Ferkans — Interactive Telecom Tutor

The Algebraic Fingerprint of DMT Optimality

The thread running through §1–§4 is the non-vanishing-determinant (NVD) property: the minimum codeword-pair determinant $\delta_{\min}$ is bounded below by a positive constant that does not shrink with the input QAM size $M$ . This property was empirical in the Golden code (a consequence of the algebraic-norm identity), formal in the CDA framework (a lattice condition within the division algebra), and pivotal in the approximate- universality proof (it absorbed the codebook cardinality into the PEP bound without letting the $M$ -dependence break DMT optimality).

This section makes NVD central. We define it precisely, prove that NVD is necessary for DMT optimality at $r \to r_{\max}$ (not merely sufficient), and discuss how practical codes either achieve it (CDA-NVD) or fail to achieve it (threaded-algebraic codes, naively-scaled Alamouti concatenations). The NVD characterisation is the final piece of the DMT-optimality puzzle and the design principle that ties §1–§4 together.

Definition:
Non-Vanishing Determinant (NVD)

A family of $n_t \times n_t$ space-time codes $\{\mathcal{C}_M\}$ with information symbols from QAM of size $M$ has the non- vanishing-determinant property if $\delta_{\min}(\mathcal{C}_M) \;\triangleq\; \min_{\ntn{X} \neq \hat{\ntn{X}} \in \mathcal{C}_M} |\det(\boldsymbol{\Delta} \boldsymbol{\Delta}^H)| \;\ge\; \delta_0 > 0$ for all $M$ , where $\delta_0$ is an $M$ -independent positive constant. Equivalently, the coding-gain sequence $\{\delta_ {\min}(\mathcal{C}_M)\}_M$ does not decay to zero as $M \to \infty$ .

The word "non-vanishing" is literal: the determinant does not vanish as we grow the constellation. Contrast this with a naive scaling scheme where the code is designed at small $M$ and then rescaled: the determinant of the difference between two adjacent codewords shrinks like $1/M$ under uniform scaling, which would give $\delta_{\min} \to 0$ as $M \to \infty$ .

,

Theorem: NVD Is Necessary for DMT-Optimality at $r \to r_{\max}$

Let $\{\mathcal{C}_M\}$ be a family of $n_t \times n_t$ space-time codes with full spatial diversity and information rate $R_M (\text{SNR}) = r \log_2 \text{SNR}$ . If the family achieves the Zheng-Tse DMT exponent $d^*(r) = (n_t - r)(n_r - r)$ at every $r \in (0, \min(n_t, n_r))$ — and in particular in a neighbourhood of $r = r_{\max}$ — then the code family must have the NVD property. Equivalently: any code family with $\delta_{\min} (\mathcal{C}_M) \to 0$ as $M \to \infty$ loses DMT optimality at high multiplexing gain.

Suppose $\delta_{\min}(\mathcal{C}_M) \doteq \text{SNR}^{-\epsilon}$ for some $\epsilon > 0$ (a polynomial decay with SNR via the relation $M \doteq \text{SNR}^{r n_t / 2}$ ). Then the union bound over the $\doteq \text{SNR}^{2 r n_t}$ codeword pairs contains a PEP prefactor $\delta_{\min}^{-n_r} \doteq \text{SNR}^{\epsilon n_r}$ which adds to the exponent — not in the good direction. The DMT exponent degrades by $\epsilon n_r$ , losing optimality. The converse statement is that only $\epsilon = 0$ preserves the Zheng-Tse exponent at high $r$ — which is exactly the NVD condition.

Show Hint

Write $M = \text{SNR}^{r n_t / 2}$ to convert between constellation size and SNR at multiplexing gain $r$ .

Suppose $\delta_{\min} \doteq \text{SNR}^{-\epsilon}$ with $\epsilon > 0$ . The PEP bound gains a factor $\delta_{\min}^{-n_r} \doteq \text{SNR}^{\epsilon n_r}$ .

Union-bound the error probability over $\text{SNR}^{2 r n_t}$ codeword pairs and show that the degraded exponent $d^*(r) - \epsilon n_r$ is strictly below the Zheng-Tse bound when $\epsilon > 0$ .

Proof

Step 1 — Parameterise constellation size in SNR

A code at multiplexing gain $r$ has rate $R = r \log_2 \text{SNR}$ , hence $M = 2^{R / n_t} \doteq \text{SNR}^{r / n_t}$ per information symbol; the codebook has $M^{n_t^2} \doteq \text{SNR}^{r n_t}$ codewords.

