Ferkans — Interactive Telecom Tutor

What If the Channel Is Not Rayleigh?

Chapter 12's Zheng-Tse DMT is stated and proved for i.i.d. Rayleigh block fading: the channel entries $(\mathbf{H})_{i j}$ are i.i.d. $\mathcal{CN}(0, 1)$ . This is the most commonly used model, but it is not the only one. Real MIMO channels can be Rician (line-of-sight component present), Nakagami- $m$ (generalised small-scale fading), log-normal (shadowing), or correlated Rayleigh (scattering geometry restricted).

A natural worry is: when we change the fading law, does the DMT curve change? Does a Zheng-Tse-optimal code lose its optimality on a different fading? For Gaussian random codes, the answer is "it depends on the distribution" — a capacity-achieving code for Rayleigh may not be capacity-achieving for Rician. But for explicit CDA-NVD codes something remarkable happens: the same code achieves the DMT curve on every fading distribution (up to an SNR-independent constant), a property Tavildar and Viswanath (2006) christened approximate universality.

The Elia-Kumar-Caire 2006 paper proves approximate universality for CDA-NVD codes. For a designer, the operational consequence is huge: you do not need a separate code for Rician or Nakagami. The same codeword matrix $\lambda(a)$ — which you constructed once — works for every physically reasonable channel. The channel-universality property is a far stronger statement than "DMT-optimal for Rayleigh"; it is a guarantee that the code is robust against mis-specification of the fading model.

,

Definition:
Approximate Universality

A family of space-time codes $\{\mathcal{C}_M\}$ with rate $R_M(\text{SNR}) = r \log_2 \text{SNR}$ is approximately universal over a class $\mathcal{F}$ of fading distributions if there exist SNR-independent constants $c_1, c_2 > 0$ such that for every $P_{\mathbf{H}} \in \mathcal{F}$ and every sufficiently large $\text{SNR}$ , $c_1 \cdot P_{\rm out}^{\mathcal{F}}(R_M, \text{SNR}) \;\le\; P_e(\mathcal{C}_M; \text{SNR}, P_{\mathbf{H}}) \;\le\; c_2 \cdot P_{\rm out}^{\mathcal{F}}(R_M, \text{SNR})$ where $P_{\rm out}^{\mathcal{F}}(R_M, \text{SNR})$ is the outage probability under $P_{\mathbf{H}}$ . Equivalently, the code achieves — up to constants — the DMT exponent of the fading $P_{ \mathbf{H}}$ : $-\lim_{\text{SNR} \to \infty} \frac{\log P_e(\mathcal{C}_M; \text{SNR}, P_{\mathbf{H}})}{\log \text{SNR}} \;=\; d_{P_{\mathbf{H}}} ^*(r)$ for every $P_{\mathbf{H}} \in \mathcal{F}$ .

Approximate universality is to fading distributions what universal coding is to source distributions: one code, optimal across the whole class. The "approximate" qualifier acknowledges that the constants $c_1, c_2$ may depend on $P_{\mathbf{H}}$ but not on $\text{SNR}$ . This is weaker than exact universality (where $c_1 = c_2 = 1$ ) but far stronger than distribution-specific optimality.

,

Definition:
Admissible Fading Class $\mathcal{F}_{\rm adm}$

The class $\mathcal{F}_{\rm adm}$ of admissible fading distributions consists of all probability distributions $P_{\mathbf{H}}$ on the matrix $\mathbf{H} \in \mathbb{C}^{n_r \times n_t}$ satisfying:

Continuity near zero. The eigenvalue density of $\mathbf{H} \mathbf{H}^{H}$ is continuous near $0$ and is bounded above by a polynomial: $f_{\lambda_i}(\epsilon) = O(\epsilon^{\alpha_i})$ as $\epsilon \to 0^+$ for some $\alpha_i > 0$ .
Decay at infinity. $\mathbb{E}[\|\mathbf{H}\|_F^{2 n_t n_r}] < \infty$ — all relevant moments are finite.
No atoms at zero. $\Pr(\mathbf{H} \text{ is singular}) = 0$ .

Examples in $\mathcal{F}_{\rm adm}$ : i.i.d. Rayleigh, i.i.d. Rician (any $K$ -factor), Nakagami- $m$ ( $m > 0$ ), log-normal (any $\sigma$ ), correlated Rayleigh with full-rank covariance, any finite mixture thereof. Excluded from $\mathcal{F}_{\rm adm}$ : any distribution that concentrates mass on rank- deficient channels (a "deep fade" atom), which would break the full-diversity premise.

