Ferkans — Interactive Telecom Tutor

Relaxing Orthogonality: The QOSTBC Idea

The Liang-Tarokh bound of §3 says: rate 1 + full diversity + linear decoding is impossible for $n_t > 2$ . What happens if we give up orthogonality to recover rate 1 at $n_t = 4$ ? Something has to give — the question is what.

Hamid Jafarkhani answered this in 2001: take two Alamouti blocks and arrange them in a $4\times 4$ matrix that is "almost" orthogonal, in the sense that the codeword Gramian $\mathbf{X}\mathbf{X}^H$ is not scalar but block-diagonal with $2\times 2$ blocks. This is the quasi- orthogonal space-time block code (QOSTBC). The price is:

Diversity drops from $n_t n_r = 4 n_r$ (full) to $2 n_r$ (half) — the rank of the worst error matrix is 2, not 4.
Decoder complexity rises from scalar slicer ( $K$ independent $M$ - ary searches) to pair-wise ML ( $K/2$ independent $M^2$ -ary searches).

In exchange, we keep rate 1. This is the rate-diversity-complexity triangle of space-time coding in its sharpest form: you can have any two, not all three. The QOSTBC is the canonical compromise for $n_t = 4$ : rate equal to V-BLAST's $n_t$ -stream rate (for $n_t = 4$ : $1$ symbol/cu), and diversity equal to Alamouti's, 2-fold per Rx — a nice middle between V-BLAST's no-diversity and OSTBC's limited rate.

Note that Jafarkhani's 2001 construction is the simplest and most celebrated, but it is not the only QOSTBC. Sharma-Papadias (2003) showed that rotating the constellation by an irrational angle recovers full diversity at the same rate — the diversity loss is therefore not fundamental but a consequence of the standard QPSK/QAM input.

Definition:
Jafarkhani's Quasi-Orthogonal STBC (QOSTBC)

The Jafarkhani QOSTBC is the rate-1 space-time code for $n_t = 4, T = 4, K = 4$ defined by $\mathbf{X}_Q(s_1, s_2, s_3, s_4) \;=\; \begin{pmatrix} \mathbf{X}_A(s_1, s_2) & \mathbf{X}_A(s_3, s_4) \\ -\mathbf{X}_A^*(s_3, s_4) & \mathbf{X}_A^*(s_1, s_2) \end{pmatrix} \;=\; \begin{pmatrix} s_1 & -s_2^* & s_3 & -s_4^* \\ s_2 & s_1^* & s_4 & s_3^* \\ -s_3^* & s_4 & s_1^* & -s_2 \\ -s_4^* & -s_3 & s_2^* & s_1 \end{pmatrix},$ where $\mathbf{X}_A$ is the $2\times 2$ Alamouti matrix. The codebook $\mathcal{Q} = \{\mathbf{X}_Q(s_1, s_2, s_3, s_4) : (s_1, \ldots, s_4) \in \mathcal{X}^4\}$ carries $K = 4$ complex symbols in $T = 4$ channel uses — rate $R = 1$ .

The code is the block Alamouti-of-Alamoutis construction: the outer $2\times 2$ arrangement is Alamouti-style in $2\times 2$ blocks. The Gramian $\mathbf{X}_Q \mathbf{X}_Q^H$ is therefore block-diagonal with $2\times 2$ scalar-identity blocks — "quasi-orthogonal" in the sense that $\{s_1, s_2\}$ and $\{s_3, s_4\}$ decouple, but within each pair the two symbols remain coupled.

Theorem: Jafarkhani QOSTBC: Rate 1, Diversity $2 n_r$ , Pair-Wise ML Decoder

The Jafarkhani QOSTBC of Definition DJafarkhani's Quasi-Orthogonal STBC (QOSTBC) satisfies:

(a) Rate: $R = 1$ complex symbol per channel use (maximum for any rate-matched linear code with $n_t = 4$ ).

(b) Diversity: The minimum rank of $\boldsymbol{\Delta} = \mathbf{X}_Q - \hat{\mathbf{X}}_Q$ over distinct codeword pairs is 2 (not full rank 4). Hence the QOSTBC achieves diversity $d = 2 n_r$ — only half the full MIMO diversity $4 n_r$ .

