Ferkans — Interactive Telecom Tutor

Generalising Alamouti: The OSTBC Idea

Alamouti's magic is a single algebraic identity: $\mathbf{X}_A\mathbf{X}_A^H \propto \mathbf{I}_2$ . The question that Tarokh, Jafarkhani, and Calderbank asked in 1999 was: for what $n_t$ can we construct a codeword matrix with the same orthogonality? If the answer is positive, we immediately get a rate- $K/T$ code with full diversity $n_t n_r$ and linear matched-filter decoding — by the exact same argument as Thm. $2 n_r$ $2 n_{r}$ with Linear Decoding" data-ref-type="theorem">TAlamouti Achieves Full Diversity $2 n_r$ with Linear Decoding.

The remarkable story of this chapter is that for the real case — real-valued symbols drawn from a PAM constellation — the question has a beautiful 19th-century answer: orthogonal designs exist at rate 1 for every $n_t$ that is a power of 2, via the Hurwitz-Radon-Eckmann family. For the complex case — the practical one, with QAM or PSK inputs — the rate must drop below 1 as soon as $n_t > 2$ , and the Liang-Tarokh 2003 bound (§3) gives the exact rate ceiling. This section constructs OSTBCs via dispersion matrices, proves full diversity, and works the two canonical examples: rate- $1/2$ for $n_t = 3, 4$ (with complex symbols) and rate- $3/4$ for $n_t = 3, 4$ (Tarokh-Jafarkhani-Calderbank 1999).

The central concept is the dispersion matrix expansion: every linear STBC can be written as $\mathbf{X} = \sum_q (\alpha_q \mathbf{A}_q + j\beta_q \mathbf{B}_q)$ for fixed complex matrices $\mathbf{A}_q, \mathbf{B}_q$ and real scalars $\alpha_q, \beta_q$ (the real and imaginary parts of the $q$ -th information symbol). OSTBCs are the subclass in which the dispersion matrices satisfy Hurwitz-Radon-Eckmann orthogonality relations; LDCs (§5) relax these relations to any linear dispersion structure.

Definition:
Dispersion Matrix Expansion of a Linear STBC

A space-time block code $\mathcal{C}$ is linear in its $K$ complex information symbols $s_q = \alpha_q + j\beta_q, q=1, \ldots, K$ if every codeword can be written as $\mathbf{X} \;=\; \sum_{q=1}^{K} \bigl(\alpha_q \mathbf{A}_q + j\beta_q \mathbf{B}_q\bigr),$ where $\{\mathbf{A}_q, \mathbf{B}_q\}_{q=1}^K \subset \mathbb{C}^{n_t \times T}$ is a fixed set of $Q = 2K$ dispersion matrices that depend only on the code (not on the data). Equivalently, every codeword lies in the real span of the $Q$ matrices $\{\mathbf{A}_q\} \cup \{j\mathbf{B}_q\}$ .

The code is fully specified by the tuple $(\{\mathbf{A}_q\}, \{\mathbf{B}_q\}, \mathcal{X})$ where $\mathcal{X}$ is the underlying constellation of the $s_q$ 's.

The dispersion-matrix representation is the most general linear structure an STBC can have. Every linear STBC is determined by its $2K$ matrices — Alamouti, OSTBCs, QOSTBCs, V-BLAST, LDCs all fit into this single framework and differ only in the algebra satisfied by the $\{\mathbf{A}_q, \mathbf{B}_q\}$ . For Alamouti, $K = 2$ and the four dispersion matrices are the real/imaginary parts of $\mathbf{A}_1 = \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}, \mathbf{B}_1 = \begin{pmatrix} 1 & 0 \\ 0 & -1 \end{pmatrix}, \mathbf{A}_2 = \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix}, \mathbf{B}_2 = \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix}$ .