Step 2 — Suppose NVD fails

Assume for contradiction that $\delta_{\min}(\mathcal{C}_M) \to 0$ as $M \to \infty$ . By scaling arguments (the code is full-diversity so $\delta_{\min}$ does not vanish at any finite $M$ ), there is $\epsilon > 0$ such that $\delta_{\min} (\mathcal{C}_M) \doteq M^{-2 \epsilon} \doteq \text{SNR}^{-2 \epsilon r / n_t}$ .

Step 3 — PEP with degraded determinant

For any pair achieving the minimum, the conditional PEP is $\exp(-\tfrac{\text{SNR}}{4 n_t} \cdot \mathrm{tr}(\mathbf{H} \boldsymbol{\Delta}\boldsymbol{\Delta}^H \mathbf{H}^{H}))$ . Averaged over Rayleigh $\mathbf{H}$ and using the eigenvalues $\lambda_i (\boldsymbol{\Delta}\boldsymbol{\Delta}^H) \ge \delta_{\min} ^{1/n_t}$ , the PEP is $\doteq \text{SNR}^{-n_r} \delta_{\min} ^{-n_r/n_t} \doteq \text{SNR}^{-n_r + 2 \epsilon r / n_t \cdot n_r/n_t}$ for the worst pair.

Step 4 — Union bound over codebook

Union-bounding over $\text{SNR}^{2 r n_t}$ pairs: $P_e \doteq \text{SNR}^{2 r n_t} \cdot \text{SNR}^{-n_r + 2 \epsilon r n_r / n_t^2}$ , giving the exponent $-d(r) = -n_r + 2 r n_t + 2 \epsilon r n_r / n_t^2$ . For Zheng-Tse optimality at the corner $r \to r_{\max}$ , this must equal $-(n_t - r)(n_r - r)$ . A direct comparison shows this requires $\epsilon = 0$ .

Step 5 — Conclude NVD is necessary

Hence $\delta_{\min}$ cannot decay polynomially in $M$ ; NVD is necessary. The Golden code and the CDA-NVD family satisfy $\epsilon = 0$ , as do the Perfect codes. The threaded-algebraic codes of Damen-Tewfik-Belfiore (2002) have $\epsilon > 0$ and lose DMT optimality at high $r$ — a gap that motivated the Perfect-codes construction in the first place. $\blacksquare$

,

Minimum Codeword-Pair Determinant $\delta_{\min}$ vs QAM Size $M$

Plot of $\delta_{\min}(\mathcal{C}_M)$ as a function of input QAM size $M$ for three $n_t = 2$ code families: the Golden code (NVD: $\delta_{\min} = 1/5$ , flat line), a naively-scaled integer lattice code (non-NVD: $\delta_{\min} \propto 1/M$ , sharp drop), and a threaded-algebraic code (partial NVD: $\delta_{\min} \propto 1/\sqrt{M}$ , slower drop). The flat line is the algebraic fingerprint of DMT optimality.

Parameters

Verify NVD for a Proposed CDA Code

Complexity:

O(R_{\max}^{2 n_t^2})

for the enumeration;

O((\log M_{\max})^{2 n_t^2})

with basis-reduction preprocessing (LLL).

Input: Cyclic algebra

\mathcal{A}(F, K, \sigma, \gamma)

with

lattice representative

\mathcal{I} \subset \mathcal{A}

of

\mathcal{O}_K

-rank

n_t

; target confidence

M_{\max}

.

Output: NVD verification result + numerical

\delta_0

.

1. Verify division algebra: Check that

\gamma \notin N_{K/F}(K^*)

, i.e. no element of

K

has norm

\gamma

.

If fails,

\mathcal{A}

is not division

\to

reject.

2. Compute the Gram matrix

G

of the lattice

\mathcal{I}

in its

F

-basis. Check

\det G \neq 0

(full rank).

3. Enumerate small non-zero lattice vectors:

for each

a \in \mathcal{I}

with

\|a\|_K \le R_{\max}

(a

radius chosen large enough to guarantee the minimum

determinant is seen):

4. Compute

\lambda(a) \in M_{n_t}(K)

(regular rep).

5. Compute

\delta(a) = |\det(\lambda(a) \lambda(a)^H)|

.

6. Track

\delta_0 \leftarrow \min(\delta_0, \delta(a))

.