(c) Decoder: The ML decoder decouples the four symbols into two pairs — $(s_1, s_3)$ and $(s_2, s_4)$ — each pair requiring a joint $M^2$ -ary search (rather than two $M$ -ary scalar searches as in OSTBC). Total decoder complexity $O(M^2)$ rather than $O(M^K) = O(M^4)$ of brute-force ML.

The "half-diversity, pair-wise decoder" result is the operational consequence of the block-orthogonal structure. The cross term between $s_1$ and $s_3$ does not vanish in $\mathbf{X}_Q\mathbf{X}_Q^H$ , so ML cannot decide them independently; but their cross term with $s_2$ and $s_4$ does vanish, so the pair $(s_1, s_3)$ decouples from $(s_2, s_4)$ . The remaining pair-wise coupling is algebraic — a $2\times 2$ lattice decoder per pair rather than four scalar slicers.

The diversity loss to 2 per Rx has the same structural origin: the error matrix $\boldsymbol{\Delta}$ corresponding to an error event affecting only $s_1$ and $s_3$ has rank 2 in $\mathbb{C}^4$ because it lives in a 2D subspace. Sharma-Papadias (2003) observed that if $(s_1, s_3)$ are rotated by an irrational angle independent of $(s_2, s_4)$ before mapping, the minimum rank of $\boldsymbol{\Delta}$ rises to 4 — recovering full diversity at the same rate.

Show Hint

Compute $\mathbf{X}_Q\mathbf{X}_Q^H$ using the $2\times 2$ Alamouti block structure; find that it is block-diagonal with $2\times 2$ blocks.

Construct a codeword pair $(\mathbf{X}_Q, \hat{\mathbf{X}}_Q)$ that differs only in $(s_1, s_3)$ ; compute the rank of the difference.

Express the ML metric using the block-orthogonal structure and observe that $(s_1, s_3)$ and $(s_2, s_4)$ decouple.

Proof

Codeword Gramian is block-diagonal

Using the $2\times 2$ Alamouti blocks and the identity $\mathbf{X}_A \mathbf{X}_A^H = (|s_1|^2 + |s_2|^2)\mathbf{I}_2$ : $\mathbf{X}_Q \mathbf{X}_Q^H \;=\; \begin{pmatrix} (|s_1|^2 + |s_2|^2 + |s_3|^2 + |s_4|^2)\mathbf{I}_2 & \mathbf{E} \\ \mathbf{E}^H & (|s_1|^2 + |s_2|^2 + |s_3|^2 + |s_4|^2)\mathbf{I}_2 \end{pmatrix},$ where the cross-block $\mathbf{E} = 2\mathrm{Re}(s_1 s_4^* - s_2 s_3^*) \mathbf{I}_2$ is not zero. So $\mathbf{X}_Q \mathbf{X}_Q^H$ is not a scalar multiple of $\mathbf{I}_4$ — it is block-diagonal with a coupled $2\times 2$ block on the anti-diagonal. This is the "quasi" of quasi-orthogonal.

ML metric decouples into two pairs

Expanding $\|\mathbf{Y} - \mathbf{H}\mathbf{X}_Q\|_F^2$ using the dispersion matrices of $\mathbf{X}_Q$ , the cross terms between $\{s_1, s_4\}$ and $\{s_2, s_3\}$ vanish while the cross terms within $\{s_1, s_4\}$ and within $\{s_2, s_3\}$ survive (these are the off-diagonal entries of $\mathbf{E}$ ). Thus the ML metric separates into $\mathcal{M}(s_1, s_4) + \mathcal{M}(s_2, s_3)$ , and each sub-metric is minimised over the pair jointly — an $M^2$ search.

Minimum rank of the error matrix is 2

Consider an error event where only $s_1$ and $s_3$ are decided incorrectly: $(\hat s_1, \hat s_3) \ne (s_1, s_3)$ but $\hat s_2 = s_2, \hat s_4 = s_4$ . The error matrix $\boldsymbol{\Delta} = \mathbf{X}_Q(s) - \mathbf{X}_Q(\hat s)$ has its nonzero contributions only in columns 1 and 3 (by the structure of $\mathbf{X}_Q$ ), and its column space is at most 2-dimensional. Hence $\mathrm{rank}(\boldsymbol{\Delta}) \le 2$ .

Equality can be achieved by a suitable error pattern, so the minimum rank is exactly 2. By Thm. [?ch10:thm-rank-determinant], the diversity order is $n_r \cdot 2 = 2 n_r$ — half the full MIMO diversity.