,

Definition:
Orthogonal Space-Time Block Code (OSTBC)

A linear STBC $\mathcal{C}$ with dispersion matrices $\{\mathbf{A}_q, \mathbf{B}_q\}$ is an orthogonal space-time block code (OSTBC) if every codeword satisfies $\mathbf{X}\mathbf{X}^H \;=\; \bigl(|s_1|^2 + |s_2|^2 + \cdots + |s_K|^2\bigr)\,\mathbf{I}_{n_t}$ for every $(s_1, \ldots, s_K) \in \mathcal{X}^K$ . Equivalently, the dispersion matrices satisfy the Hurwitz-Radon-Eckmann relations: $\mathbf{A}_p \mathbf{A}_q^H + \mathbf{A}_q \mathbf{A}_p^H \;=\; 2 \delta_{pq}\,\mathbf{I}_{n_t},$ $\mathbf{B}_p \mathbf{B}_q^H + \mathbf{B}_q \mathbf{B}_p^H \;=\; 2 \delta_{pq}\,\mathbf{I}_{n_t},$ $\mathbf{A}_p \mathbf{B}_q^H + \mathbf{B}_q \mathbf{A}_p^H \;=\; \mathbf{0} \quad \text{for all } p, q.$

These three relations are exactly what is needed to make the cross terms in $\mathbf{X}\mathbf{X}^H$ vanish after expansion. They are called Hurwitz-Radon-Eckmann because they are the anti-commutation relations satisfied by the generators of a real Clifford algebra — the same algebra that underlies the classical sums-of-squares identities of Hurwitz (1898) and Radon (1922). The maximum number of such matrices of size $n$ is the Hurwitz-Radon number $\rho(n)$ , which is $\le n$ for all $n$ and equals $n$ only for $n \in \{1, 2, 4, 8\}$ . This is the deep algebraic reason why complex OSTBCs lose rate beyond $n_t = 2$ — see §3.

,

Theorem: Every OSTBC Achieves Full Diversity $n_t n_r$

Let $\mathcal{C}$ be an OSTBC with $K$ information symbols drawn from a constellation $\mathcal{X}$ and block length $T$ over the quasi-static i.i.d. Rayleigh MIMO channel. Then:

(a) The ML decoder decouples into $K$ independent scalar decisions, one per information symbol $s_q$ : for each $q$ , the sufficient statistic is $\tilde y_q = \|\mathbf{H}\|_F^2 \cdot s_q + \tilde w_q$ with $\tilde w_q$ white Gaussian.

(b) Each scalar decision has effective SNR $\text{SNR}_{\mathrm{eff}} = (\text{SNR}/n_t)\sum_{i,j}|h_{i,j}|^2$ — a sum of $n_t n_r$ Rayleigh magnitudes.

(c) Consequently, $\mathcal{C}$ achieves diversity $d = n_t n_r$ — the full MIMO diversity order.

This theorem is a one-paragraph lift of the Alamouti argument. The OSTBC orthogonality $\mathbf{X}\mathbf{X}^H \propto \mathbf{I}$ means that after any linear receiver processing, each information symbol appears with a noise that is independent of the other symbols and whose variance is scaled by $\|\mathbf{H}\|_F^2$ — a sum of $n_t n_r$ i.i.d. squared Rayleigh magnitudes. The scaling $1/n_t$ is the transmit-power-splitting penalty (generalising the 3 dB of Alamouti to $10\log_{10} n_t$ ).

Show Hint

Compute $\|\mathbf{Y} - \mathbf{H}\mathbf{X}\|_F^2$ and expand $\mathbf{X}\mathbf{X}^H = \|\mathbf{s}\|^2 \mathbf{I}$ .

Use the cross-term Hurwitz-Radon-Eckmann relations to separate the metric into a sum of $K$ per-symbol metrics.

Deduce that each metric depends on $s_q$ alone; minimising over $s_q$ is a scalar matched-filter decision.

Proof

Expand the ML metric

The ML decoder maximises $-\|\mathbf{Y} - \mathbf{H}\mathbf{X}\|_F^2$ over $(s_1, \ldots, s_K) \in \mathcal{X}^K$ . Expand: $\|\mathbf{Y} - \mathbf{H}\mathbf{X}\|_F^2 \;=\; \|\mathbf{Y}\|_F^2 - 2\mathrm{Re}\,\mathrm{tr}(\mathbf{Y}^H\mathbf{H}\mathbf{X}) + \mathrm{tr}(\mathbf{X}^H\mathbf{H}^{H}\mathbf{H}\mathbf{X}).$ The first term is constant in $(s_q)$ and drops out. The third term expands via the dispersion matrix representation $\mathbf{X} = \sum_q (\alpha_q \mathbf{A}_q + j\beta_q\mathbf{B}_q)$ .