Theorem: CDA-NVD Codes Are Approximately Universal Over $\mathcal{F}_{\rm adm}$

Let $\{\mathcal{C}_M\}$ be a CDA space-time code family of degree $n_t$ with the non-vanishing-determinant property $\delta_{\min} (\mathcal{C}_M) \ge \delta_0 > 0$ independent of $M$ , transmitted over an $n_t \times n_r$ block-fading channel with block length $L = n_t$ . Then $\{\mathcal{C}_M\}$ is approximately universal over $\mathcal{F}_{\rm adm}$ : for every $P_{\mathbf{H}} \in \mathcal{F}_{\rm adm}$ and every $r \in [0, \min(n_t, n_r)]$ , the code achieves the DMT exponent $d_{P_{\mathbf{H}}}^*(r)$ associated with that fading distribution.

The proof is a worst-case exponent argument: one shows that the NVD condition together with the full-diversity property bounds the error probability by a quantity that depends on $P_{\mathbf{H}}$ only through its outage-probability behaviour. Since the outage probability is — by definition — the minimum achievable error probability at any rate $R$ , matching the outage exponent up to a constant gives optimality for that distribution. The key fact is that NVD controls the worst-case PEP uniformly over codewords and fading realisations, which is exactly what approximate universality requires.

Show Hint

Start from the conditional PEP $P(\ntn{X} \to \hat{\ntn{X}} | \mathbf{H}) \le \exp(- \tfrac{\text{SNR}}{4 n_t} \mathrm{tr}(\mathbf{H} \boldsymbol{\Delta} \boldsymbol{\Delta}^H \mathbf{H}^{H}))$ .

Use the NVD property to bound the eigenvalues of $\boldsymbol{\Delta} \boldsymbol{\Delta}^H$ uniformly: $\lambda_i \ge \delta_0^{1/n_t}$ .

Average over $P_{\mathbf{H}}$ using the eigenvalue density of $\mathbf{H}\mathbf{H}^{H}$ under $P_{\mathbf{H}}$ — this is where the fading-specific outage exponent enters.

Apply Laplace's method at high SNR; the exponent is determined entirely by the rate at which $P_{\mathbf{H}}$ places mass near the rank-deficient region.

Proof

Step 1 — Conditional PEP bound from NVD

For any two CDA codewords $\ntn{X}, \hat{\ntn{X}}$ , the NVD bound gives $\det(\boldsymbol{\Delta}\boldsymbol{\Delta}^H) \ge \delta_0^2$ . By AM-GM, the eigenvalues satisfy $\prod_i \lambda_i \ge \delta_0^2$ , so at least one $\lambda_i \ge \delta_0^{2/n_t}$ . The conditional PEP is bounded by $\exp(- \tfrac{\text{SNR}}{4 n_t} \sum_i \lambda_i(\mathbf{H}) \lambda_i(\boldsymbol{\Delta} \boldsymbol{\Delta}^H))$ .

Step 2 — Union bound with NVD

Union-bounding over the $\doteq \text{SNR}^{2 r n_t}$ codeword pairs (this is where the rate enters) and using NVD to replace $\delta_{\min}$ by $\delta_0$ (an $M$ -independent constant), the conditional block-error probability becomes $P_e \mid \mathbf{H} \le \text{SNR}^{2 r n_t} \exp(- \tfrac{ \delta_0 \text{SNR}}{4 n_t} \cdot \lambda_{\min}(\mathbf{H} \mathbf{H}^{H})^{n_t})$ .

Step 3 — Average over $P_{\ntn{ch}}$

Taking expectation over $P_{\mathbf{H}} \in \mathcal{F}_{\rm adm}$ : the expectation is controlled by the small-eigenvalue behaviour of $\mathbf{H}\mathbf{H}^{H}$ under $P_{\mathbf{H}}$ , which — for admissible distributions — satisfies $\Pr(\lambda_{ \min} \le \text{SNR}^{-\alpha}) \doteq \text{SNR}^{-d_{P} (\alpha)}$ where $d_P(\cdot)$ is the fading-specific large- deviations rate function.

Step 4 — Matching the outage exponent

Laplace's method evaluates the expectation: the dominating exponent is exactly the Legendre transform of $d_P(\cdot)$ at rate $r$ — which is the outage-DMT exponent $d_{P_{\mathbf{H}}}^*(r)$ for $P_{\mathbf{H}}$ . Hence $P_e \le c_2 \cdot \text{SNR}^{-d_{P_{\ntn{ch}}}^*(r)}$ .