7. Test scaling invariance: Repeat steps 3–6 at

M = 4, 16, 64, 256, \ldots, M_{\max}

. If

\delta_0

is constant

(modulo numerical precision), NVD holds at confidence

M_{\max}

. If

\delta_0

decreases with

M

, NVD fails.

8. Return

(\text{NVD: yes/no}, \delta_0)

.

Step 1 is a number-theoretic check (a finite algorithm using the class group of $K$ ). Step 3 is the expensive step; in practice it is replaced by an LLL-reduction of the lattice $\mathcal{I}$ followed by evaluation on the reduced basis. For the Golden code, this algorithm produces $\delta_0 = 1/5$ exactly after a single iteration.

⚠️Engineering Note

CDA Codes vs 5G NR and WLAN MIMO Precoding

The CDA-NVD framework of this chapter gives the theoretically optimal linear space-time code for every $(n_t, n_r)$ . In practice, 5G NR (Release 15+) and WLAN 802.11ax/be do not implement full CDA codes. Instead they use:

Codebook-based precoding: The transmitter picks a precoder from a finite codebook (Type-I for low complexity, Type-II for high performance in 5G NR), often based on DFT columns or Grassmannian designs. This precoding reduces MIMO to a small number of parallel streams rather than coding across the full matrix.
BICM + LDPC or polar outer codes: The coded bits are interleaved and mapped to QAM symbols on each stream independently; the channel code handles the diversity (over the BICM bit channels), not the space-time code.
MIMO + HARQ rather than full-diversity STBC: When the channel is in a deep fade, the receiver requests a retransmission (next chapter); the ARQ-DMT gives higher multiplexing at equivalent diversity.

Why the practical detour? Three reasons: (i) decoder complexity — sphere decoding of a $4 \times 4$ CDA code with $M = 64$ is $O(64^{8}) = 10^{14}$ candidates on average, impractical for real-time decoders; (ii) rate flexibility — CDA codes are rate- rigid ( $n_t$ streams, full), whereas codebook precoding allows rank adaptation from $1$ to $n_t$ ; (iii) HARQ integration — CDA codes do not naturally support incremental redundancy, while LDPC codes with rate-matching do. DVB-NGH (2012) is the rare standard that adopted a CDA code (the Golden code) — and it used it as an optional mode, not default.

For a practitioner, the CDA framework is benchmark: it tells you how well any linear scheme could do at high SNR on any fading distribution. Your actual code is likely simpler, but its DMT gap to the CDA benchmark is your optimisation target.

Practical Constraints

•
5G NR (Rel-15+) uses Type-I/Type-II codebook precoding (DFT + power allocation), not CDA
•
WLAN 802.11ax/be uses spatial mapping matrices similar to 5G Type-I
•
DVB-NGH (2012) is the rare standard that adopted Golden code (optional mode)
•
Sphere-decoder complexity $O(M^{n_t^2/2})$ is prohibitive for $n_t \ge 4$ at common QAM sizes

📋 Ref: 3GPP TS 38.211 §6.3.1; 3GPP TS 38.214 §5.2; DVB-NGH (ETSI EN 303 105-1)

,

Historical Note: The Oggier-Rekaya-Belfiore-Viterbo Perfect Codes (2006)

2006

Six months after the Golden code paper, Frederique Oggier, Ghaya Rekaya, Jean-Claude Belfiore, and Emanuele Viterbo (the last three being the Golden-code authors) published "Perfect Space- Time Block Codes" in IEEE Transactions on Information Theory (September 2006). The paper unified the Golden code into a family of four — $n_t = 2, 3, 4, 6$ — through the strict Perfect-code definition (full rate, full diversity + NVD, uniform average energy, cubic shaping). The $n_t = 2$ case is the Golden code; $n_t = 3$ uses $\mathbb{Q}(\zeta_3)$ and a degree-3 cyclic extension; $n_t = 4$ uses $\mathbb{Q}(j)$ and a degree-4 cyclic extension; $n_t = 6$ uses $\mathbb{Q} (\zeta_3)$ and a degree-6 cyclic extension.

Three months later, Elia, K. R. Kumar, Pawar, P. V. Kumar, Lu, and Caire generalised the construction to any $n_t$ at the cost of sacrificing uniform average energy and cubic shaping — giving a CDA-NVD code family that achieves the DMT in every dimension. Together, these two 2006 papers closed the DMT- optimal-construction problem for linear space-time block codes.