Orthogonality kills cross terms

Using the Hurwitz-Radon-Eckmann relations, the third term simplifies to $\mathrm{tr}(\mathbf{X}^H\mathbf{H}^{H}\mathbf{H}\mathbf{X}) \;=\; \|\mathbf{H}\|_F^2 \sum_{q=1}^K |s_q|^2 + \text{cross terms that vanish}.$ Concretely: the cross-term coefficients are of the form $\alpha_p\alpha_q(\mathbf{A}_p\mathbf{A}_q^H + \mathbf{A}_q\mathbf{A}_p^H)$ for $p\ne q$ , which is zero by orthogonality, and similarly for the $\mathbf{A}\mathbf{B}$ and $\mathbf{B}\mathbf{B}$ cross products.

The metric separates into $K$ scalar metrics

Combining the two reductions, the ML metric becomes $\mathcal{M}(s_1, \ldots, s_K) \;=\; \sum_{q=1}^K \Bigl[\|\mathbf{H}\|_F^2 |s_q|^2 - 2\mathrm{Re}\,\bigl(s_q^* \cdot z_q\bigr)\Bigr],$ where $z_q$ is a linear functional of $(\mathbf{Y}, \mathbf{H}, \mathbf{A}_q, \mathbf{B}_q)$ — the per-symbol matched-filter output. Each summand depends on $s_q$ alone, so the minimisation decouples: $\hat s_q = \arg\min_{s\in\mathcal{X}} [\|\mathbf{H}\|_F^2 |s|^2 - 2\mathrm{Re}(s^* z_q)]$ , which is a scalar slicer on the constellation.

Effective SNR and diversity

Completing the square, the per-symbol sufficient statistic is $\tilde y_q = z_q / \|\mathbf{H}\|_F^2$ and the effective SNR is $\text{SNR}_{\mathrm{eff}} \;=\; \frac{\|\mathbf{H}\|_F^2 \cdot \mathbb{E}[|s_q|^2]}{\sigma^2} \;=\; \frac{\text{SNR}}{n_t} \sum_{i=1}^{n_r}\sum_{j=1}^{n_t} |h_{i,j}|^2,$ using the normalisation $\mathbb{E}[|s_q|^2] = E_s/K$ and adjusting for the per-antenna energy split $1/n_t$ . This is a sum of $n_t n_r$ i.i.d. exponentially distributed squared-magnitudes, so the error probability decays as $\text{SNR}^{-n_t n_r}$ — diversity $n_t n_r$ , full MIMO diversity. $\blacksquare$

,

Example: Alamouti with QPSK: Two Scalar QPSK Decisions

Specialise the Alamouti scheme of §1 to QPSK input ( $\mathcal{X} = \{\pm 1 \pm j\}/\sqrt{2}$ ), $n_t = 2, n_r = 2$ , i.i.d. Rayleigh channel at SNR $\text{SNR} = 10$ dB per receive antenna. Compute the ML decoder decision rule, the expected effective SNR per symbol, and the approximate symbol-error probability $P_e$ .

Solution

ML decoder: two scalar QPSK slicers

By Thm. $2 n_r$ $2 n_{r}$ with Linear Decoding" data-ref-type="theorem">TAlamouti Achieves Full Diversity $2 n_r$ with Linear Decoding, the ML decoder forms $\tilde y_q = z_q / \|\mathbf{H}\|_F^2$ for $q = 1, 2$ and slices each separately on the QPSK constellation. The slicing is: $\hat s_q = \mathrm{sign}(\mathrm{Re}\,\tilde y_q) + j\,\mathrm{sign} (\mathrm{Im}\,\tilde y_q)$ , scaled by $1/\sqrt{2}$ .

Expected effective SNR

$\mathbb{E}[\text{SNR}_{\mathrm{eff}}] = (\text{SNR}/2) \cdot \mathbb{E}[\sum_{i,j} |h_{i,j}|^2] = (\text{SNR}/2) \cdot n_t n_r = 2 \text{SNR}$ for $n_t = n_r = 2$ . At $\text{SNR} = 10$ dB $= 10$ linear, $\mathbb{E}[\text{SNR}_{\mathrm{eff}}] = 20 \approx 13$ dB — a 3 dB increase over per-receive-antenna SNR thanks to the $n_t n_r = 4$ -fold summation.