Pair-wise ML decoder complexity

The decoder solves two independent $M^2$ searches — one for $(s_1, s_4)$ , one for $(s_2, s_3)$ . Complexity $O(2M^2) = O(M^2)$ , not $O(M^4)$ as brute-force ML would be.

For QPSK ( $M = 4$ ), the per-pair search is $M^2 = 16$ metric evaluations — cheap. For 16-QAM ( $M = 16$ ), it is $256$ per pair, still manageable. For 256-QAM the search grows to $65\,536$ per pair and sphere-decoding heuristics become essential. $\blacksquare$

,

Example: Jafarkhani QOSTBC for QPSK: The 4-Symbol Codeword

Write the Jafarkhani QOSTBC codeword for the specific input $(s_1, s_2, s_3, s_4) = ((1+j), (1-j), (-1+j), (-1-j))/\sqrt{2}$ (unit-energy QPSK symbols). Compute $\mathbf{X}_Q \mathbf{X}_Q^H$ and verify it is block-diagonal with the predicted off-block structure.

Solution

Build the codeword

Substituting into Definition DJafarkhani's Quasi-Orthogonal STBC (QOSTBC): $\mathbf{X}_Q \;=\; \frac{1}{\sqrt 2}\begin{pmatrix} 1+j & -(1+j) & -1+j & -(-1+j) \\ 1-j & 1-j & -1-j & -1+j \\ -(-1-j) & -1-j & 1+j & -(1-j) \\ -(-1+j) & -(-1+j) & 1+j & 1-j \end{pmatrix} \;=\; \frac{1}{\sqrt 2}\begin{pmatrix} 1+j & -1-j & -1+j & 1-j \\ 1-j & 1-j & -1-j & -1+j \\ 1+j & -1-j & 1+j & -1+j \\ 1-j & 1-j & 1+j & 1-j \end{pmatrix}.$

Compute $\mathbf{X}_Q\mathbf{X}_Q^H$

Each QPSK symbol has $|s_q|^2 = 1$ , so $\sum_{q=1}^4 |s_q|^2 = 4$ . By Thm. $2 n_r$ $2 n_{r}$ , Pair-Wise ML Decoder" data-ref-type="theorem">TJafarkhani QOSTBC: Rate 1, Diversity $2 n_r$ , Pair-Wise ML Decoder, $\mathbf{X}_Q \mathbf{X}_Q^H = \begin{pmatrix} 4 \mathbf{I}_2 & \mathbf{E} \\ \mathbf{E}^H & 4 \mathbf{I}_2 \end{pmatrix}$ with $\mathbf{E} = 2\mathrm{Re}(s_1 s_4^* - s_2 s_3^*)\mathbf{I}_2$ . Evaluate: $s_1 s_4^* = (1+j)(-1+j)/2 = (-1-j+j+j^2)/2 = (-2)/2 = -1$ ; $s_2 s_3^* = (1-j)(-1-j)/2 = (-1-j+j+j^2)/2 = (-2)/2 = -1$ . So $s_1 s_4^* - s_2 s_3^* = 0$ and $\mathbf{E} = 0$ .

The specific input is lucky

For this particular input the off-block happens to vanish — the QOSTBC Gramian is scalar, $\mathbf{X}_Q\mathbf{X}_Q^H = 4 \mathbf{I}_4$ , as if it were a rate-1 full-diversity OSTBC. But for a generic QPSK input (say $s_1 s_4^* = 1, s_2 s_3^* = -1$ ), we would get $\mathbf{E} = 2\mathrm{Re}(1 - (-1))\mathbf{I}_2 = 4 \mathbf{I}_2$ and the Gramian would have a large off-diagonal block.

The QOSTBC Gramian is input-dependent — unlike OSTBC, where the Gramian is always scalar. This input-dependence is what forces the pair-wise ML decoder and creates the diversity loss: for worst-case error patterns the Gramian coupling prevents rank-4 error matrices.

