Diversity vs Degrees of Freedom in Gaussian Fading Channels

Technical Summary: Diversity vs. Degrees of Freedom for Gaussian Fading Channels

Problem Statement
The paper addresses a fundamental limitation in the classical definitions of Degrees of Freedom (DOF) and Diversity (DIV) for Gaussian fading channels. Traditionally, these metrics are extracted using a fixed gauge of $\log \rho$ (where $\rho$ is the signal-to-noise ratio): DOF is defined as $\lim_{\rho \to \infty} C(\rho)/\log \rho$, and DIV as $\lim_{\rho \to \infty} -\log P_e(\rho)/\log \rho$.

The author argues that this approach conflates two distinct operations: gauge selection (identifying the correct asymptotic scale on which capacity or reliability grows) and atomic normalization (dividing the raw growth coefficient by the coefficient of a single "atom" or unit of information/reliability). In the classical coherent complex MIMO case, the gauge happens to be $\log \rho$ and the relevant atom coefficients happen to be 1, rendering these steps numerically invisible. However, for channels where the growth scale differs from $\log \rho$ (e.g., noncoherent fast fading where capacity grows as $\log \log \rho$) or where the raw coefficient includes non-unit constants (e.g., real AWGN or fixed-H channels), the classical definitions yield uninformative or incorrect results (e.g., zero prelog or infinite diversity slope).

Methodology
The paper proposes a two-step framework to decouple gauge selection from atomic normalization, utilizing the Bhattacharyya frontier as a unifying geometric tool.

1. Bhattacharyya Frontier Construction: The core tool is the Bhattacharyya frontier, $\Delta^*_B(K; \rho)$, defined as the maximum, over all admissible $K$-point codebooks, of the minimum pairwise Bhattacharyya distance between induced output laws.

   • Capacity Side (Rate): Fixed-separation packing of codewords identifies the capacity gauge. Theorem 1 (Capacity–Packing Sandwich) demonstrates that if fixed-resolution output-law packing and covering have matching entropy coefficients, the capacity gauge and the raw capacity coefficient are recovered.
   • Diversity Side (Reliability): The binary endpoint ($K=2$) of the frontier identifies the zero-rate diversity gauge. Theorem 2 establishes that the binary endpoint is gauge-equivalent to the optimal two-message testing error exponent.

2. Atomic Normalization: Once the correct gauge and the raw coefficient are identified, the paper defines "atoms" to normalize these values:

   • Communication Atom: One independently controllable input degree of freedom (e.g., a SISO input direction or a Grassmannian subspace).
   • Reliability Atom: One independent unit-normalized output reliability look (e.g., a unit-gain noisy observation).
   • Endpoint DOF/DIV: The final metrics are obtained by dividing the raw coefficient by the corresponding atom coefficient ($\alpha_{comm}$ or $\alpha_{rel}$).

Key Contributions and Results
The paper applies this framework to seven channel families, distinguishing between "same-gauge" cases (where classical definitions hold) and "cross-gauge" cases (where they fail).

• Fixed Deterministic H (Cross-Gauge):

  • Capacity: Grows on the $\log \rho$ gauge. The block endpoint DOF is $T \cdot \text{rank}(H)$.
  • Diversity: Grows on the linear $\rho$ gauge (not $\log \rho$). The binary endpoint yields a raw coefficient of $T\sigma_1^2(H)/\ln 2$. After normalization by the fixed-AWGN reliability atom ($1/\ln 2$), the endpoint DIV is $T\sigma_1^2(H)$.
  • Insight: This channel is "cross-gauge." The capacity is determined by covering the input image (input dimensions), while diversity is determined by expanding pairwise output separation (output looks). The endpoint DIV is gain-dependent for unnormalized $H$.

• Noncoherent Fast Fading (Cross-Gauge):

  • Capacity: Grows on the $\log \log \rho$ gauge (recovering Lapidoth–Moser results). The endpoint DOF is 1.
  • Diversity: Grows on the $\log \rho$ gauge. The binary endpoint has a raw coefficient of $N/2$. After normalization by the fast-fading reliability atom ($1/2$), the endpoint DIV is $N$.
  • Load-r Frontier: The paper derives an exact load-r frontier gauge of $(\log \rho)^{1-r}$ with a coefficient of $N/2$.

• Same-Gauge Calibration Cases:

  • Coherent Rayleigh MIMO and Noncoherent Block Fading are shown to be same-gauge cases where both capacity and diversity grow on the $\log \rho$ scale. The raw coefficients match the classical DOF and DIV counts (e.g., zero-rate DIV $MN$) because the relevant atom coefficients are 1.

• Open and Conditional Cases:

  • Multipath Fading: A guarded construction provides a lower bound on the diversity gauge ($\log \rho$) and a lower bound on endpoint DIV, but the exact unguarded frontier remains open.
  • Parallel Fractional-Log Fading: Provides a conditional lower-bound benchmark based on Gallager's $E_0$ function, assuming a product input ensemble.
  • Stationary Toeplitz Fading: The capacity gauge is $(\log \rho)^\beta$ (fractional log), but the zero-rate diversity gauge and load-r frontier remain open due to the non-commutativity of covariance matrices across codewords, which prevents coordinatewise factorization.

Significance and Claims
The paper claims to clarify the definitions of DOF and DIV by explicitly separating the identification of the asymptotic growth scale (gauge) from the normalization of the growth coefficient (atomic count).

• Correction of Misconceptions: It demonstrates that "zero prelog" or "infinite diversity slope" in certain channels is not a lack of signal structure but a mismatch of the gauge (using $\log \rho$ when the channel grows as $\log \log \rho$ or $\rho$).
• Unified Framework: It introduces the Bhattacharyya frontier as a gauge-free geometric object that consistently extracts both capacity and diversity metrics across diverse channel models.
• Precision: It provides exact endpoint DOF and DIV values for cross-gauge channels (Fixed-H and Noncoherent Fast Fading) where classical definitions fail, and establishes exact load-r frontiers for the fast-fading case.
• Modesty: The paper explicitly acknowledges that for several models (Multipath, Stationary Toeplitz), the exact diversity gauges and load-r frontiers are not yet determined. It frames the results for these cases as lower bounds, conditional benchmarks, or open problems, avoiding claims of exact operational reliability functions where they have not been proven.

In summary, the paper argues that the classical coherent MIMO convention has obscured the necessity of gauge selection and atomic normalization. By making these steps explicit, the framework recovers classical results where appropriate and provides rigorous, corrected metrics for channels where the classical $\log \rho$ gauge is insufficient.