Approximate SER

For QPSK at effective SNR $\gamma$ (post-MF), the symbol-error probability is $2Q(\sqrt{\gamma}) - Q^{2}(\sqrt{\gamma}) \approx 2 Q(\sqrt{\gamma})$ . Averaging over the $\chi^2_4$ distribution of $\sum|h_{i,j}|^2$ : $P_e \;\approx\; \int_0^\infty 2\,Q(\sqrt{\gamma})\, f_{\chi^2_4}(2\gamma/\text{SNR})\,\mathrm{d}\gamma \;\approx\; \binom{2 n_r - 1}{n_r - 1} \left(\frac{1}{4\text{SNR}}\right)^{n_r} \;=\; \frac{3}{16 \text{SNR}^{2}}$ for $n_r = 2$ (using the high-SNR diversity- $2 n_r$ tail formula). At $\text{SNR} = 10$ : $P_e \approx 3/1600 \approx 2\times 10^{-3}$ . The alamouti_ber plot of §1 confirms this numerically.

Sanity check: same slope as MRC, 3 dB worse

The $1\times 2$ MRC at the same total transmit power has effective SNR $\text{SNR}\sum|h_j|^2$ — exactly twice what Alamouti has with the same $\sum|h_{i,j}|^2$ . So MRC's $P_e$ is Alamouti's evaluated at $2\text{SNR}$ : ${P_e}_{\rm MRC} \approx 3/(16(2\text{SNR})^2) = 3/64\text{SNR}^{2}$ , a factor of 4 smaller ( $\approx 6$ dB shift). But wait — $n_r$ for MRC here is 2, whereas for Alamouti $2\times 2$ it is $n_t n_r = 4$ . The MRC benchmark for Alamouti's full-diversity $2 n_r = 4$ is $1\times 4$ MRC. With $1\times 4$ MRC we have $n_r = 4$ and $P_e \approx \binom{7}{3}/(4\text{SNR})^4 = 35/ (256 \text{SNR}^{4})$ — deep below Alamouti's $3/(16 \text{SNR}^{2})$ at any meaningful SNR. Alamouti matches $1\times 2$ MRC in slope, not in diversity order; it is a $2\times 1$ code, not a $1\times 4$ equivalent.

,

Example: Rate- $3/4$ OSTBC for $n_t = 4$ : Tarokh-Jafarkhani-Calderbank 1999

Write down the rate- $3/4$ OSTBC of Tarokh-Jafarkhani-Calderbank 1999 for $n_t = 4, T = 4, K = 3$ . Verify the orthogonality property $\mathbf{X}\mathbf{X}^H = (|s_1|^2 + |s_2|^2 + |s_3|^2)\mathbf{I}_4$ . Compare the rate to the Liang-Tarokh upper bound (§3).

Solution

The TJC code

The rate- $3/4$ OSTBC for $n_t = T = 4$ is $\mathbf{X} \;=\; \begin{pmatrix} s_1 & -s_2^* & \tfrac{s_3^*}{\sqrt{2}} & \tfrac{s_3^*}{\sqrt{2}} \\ s_2 & s_1^* & \tfrac{s_3^*}{\sqrt{2}} & -\tfrac{s_3^*}{\sqrt{2}} \\ \tfrac{s_3}{\sqrt{2}} & \tfrac{s_3}{\sqrt{2}} & \tfrac{-s_1 - s_1^* + s_2 - s_2^*}{2} & \tfrac{s_2 + s_2^* + s_1 - s_1^*}{2} \\ \tfrac{s_3}{\sqrt{2}} & -\tfrac{s_3}{\sqrt{2}} & \tfrac{s_2 + s_2^* - s_1 + s_1^*}{2} & -\tfrac{s_1 + s_1^* + s_2 - s_2^*}{2} \end{pmatrix}.$ It carries $K = 3$ complex symbols in $T = 4$ channel uses; the rate is $R_{\mathrm{OSTBC}} = 3/4$ complex symbols per channel use.