BER: OSTBC vs QOSTBC vs V-BLAST at $n_t = 4$

Side-by-side BER at $n_t = n_r = 4$ of three $n_t = 4$ space-time schemes, at the same spectral efficiency (2 bits/channel use for illustration):

Rate- $3/4$ TJC OSTBC with 16-QAM (full diversity 16, linear MF decoder): steepest slope, asymptotically dominant below a target BER of $\sim 10^{-4}$ .
Jafarkhani QOSTBC with QPSK (diversity 8, pair-wise ML decoder): same rate (2 bits/cu = 1 symbol/cu × 2 bits/symbol), but flatter slope due to half-diversity; wins in the high-BER, moderate-SNR regime.
V-BLAST with QPSK + ZF-SIC (rate 4 symbols/cu raw, here rate- matched with half-rate code; diversity 1 per layer): shallowest slope, an error-floor-dominated regime.

The message: at fixed rate, diversity order determines the BER slope — OSTBC wins asymptotically, QOSTBC wins at moderate SNR, V-BLAST wins only at extremely high SNR (or with outer coding to exploit multiplexing). The crossover points between the curves map out the Zheng-Tse DMT curve which Chapter 12 develops fully.

Parameters

Common Mistake: QOSTBC Is Not Full-Diversity

Mistake:

Reading "QOSTBC is rate 1 at $n_t = 4$ " as the claim that QOSTBC is a drop-in replacement for Alamouti at $n_t = 4$ — i.e., rate 1, full diversity, linear decoder.

Correction:

QOSTBC at $n_t = 4$ has diversity $2 n_r$ , which is half the full diversity $4 n_r$ . If your target BER is $10^{-6}$ at $\text{SNR} = 15$ dB with $n_r = 4$ , QOSTBC (diversity 8) and OSTBC (diversity 16) give wildly different answers: OSTBC delivers it; QOSTBC sits $\sim 6$ dB worse at that operating point. The pair-wise decoder is also more expensive than OSTBC's scalar one — both costs compound.

Sharma-Papadias (2003) "rotation-based QOSTBC" restores full diversity at the same rate by pre-rotating half the symbols by an irrational angle. This is sometimes called "rotated QOSTBC" or "optimal QOSTBC" and is the production-grade variant; the unrotated Jafarkhani QOSTBC of Definition DJafarkhani's Quasi-Orthogonal STBC (QOSTBC) is the pedagogical baseline.

In cellular standards, Alamouti at $n_t = 2$ is the only truly-deployed orthogonal code; $n_t = 4$ systems use codebook-based precoding (closed- loop) rather than QOSTBC (open-loop) because the precoding gain is larger than the QOSTBC diversity at typical cellular CSI qualities.

,

🔧Engineering Note

QOSTBC in Wi-Fi and Why Standards Prefer Other Paths

Jafarkhani's 2001 QOSTBC, and its rotated variant by Sharma-Papadias 2003, were in the running for 4-Tx transmit-diversity modes in 802.11n/ac and LTE-Release 9 during 2006-2009. In the end, neither was adopted.

The reasons are:

Codebook precoding wins when CSI is imperfect but available. In a base-station downlink, the UE can feed back a PMI (precoding-matrix indicator) that the BS uses to beamform. The beamforming gain is typically 2-4 dB at low-to-moderate UE speeds, exceeding the diversity gain of an open-loop QOSTBC.
Spatial multiplexing wins at high SNR. At $\text{SNR} > 15$ dB, cellular systems serve rate-limited users anyway; V-BLAST-style spatial multiplexing (4 streams at $n_t = n_r = 4$ ) with LDPC coding dominates QOSTBC at the same spectral efficiency.
Implementation simplicity. 2-antenna Alamouti is fully standardised and carries minimal complexity; 4-antenna QOSTBC requires a new dedicated pair-wise decoder, and vendors preferred to spend silicon area on extra Rx antennas rather than QOSTBC circuitry.

QOSTBC remains an important pedagogical construction — it is the clearest demonstration of the rate-diversity-complexity triangle in space-time coding — and it is occasionally used in specialised applications (deep-space, military, some sensor networks). But in mainstream cellular / WLAN standards, Alamouti-2 + codebook-precoding-4 is the dominant architecture.