The Perfect-codes paper was awarded the IEEE Information Theory Society Paper Award in 2008. The Elia-Kumar-Caire paper was cited by the DVB-NGH standard's space-time-coding section (2012) as the theoretical foundation. Both constructions remain the benchmark against which all subsequent MIMO code designs — LAST codes (Ch. 17), compute-and-forward (Ch. 18), and the non-coherent space-time codes of Chapter 22 — are measured.

Quick Check

A code family has full diversity $n_t n_r$ at $r = 0$ but $\delta_{\min}(\mathcal{C}_M) \propto 1/M$ . At what multiplexing gain $r$ does it first lose DMT optimality?

At $r = 0$

As soon as $r > 0$

Only at $r = r_{\max}$

Never — full diversity at $r = 0$ is enough

Correction:

As soon as

r > 0

Correct. Theorem $r \to r_{\max}$ $r \to r_{m a x}$ " data-ref-type="theorem">TNVD Is Necessary for DMT-Optimality at $r \to r_{\max}$ shows that any polynomial decay of $\delta_{\min}$ with $M$ degrades the DMT exponent by $\epsilon n_r > 0$ for all $r > 0$ . The Golden code's flat $\delta_{\min} = 1/5$ preserves the full curve.

Non-Vanishing Determinant (NVD)

A property of a space-time code family: the minimum codeword-pair determinant $\delta_{\min}$ is bounded below by a positive constant $\delta_0$ that does not depend on the input QAM size $M$ . NVD is necessary and sufficient for a full-diversity CDA code to achieve the Zheng-Tse DMT curve and — under additional regularity on the fading — to be approximately universal.

Cyclic Division Algebra (CDA)

An $F$ -algebra $\mathcal{A}(F, K, \sigma, \gamma)$ of dimension $n^2$ over the base field $F$ , where $K/F$ is a cyclic Galois extension of degree $n$ , $\sigma$ is the Galois generator, and $\gamma \in F^*$ is a non-norm element. Every non-zero element of $\mathcal{A}$ is invertible (division), which through the regular representation gives a full-rank $n \times n$ codeword matrix.

Approximate Universality

A code family that achieves the DMT exponent of every fading distribution in the admissible class $\mathcal{F}_{\rm adm}$ — not just the one it was designed for — up to SNR-independent constants. CDA-NVD codes are approximately universal over $\mathcal{F}_{\rm adm}$ (Tavildar-Viswanath 2006, Elia-Kumar- Caire 2006).

Golden Code

The $n_t = 2$ Perfect space-time code, constructed over the cyclic division algebra $\mathbb{Q}(j, \theta)/\mathbb{Q}(j)$ where $\theta = (1 + \sqrt{5})/2$ is the golden ratio. Full rate, full diversity, NVD with $\delta_{\min} = 1/5$ . Introduced by Belfiore-Rekaya-Viterbo (2005); the first explicit DMT-optimal $2 \times 2$ space-time code.

Division Algebra

A non-commutative ring in which every non-zero element has a two-sided multiplicative inverse. The classical example is the Hamilton quaternions $\mathbb{H}$ . The cyclic algebra $\mathcal{A} (F, K, \sigma, \gamma)$ is a division algebra iff $\gamma$ (and its powers up to $n - 1$ ) are non-norms in $F$ .

Related: CDA Codewords Are Full Rank

Why This Matters: From CDA Codes to LAST Codes

CDA-NVD codes achieve the Zheng-Tse DMT for square $(n_t \times n_t)$ MIMO but are less natural for asymmetric channels $n_r \ne n_t$ — the cyclic extension degree is fixed at $n_t$ , and increasing $n_r$ only brings coding-gain improvements via the algebraic receiver structure. For arbitrary $(n_t, n_r)$ with possibly asymmetric dimensions, Chapter 17 introduces LAST codes (Lattice Space-Time codes) of El Gamal, Caire, and Damen (2004), which generalise the CDA idea to lattice coding and achieve the full Zheng-Tse DMT with a universal MMSE-GDFE receiver — a different but equally powerful construction strategy.

Chapter 14 meanwhile extends the DMT framework in the opposite direction: toward ARQ-enabled MIMO, where the ARQ-DMT curve strictly exceeds the Zheng-Tse curve at the cost of feedback. The CommIT contribution of Chapter 14 — El Gamal- Caire-Damen 2006 IR-LAST codes — combines the ARQ-DMT with lattice space-time codes, a synthesis of the CDA and lattice threads that runs through Chapters 13–17 of this book.

The Non-Vanishing-Determinant Property