Orthogonality check

Direct calculation (tedious but straightforward — use a symbolic algebra system) confirms $\mathbf{X}\mathbf{X}^H = (|s_1|^2 + |s_2|^2 + |s_3|^2)\,\mathbf{I}_4$ for every $(s_1, s_2, s_3) \in \mathcal{X}^3$ . Hence by Thm. $n_t n_r$ $n_{t} n_{r}$ " data-ref-type="theorem">TEvery OSTBC Achieves Full Diversity $n_t n_r$ , the code achieves diversity $n_t n_r = 4 n_r$ , with 3 scalar slicer decisions at the receiver.

Rate vs. the Liang-Tarokh bound

The Liang-Tarokh upper bound (§3, Thm. TLiang-Tarokh Rate Upper Bound for Complex OSTBCs) for $n_t = 4$ is $R_{\max} = 3/4$ . The TJC code meets this bound with equality — it is a rate-optimal complex OSTBC for $n_t = 4$ . For $n_t > 4$ , the bound starts dropping below $3/4$ and the rate- optimal designs become more complicated. This is the best complex OSTBC rate for $n_t = 4$ , period.

,

Common Mistake: Codeword Orthogonality Is Not Channel Orthogonality

Mistake:

Reading "OSTBC has $\mathbf{X}\mathbf{X}^H = \|\mathbf{s}\|^2 \mathbf{I}$ " as the claim that the channel $\mathbf{H}$ becomes orthogonal somehow, or that the code requires $\mathbf{H}\mathbf{H}^{H}$ to be diagonal.

Correction:

The orthogonality is purely a codebook property — it holds for every channel realisation $\mathbf{H}$ . The receiver uses this codebook property to build an effective virtual channel whose columns are orthogonal (after symbol-wise conjugate-stacking as in the Alamouti proof). The actual MIMO channel $\mathbf{H}$ remains whatever it is — typically non-orthogonal i.i.d. Rayleigh.

The gain from OSTBC comes from exploiting channel randomness (the $n_t n_r$ -fold diversity sum), not from orthogonalising it. Confusing the two would lead to the wrong conclusion that the OSTBC receiver needs to pre-whiten $\mathbf{H}$ , which it does not.

,

Quick Check

Which algebraic condition on the dispersion matrices $\{\mathbf{A}_q, \mathbf{B}_q\}$ distinguishes an OSTBC from a general linear dispersion code?

The $2K$ matrices satisfy Hurwitz-Radon-Eckmann anti-commutation relations (OSTBC only)

The dispersion matrices are all unitary

The codeword matrix $\mathbf{X}$ itself is unitary

The rate $K/T$ equals 1 (one complex symbol per channel use)

Correction:

The

2K

matrices satisfy Hurwitz-Radon-Eckmann anti-commutation relations (OSTBC only)

This is the defining property. The cross-term-killing conditions $\mathbf{A}_p\mathbf{A}_q^H + \mathbf{A}_q\mathbf{A}_p^H = 2\delta_{pq} \mathbf{I}$ (and the analogous $\mathbf{B}\mathbf{B}$ and cross conditions) are what produce $\mathbf{X}\mathbf{X}^H = \|\mathbf{s}\|^2 \mathbf{I}$ . Without them, the codeword Gramian is not scalar and the ML decoder does not decouple.

Orthogonal Space-Time Block Code (OSTBC)

A linear STBC whose codewords satisfy $\mathbf{X}\mathbf{X}^H = \|\mathbf{s}\|^2 \mathbf{I}_{n_t}$ for every symbol choice, equivalently whose dispersion matrices satisfy Hurwitz-Radon-Eckmann anti-commutation. OSTBCs achieve full diversity $n_t n_r$ with linear matched-filter decoding (Tarokh-Jafarkhani-Calderbank 1999).

Dispersion Matrix

Fixed complex matrix $\mathbf{A}_q$ or $\mathbf{B}_q \in \mathbb{C}^{n_t \times T}$ that multiplies the real (respectively imaginary) part of the $q$ -th information symbol in a linear STBC's codeword expansion $\mathbf{X} = \sum_q (\alpha_q \mathbf{A}_q + j\beta_q \mathbf{B}_q)$ . Introduced in the Hassibi-Hochwald 2002 LDC framework.

Orthogonal Space-Time Block Codes