Step 5 — Lower bound via Fano

The outage probability is a fundamental lower bound: no code at rate $R = r \log \text{SNR}$ can have error probability below $P_{\rm out}^{P_{\mathbf{H}}}(R, \text{SNR}) \doteq \text{SNR}^{-d_{P_{\ntn{ch}}}^*(r)}$ (Fano-type argument). Hence $P_e \ge c_1 \cdot \text{SNR}^{-d_{P_{\ntn{ch}}}^*(r)}$ . Combining steps 4 and 5 yields approximate universality. $\blacksquare$

,

Approximate Universality: CDA-NVD Across Fading Distributions

Achieved DMT exponent $d(r)$ of a CDA-NVD code across four fading distributions — i.i.d. Rayleigh (the baseline), i.i.d. Rician ( $K = 5$ dB), Nakagami- $m$ ( $m = 2$ ), and log-normal ( $\sigma = 4$ dB) — on an $n_t \times n_r$ channel. Observe that all four distributions produce the same DMT slope $-d^*(r) = -(n_t - r)(n_r - r)$ : approximate universality at work. The vertical offset between the curves is the SNR-independent constant $c_2$ of Definition DApproximate Universality — distribution- specific but SNR-independent.

Parameters

n_t

2

n_r

2

Approximate Universality: Same Slope, Different Fading

Animation sweeping through four fading distributions (Rayleigh, Rician, Nakagami-

m

, log-normal) and showing the Golden code's BER curve on each. The slope

d^*(r)

is invariant across distributions; only the prefactor (the SNR-independent constant) changes. This is approximate universality made visual.

Top-left: i.i.d. Rayleigh. Top-right: Rician (

K = 5

dB). Bottom-left: Nakagami-

m

(

m = 2

). Bottom-right: log-normal (

\sigma = 4

dB). All four BER curves share the same asymptotic slope

-d^*(r)

predicted by Theorem

\mathcal{F}_{\rm adm}

F_{adm}

" data-ref-type="theorem">TCDA-NVD Codes Are Approximately Universal Over

\mathcal{F}_{\rm adm}

.

Example: Verifying Approximate Universality on Rician Fading

Show that the DMT exponent of an $n_t \times n_r$ Rician channel (with LOS factor $K > 0$ ) equals the Zheng-Tse exponent for Rayleigh, i.e. $(n_t - r)(n_r - r)$ , and verify that the Golden code therefore achieves it approximately universally.

Solution

Rician eigenvalue density near zero

For Rician fading, $\mathbf{H} = \sqrt{K/(1+K)} \mathbf{H}_{\rm LOS} + \sqrt{1/(1+K)} \mathbf{H}_{\rm NLOS}$ with $\mathbf{H}_{ \rm NLOS}$ i.i.d. $\mathcal{CN}(0, 1)$ . Conditioning on the LOS component, the conditional distribution of $\mathbf{H} \mathbf{H}^{H}$ is a non-central Wishart. The eigenvalue density near zero is bounded below by that of a central Wishart (the LOS term only shifts mass away from zero).

Outage exponent unchanged

Since the small-eigenvalue behaviour of the Rician Wishart matches the central Rayleigh Wishart up to a constant factor (the LOS contribution $\mathbf{H}_{\rm LOS}$ is deterministic and does not contribute to the outage exponent), the large- deviations rate function is the same. Hence $d_{\rm Rician}^*(r) = d_{\rm Rayleigh}^*(r) = (n_t - r)(n_r - r)$ .

Apply Theorem <a href="#thm-cda-approx-universal" class="ferkans-ref" title="Theorem: CDA-NVD Codes Are Approximately Universal Over $\mathcal{F}_{\rm adm}$" data-ref-type="theorem"><span class="ferkans-ref-badge">T</span>CDA-NVD Codes Are Approximately Universal Over $\mathcal{F}_{\rm adm}$</a>

Since Rician fading lies in $\mathcal{F}_{\rm adm}$ (eigenvalue density is continuous near zero, all moments finite, no rank-deficient atom), the Golden code (a CDA-NVD code) achieves this exponent. The constant $c_2$ depends on $K$ — large $K$ gives better prefactor because the LOS term boosts the effective SNR — but the exponent is invariant. $\blacksquare$

Common Mistake: Approximate Universality Is Approximate

Mistake:

A tempting but wrong reading of approximate universality is: "The Golden code is exactly optimal on any fading channel." One might then compare BER curves at moderate SNR and be surprised that they differ by several dB.

