Ferkans — Interactive Telecom Tutor

From Existence to Construction

Chapter 12 closed with the Zheng-Tse theorem: the DMT curve of an $n_t \times n_r$ i.i.d. Rayleigh channel is $d^*(r) = (n_t - r)(n_r - r)$ , achievable in principle by a random Gaussian codebook of block length $L \ge n_t + n_r - 1$ . The point is that this is a non-constructive existence proof — a capacity-achieving argument that tells us what is achievable but not how to build the code. The practical question is: can we write down an explicit codebook — a structured, algebraically clean family of $n_t \times L$ complex matrices — that achieves the full Zheng-Tse curve, decodes in polynomial-in- $M$ complexity, and does not require the random-coding infinities?

For $n_t = 2$ , Belfiore, Rekaya, and Viterbo answered this in 2005 with a breathtakingly simple construction: the Golden code. Four complex information symbols $a, b, c, d$ drawn from a QAM constellation are packed into a $2 \times 2$ matrix built from the golden ratio $\theta = (1 + \sqrt{5})/2$ — the same irrational number that governs Fibonacci growth and classical æsthetic proportions, now pressed into service as the algebraic basis for a DMT-optimal space-time code. The code achieves rate $2$ complex symbols per channel use (full multiplexing on a $2 \times 2$ channel), full diversity $4$ at $r = 0$ , and — most importantly — a non- vanishing determinant $\delta_{\min} = 1/5$ that does not shrink as we grow the input constellation. This last property is the algebraic fingerprint of DMT optimality; it is what separates the Golden code from naive lattice codes whose determinants collapse at large QAM sizes.

This section defines the Golden code, proves the non-vanishing determinant via the algebraic-norm argument, and unpacks what "coding gain" means when the constellation scales. §2 generalises to arbitrary $n_t$ via cyclic division algebras; §3 presents the Perfect-codes family; §4 proves approximate universality; §5 makes the NVD property precise.

, ,

Historical Note: The 2005 Golden Code

2005

The Golden code was introduced in a 4-page IEEE Transactions on Information Theory correspondence by Jean-Claude Belfiore, Ghaya Rekaya, and Emanuele Viterbo (April 2005). At the time, Zheng and Tse's DMT theorem (2003) was only two years old and the question of whether any explicit full-rate $2 \times 2$ space-time code could achieve the full DMT curve was open. Prior full-rate designs (threaded algebraic space-time codes, the Damen-Tewfik- Belfiore code) had determinants that decayed as $1/\sqrt{M}$ with constellation size — acceptable at small $M$ , fatal at large $M$ .

Belfiore-Rekaya-Viterbo's key move was to use the number field $\mathbb{Q}(j, \sqrt{5}) = \mathbb{Q}(j, \theta)$ — a degree-2 extension of the Gaussian rationals $\mathbb{Q}(j)$ — and exploit the fact that the algebraic norm on this field satisfies $N(a + \theta b) = (a + \theta b)(a + \bar\theta b) = a^2 + ab - b^2 \in \mathbb{Z}[j]$ , a non-zero Gaussian integer whenever $a + \theta b \neq 0$ . The minimum absolute value of a non-zero Gaussian integer is $1$ — hence the determinant of any non-zero codeword difference is bounded below by $1/|5|$ , and the $1/5$ factor comes from the overall normalisation chosen to make the average codeword energy equal to the transmit power.

The paper christened the construction "the Golden code" in a reference to the golden ratio $\theta$ that sits at its algebraic heart. It became the most-cited space-time code in wireless communications, adopted as a reference design in the DVB-NGH mobile TV standard (2012) and used as a benchmark for every subsequent DMT-optimal construction. Its fame is deserved: the Golden code is, in Tse and Viswanath's phrase, "as close to a perfect space-time code as one can hope to find."

Definition:
The Golden Code

Let $\theta = (1 + \sqrt{5})/2$ (the golden ratio) with Galois conjugate $\bar\theta = 1 - \theta = (1 - \sqrt{5})/2$ . Let $\alpha = 1 + j\bar\theta$ and $\bar\alpha = 1 + j\theta$ be the two units of the Golden-code construction. Let $a, b, c, d$ be four information symbols drawn from a scaled QAM constellation $\mathcal{Q}_M \subset \mathbb{Z}[j]$ of size $M$ .

The Golden codeword matrix is the $2 \times 2$ complex matrix $\ntn{X}_{\rm Gold} = \frac{1}{\sqrt{5}} \begin{pmatrix} \alpha (a + \theta b) & \alpha (c + \theta d) \\ j \, \bar\alpha (c + \bar\theta d) & \bar\alpha (a + \bar\theta b) \end{pmatrix}.$

The factor $1/\sqrt{5}$ is the energy normalisation; $|\alpha|^2 = 1 + \bar\theta^2$ and $|\bar\alpha|^2 = 1 + \theta^2$ satisfy $|\alpha|^2 + |\bar\alpha|^2 = 5$ . The golden code transmits two complex symbols per channel use over $n_t = 2$ antennas during $L = 2$ channel uses — it is full-rate ( $\log_2 M^4 / 2 = 2 \log_2 M$ bits per channel use).

