}},
  year = {2012},
  eprint = {1208.4656},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1208.4656},
}
Figure