Abstract
The high-SNR capacity of the noncoherent MIMO channel has traditionally been derived for independent and identically distributed (IID) Rayleigh block fading by exploiting the Gaussianity of the channel matrix. This derivation establishes the optimal degrees of freedom (DoF)—the capacity pre-log factor. However, IID Rayleigh fading is a specific condition that may not reflect all physical environments. This paper demonstrates that the optimal DoF for the IID Rayleigh block fading channel is also the optimal DoF for a much broader class of generic block fading channels, characterized simply by finite power and finite differential entropy. We provide a novel converse proof using the duality approach to establish this fundamental limit.
Introduction
In wireless communication, Multiple-Input Multiple-Output (MIMO) technology is essential for achieving high data rates. Most theoretical frameworks assume that either the transmitter or the receiver possesses perfect Channel State Information (CSI). However, in many practical, high-mobility, or low-latency scenarios, the channel changes too rapidly for accurate estimation, leading to noncoherent communication.
The fundamental metric for evaluating these systems at high Signal-to-Noise Ratio (SNR) is the Degrees of Freedom (DoF), which represents the number of independent signaling dimensions available. While the DoF for noncoherent MIMO was well-established for Gaussian (Rayleigh) channels, it remained unclear whether these limits were a byproduct of the Gaussian distribution or a universal property of the MIMO architecture. This research bridges that gap by proving the robustness of the DoF result across a wide range of fading distributions.
The Generic Fading Model
The authors move beyond the restrictive IID Rayleigh model to a "generic" block fading model. A channel matrix $\mathbf{H}$ is considered generic if it satisfies two mild conditions:
- Finite Power: $E[\|\mathbf{H}\|_F^2] < \infty$.
- Finite Differential Entropy: $h(\mathbf{H}) > -\infty$.
These conditions essentially exclude channels that are infinitely powerful or deterministic (zero noise/entropy). By proving the DoF for this class, the authors show that the specific "shape" of the fading distribution does not affect the scaling law of the capacity at high SNR.
Methodology: The Duality Approach
The core contribution of this work is a novel converse proof—an upper bound on the maximum achievable rate. To achieve this, the authors utilize the duality approach in information theory.
Instead of directly maximizing the mutual information $I(\mathbf{X}; \mathbf{Y})$, which is mathematically intractable for generic distributions, the duality approach provides an upper bound by choosing an "auxiliary" output distribution $q(\mathbf{Y})$. The capacity $C$ is bounded by: $$C \leq E[D(p(\mathbf{Y}|\mathbf{X}) \| q(\mathbf{Y}))]$$ where $D(\cdot\|\cdot)$ is the Kullback-Leibler divergence.
The technical innovation lies in the construction of this auxiliary distribution. By performing a Singular Value Decomposition (SVD) on the received signal matrix, the authors separate the signal into its geometric components. They then construct a distribution for the singular values that accounts for the geometric structure of the Stiefel manifold (the space of unitary matrices). This geometric approach allows them to bound the entropy of the received signal without needing to know the exact distribution of the channel $\mathbf{H}$.
Key Results and Interpretation
The paper proves that for a noncoherent $M \times N$ MIMO channel with a coherence block of $T$ symbols, the optimal DoF ($d_{opt}$) is: $$d_{opt} = M^* \left( 1 - \frac{M^*}{T} \right)$$ where $M^* = \min\{M, N, \lfloor T/2 \rfloor\}$.
Implications of the Formula
- The Cost of Training: The term $(1 - M^*/T)$ represents the fraction of time available for data transmission after the "cost" of implicitly or explicitly learning the channel is subtracted.
- Antenna Diminishing Returns: The result $M^* = \min\{M, N, \lfloor T/2 \rfloor\}$ implies that adding more antennas only helps up to a certain point. If the number of antennas exceeds half the coherence time ($T/2$), the overhead of estimating the channel between those antennas outweighs the spatial multiplexing gain.
- Universality: Because this formula holds for all generic fading, it confirms that pilot-based schemes—where a portion of the block is dedicated to known training symbols—are not just practical heuristics but are fundamentally near-optimal for achieving the capacity slope.
Conclusion and Future Work
This work settles the question of the optimal DoF for noncoherent MIMO under general conditions, proving that the Gaussian assumption is sufficient but not necessary for the standard DoF result. However, several challenges remain. While the slope (DoF) is universal, the intercept (the fading number or constant offset) remains highly dependent on the specific channel distribution. Additionally, for very small coherence blocks ($T$), the approximations used in high-SNR analysis become less reliable. Finally, the authors suggest that their SVD-based duality proof offers a potential roadmap for solving the DoF of noncoherent Multiple-Access Channels (MAC), where multiple users share the same spectrum.
Summary updated