Intuitively, what happens is this. The top row is built from the number-field element $a + \theta b$ on the first channel use and $c + \theta d$ on the second. The bottom row is the Galois conjugate of the top row — $a + \theta b$ becomes $a + \bar\theta b$ — which provides the diversity. The factor $j$ in the bottom- left entry is what breaks the "transposed" symmetry and lets the determinant stay away from zero; the units $\alpha, \bar\alpha$ are what make the columns and rows orthogonal on average so the code has uniform average energy across antennas.

,

Example: An Explicit Golden Codeword

Compute the Golden codeword $\ntn{X}_{\rm Gold}$ for information symbols $(a, b, c, d) = (1, j, -1, 1 + j)$ and verify that $|\det(\ntn{X}_{\rm Gold})|^2 \ne 0$ .

Solution

Compute $a + \theta b$ and $a + \bar\theta b$

With $\theta \approx 1.618$ , $\bar\theta \approx -0.618$ : $a + \theta b = 1 + j\theta \approx 1 + 1.618 j$ ; $a + \bar\theta b = 1 + j\bar\theta \approx 1 - 0.618 j$ .

Compute $c + \theta d$ and $c + \bar\theta d$

$c + \theta d = -1 + \theta(1 + j) = (\theta - 1) + \theta j \approx 0.618 + 1.618 j$ (using $\theta - 1 = \bar\theta^{-1} = 1/\theta \approx 0.618$ ); $c + \bar\theta d = -1 + \bar\theta (1 + j) = (\bar\theta - 1) + \bar\theta j \approx -1.618 - 0.618 j$ .

Compute $\alpha$ and $\bar\alpha$

$\alpha = 1 + j\bar\theta \approx 1 - 0.618 j$ ; $\bar\alpha = 1 + j\theta \approx 1 + 1.618 j$ .

Assemble the $2 \times 2$ matrix

$\ntn{X}_{\rm Gold} = \tfrac{1}{\sqrt{5}} \begin{pmatrix} \alpha (1 + j\theta) & \alpha ((\theta - 1) + \theta j) \\ j \bar\alpha ((\bar\theta - 1) + \bar\theta j) & \bar\alpha (1 + j\bar\theta) \end{pmatrix}.$ All four entries are non-zero complex numbers.

Verify $\det \ne 0$

The determinant lies in $\tfrac{1}{5} \mathbb{Z}[j]$ (the $1/\sqrt{5}$ normalisation gives $1/5$ on the determinant of the $2\times2$ ). Explicit computation gives a non-zero Gaussian integer divided by $5$ — for these inputs, $|\det(\ntn{X}_{\rm Gold})|^2 \approx 0.48 \ne 0$ , confirming full rank. The next theorem shows this is not a coincidence: the minimum over all codeword PAIRS of this quantity is exactly $1/5$ , independent of QAM size. $\blacksquare$

Golden Codeword Entries Viewed as a $\mathbb{C}$ -Valued Constellation

Each information quadruple $(a, b, c, d) \in \mathcal{Q}_M^4$ produces a $2 \times 2$ complex Golden codeword with four complex-valued entries. This plot scatters those four entries in the complex plane for the full constellation $\mathcal{Q}_M^4$ ( $M = 4, 16, 64$ ) and reveals the dense lattice structure of the Golden-code output — the algebraic substrate that makes the non-vanishing-determinant property work.

Parameters

Input QAM size

M

Golden Code: Sphere Decoding of a $2 \times 2$ Codeword

Animation of the Golden-code receiver: an information quadruple

(a, b, c, d)

is encoded into

\ntn{X}_{\rm Gold}

, corrupted by the

2 \times 2

i.i.d. Rayleigh channel

\mathbf{H}

and AWGN

\mathbf{w}

, and decoded by the Viterbo-Boutros sphere decoder. The sphere decoder searches through the lattice cosets

\{\mathbf{H} \ntn{X} : \ntn{X} \in \mathcal{C}\}

inside a shrinking radius until the nearest lattice point is found — the ML estimate.

The sphere radius (red circle) shrinks after each candidate examined; the ML codeword (green) is the unique lattice point inside the final sphere. Complexity scales as

O(M^{n_t^2/2}) = O(M^2)

for

n_t = 2

on average.

Theorem: Non-Vanishing Determinant of the Golden Code

Let $\ntn{X}_{\rm Gold}, \hat{\ntn{X}}_{\rm Gold}$ be two distinct Golden codewords built from information quadruples $(a, b, c, d)$ and $(\hat a, \hat b, \hat c, \hat d)$ in $\mathcal{Q}_M^4 \subset \mathbb{Z}[j]^4$ . Let $\boldsymbol{\Delta} = \ntn{X}_{\rm Gold} - \hat{\ntn{X}}_{\rm Gold}$ . Then $|\det(\boldsymbol{\Delta})|^2 \;\ge\; \frac{1}{5},$ with the bound achieved; equivalently $\delta_{\min} = 1/5$ . The bound is independent of the constellation size $M$ — it does not shrink as $M$ grows. Hence the Golden code has the non-vanishing-determinant (NVD) property.