Correction:

Approximate universality says the asymptotic DMT exponent — the slope at $\text{SNR} \to \infty$ — is invariant across $\mathcal{F}_{\rm adm}$ . The prefactor $c_2 / c_1$ can vary by several dB depending on the distribution. At moderate SNR (10–30 dB), the prefactor dominates and the BER curves visibly separate. Approximate universality is a high-SNR guarantee; it does not say that code performance is distribution-invariant at finite SNR. For finite-SNR work, one must still measure the coding-gain prefactor on each distribution of interest.

Common Mistake: DMT Optimality Is Asymptotic

Mistake:

A designer reading "the Golden code achieves the DMT curve" might conclude that at every SNR, every rate, the Golden code has the lowest error probability among all $n_t = 2$ codes. This is false.

Correction:

The DMT is a first-order asymptotic — it measures the slope $\log P_e / \log \text{SNR}$ as $\text{SNR} \to \infty$ . Two codes with the same $d^*(r)$ can differ by many dB at moderate SNR because they differ in the coding-gain constant $\delta_ {\min}$ (and higher-order prefactors). The Golden code has $\delta_{\min} = 1/5$ ; a naively-scaled CDA might have $\delta_{\min} = 1/100$ , giving a $13$ -dB SNR penalty at moderate SNR even though both are DMT-optimal. DMT is a first- order design criterion; for finite-SNR work, one must also evaluate coding gain and decode via Monte Carlo. The DMT tells you which corner $(r, d)$ to target; the coding gain tells you how far inside the corner you actually land.

,

Quick Check

Which fading distribution is not in the admissible class $\mathcal{F}_{\rm adm}$ over which CDA-NVD codes are approximately universal?

Rician with $K = 10$ dB

Log-normal shadowing with $\sigma = 8$ dB

Rank-1 deterministic LOS (specular $\mathbf{H}_{\rm LOS}$ fixed, no fading)

Correlated Rayleigh with full-rank $\mathbf{R}_t$

Correction:

Rank-1 deterministic LOS (specular

\mathbf{H}_{\rm LOS}

fixed, no fading)

Correct. A rank-1 deterministic channel has a singular atom at rank-deficient matrices: $\Pr(\mathbf{H} \text{ singular}) = 1$ . This violates condition 3 of Definition $\mathcal{F}_{\rm adm}$ $F_{adm}$ " data-ref-type="definition">DAdmissible Fading Class $\mathcal{F}_{\rm adm}$ . The code's diversity collapses because $\mathrm{rank}(\mathbf{H}) = 1$ regardless of the codeword.

Key Takeaway

Approximate universality. CDA-NVD codes achieve the DMT on every fading distribution $P_{\mathbf{H}} \in \mathcal{F}_{\rm adm}$ (continuous eigenvalue density near zero, finite moments, no rank-deficient atom), with an SNR-independent constant depending on $P_{\mathbf{H}}$ but not on $\text{SNR}$ . Operationally: one code, every fading. This is the strongest achievability result possible for linear space-time codes; it is the defining contribution of the Elia-Kumar-Caire 2006 paper.

Approximate Universality Over Arbitrary Fading

What If the Channel Is Not Rayleigh?

Definition: Approximate Universality

Definition: Admissible Fading Class Fadm\mathcal{F}_{\rm adm}Fadm​

Theorem: CDA-NVD Codes Are Approximately Universal Over Fadm\mathcal{F}_{\rm adm}Fadm​

Step 1 — Conditional PEP bound from NVD

Step 2 — Union bound with NVD

Step 3 — Average over $P_{\ntn{ch}}$

Step 4 — Matching the outage exponent

Step 5 — Lower bound via Fano

Approximate Universality: CDA-NVD Across Fading Distributions

Parameters

Approximate Universality: Same Slope, Different Fading

Example: Verifying Approximate Universality on Rician Fading

Rician eigenvalue density near zero

Outage exponent unchanged

Apply Theorem <a href="#thm-cda-approx-universal" class="ferkans-ref" title="Theorem: CDA-NVD Codes Are Approximately Universal Over $\mathcal{F}_{\rm adm}$" data-ref-type="theorem"><span class="ferkans-ref-badge">T</span>CDA-NVD Codes Are Approximately Universal Over $\mathcal{F}_{\rm adm}$</a>

Common Mistake: Approximate Universality Is Approximate

Common Mistake: DMT Optimality Is Asymptotic

Quick Check

Key Takeaway

Definition:
Approximate Universality

Definition:
Admissible Fading Class $\mathcal{F}_{\rm adm}$

Theorem: CDA-NVD Codes Are Approximately Universal Over $\mathcal{F}_{\rm adm}$