The determinant of a Golden codeword factors through the algebraic norm on the number field $\mathbb{Q}(\theta)$ . Norms of non-zero algebraic integers are non-zero rational integers; the minimum absolute value of a non-zero Gaussian integer is $1$ . The $1/5$ prefactor is the square of the $1/\sqrt{5}$ energy normalisation. Crucially, scaling to larger QAM only changes the magnitudes within $\mathbb{Z}[j]$ — it never produces smaller non-zero integers — so the bound survives at every $M$ .

Show Hint

Express $\det(\ntn{X}_{\rm Gold})$ in terms of $(a + \theta b)(a + \bar\theta b) - (c + \theta d)(c + \bar\theta d) \cdot$ (unit factors).

Recognise the first product as $N_{\mathbb{Q}(\theta)/\mathbb{Q}}(a + \theta b) = a^2 + ab - b^2$ and similarly for the second.

Show that $|\alpha|^2 |\bar\alpha|^2 = (1 + \bar\theta^2)(1 + \theta^2) = 5$ .

Conclude that $|\det(\boldsymbol{\Delta})|^2 = \tfrac{1}{5} |N(a - \hat a + \theta(b - \hat b)) - j N(c - \hat c + \theta(d - \hat d))|^2$ — a non-zero Gaussian integer divided by $5$ .

Proof

Step 1 — Expand the determinant

Direct computation of the $2 \times 2$ determinant of $\ntn{X}_{\rm Gold}$ gives $\det(\ntn{X}_{\rm Gold}) = \tfrac{1}{5} \big[ \alpha \bar\alpha (a + \theta b)(a + \bar\theta b) - j \alpha \bar\alpha (c + \theta d)(c + \bar\theta d) \big].$ The factor $\alpha \bar\alpha$ is constant; the two quadratic forms are precisely the algebraic norms of the number-field elements.

Step 2 — Identify the algebraic norms

The extension $\mathbb{Q}(\theta)/\mathbb{Q}$ has Galois generator $\sigma: \theta \mapsto \bar\theta$ , so for any $x, y \in \mathbb{Q}$ , $N_{\mathbb{Q}(\theta)/\mathbb{Q}}(x + \theta y) = (x + \theta y)(x + \bar\theta y) = x^2 + xy(\theta + \bar\theta) + \theta\bar\theta y^2 = x^2 + xy - y^2$ (using $\theta + \bar\theta = 1$ and $\theta\bar\theta = -1$ ). For $x, y \in \mathbb{Z}[j]$ , this norm is a Gaussian integer.

Step 3 — Compute $|\alpha|^2 |\bar\alpha|^2$

Step 4 — Bound the determinant difference

Setting $u = N(a - \hat a + \theta (b - \hat b)) \in \mathbb{Z}[j]$ and $v = N(c - \hat c + \theta (d - \hat d)) \in \mathbb{Z}[j]$ , $\det(\boldsymbol{\Delta}) = \tfrac{1}{5} \alpha \bar\alpha (u - j v)$ . Hence $|\det(\boldsymbol{\Delta})|^2 = \tfrac{|\alpha|^2 |\bar\alpha|^2}{25} |u - j v|^2 = \tfrac{1}{5} |u - j v|^2.$ Since $u, v \in \mathbb{Z}[j]$ and the codewords differ, at least one of $u, v$ is non-zero, so $|u - j v|^2 \ge 1$ (it is a non-zero Gaussian integer). Therefore $|\det(\boldsymbol{\Delta})|^2 \ge 1/5$ , with equality when $(u, v) = (1, 0)$ or $(0, 1)$ (achieved for adjacent QAM points). $\blacksquare$

,

What the Golden Code Achieves

The Golden code is full-rate ( $r_{\max} = 2$ , the channel's spatial degrees of freedom), full-diversity ( $d_{\max} = 4 = n_t n_r$ for $2 \times 2$ ), and has non-vanishing determinant ( $\delta_{\min} = 1/5$ at every $M$ ). Combined with the rank and determinant criteria of Chapter 11 and the Zheng-Tse DMT upper bound of Chapter 12, these three properties imply that the Golden code achieves the Zheng-Tse DMT curve $d^*(r) = (2 - r)(2 - r)$ for all $r \in [0, 2]$ . It is the first explicit $2 \times 2$ construction with this property, and — for its dimension — the benchmark against which all subsequent space-time codes are measured.

,

Key Takeaway

The Golden Code. $\ntn{X}_{\rm Gold} = \tfrac{1}{\sqrt{5}} \begin{pmatrix} \alpha(a + \theta b) & \alpha(c + \theta d) \\ j\bar\alpha(c + \bar\theta d) & \bar\alpha(a + \bar\theta b) \end{pmatrix}$ over $\mathbb{Q}(j, \theta)$ is full-rate, full- diversity, and has $\delta_{\min} = 1/5$ independent of $M$ . It is the first explicit DMT-optimal code for $n_t = 2$ . The next section generalises this construction to arbitrary dimensions via cyclic division algebras.

The Golden Code for 2×22 \times 22×2 MIMO