Practical Constraints

•
QOSTBC at $n_t = 4$ delivers rate 1 but half-diversity; full-diversity variants need constellation rotation
•
Pair-wise ML decoder has $O(M^2)$ per pair vs $O(M)$ per symbol for OSTBC
•
Not deployed in 3GPP / IEEE 802.11 cellular / WLAN standards

Historical Note: Jafarkhani 2001: Quasi-Orthogonal STBCs and the Rate-Diversity Trade

2001

Hamid Jafarkhani's 2001 IEEE Trans. Commun. paper, "A quasi-orthogonal space-time block code" (vol. 49, no. 1, Jan. 2001, pp. 1-4), introduced the Alamouti-of-Alamoutis construction at $n_t = 4$ . Jafarkhani — then at AT&T Labs, later at UC Irvine — was a co-author (with Tarokh and Calderbank) of the 1999 OSTBC paper, and his 2001 work was explicitly a response to the Liang-Tarokh rate ceiling: "If we cannot have rate 1 + full diversity + orthogonal, let's see what combinations we can have."

The paper crystallised the rate-diversity-complexity triangle that every subsequent space-time coding paper implicitly invokes. Within two years, Sharma and Papadias (2003) showed that a simple constellation rotation restores full diversity in the Jafarkhani construction — at the cost of inseparable pair-wise decoding over a rotated QAM (not the original QAM). This made "rotated QOSTBC" the reference rate-1 full-diversity code for $n_t = 4$ , though it remained a research curiosity rather than a standard feature.

The bigger historical contribution of Jafarkhani 2001 was sociological: it legitimised non-orthogonal constructions as first-class space- time codes. Until 2001, "space-time coding" meant "OSTBC" in most of the community. After 2001, everyone knew that OSTBC was a special case, and the rate-diversity trade-off was negotiable. This cleared the ground for linear dispersion codes (Hassibi-Hochwald 2002), the Golden code (Belfiore-Rekaya 2003), and Perfect codes (Oggier et al. 2006) — the whole Ch. 13 story.

,

Quick Check

Jafarkhani's QOSTBC at $n_t = 4$ achieves rate 1. What is its diversity order on an i.i.d. Rayleigh channel with $n_r$ receive antennas?

$2 n_r$ (half the full diversity)

$4 n_r$ (full MIMO diversity, same as OSTBC)

$n_r$ (same as a single-antenna SIMO)

$2$ (independent of $n_r$ )

Correction:

2 n_r

(half the full diversity)

Exactly right. The minimum rank of the error matrix $\boldsymbol{\Delta} = \mathbf{X}_Q - \hat{\mathbf{X}}_Q$ is 2, not 4, because error events that flip only $(s_1, s_3)$ (or only $(s_2, s_4)$ ) produce rank-2 differences. Hence diversity $= 2 n_r$ , half of the full $4 n_r$ . See Thm. $2 n_r$ $2 n_{r}$ , Pair-Wise ML Decoder" data-ref-type="theorem">TJafarkhani QOSTBC: Rate 1, Diversity $2 n_r$ , Pair-Wise ML Decoder.

Quasi-Orthogonal Space-Time Block Code (QOSTBC)

A linear STBC whose codeword Gramian $\mathbf{X}\mathbf{X}^H$ is block-diagonal (but not scalar), allowing pair-wise rather than scalar ML decoding. The canonical example is Jafarkhani's 2001 $n_t = 4$ construction: rate 1, diversity $2 n_r$ , $O(M^2)$ -per-pair decoder. A rotated variant (Sharma-Papadias 2003) recovers full diversity at the same rate.

Quasi-Orthogonal Space-Time Block Codes

Relaxing Orthogonality: The QOSTBC Idea

Definition: Jafarkhani's Quasi-Orthogonal STBC (QOSTBC)

Theorem: Jafarkhani QOSTBC: Rate 1, Diversity 2nr2 n_r2nr​, Pair-Wise ML Decoder

Codeword Gramian is block-diagonal

ML metric decouples into two pairs

Minimum rank of the error matrix is 2

Pair-wise ML decoder complexity

Example: Jafarkhani QOSTBC for QPSK: The 4-Symbol Codeword

Build the codeword

Compute $\mathbf{X}_Q\mathbf{X}_Q^H$

The specific input is lucky

BER: OSTBC vs QOSTBC vs V-BLAST at nt=4n_t = 4nt​=4

Parameters

Common Mistake: QOSTBC Is Not Full-Diversity

QOSTBC in Wi-Fi and Why Standards Prefer Other Paths

Historical Note: Jafarkhani 2001: Quasi-Orthogonal STBCs and the Rate-Diversity Trade

Quick Check

Quasi-Orthogonal Space-Time Block Code (QOSTBC)

Definition:
Jafarkhani's Quasi-Orthogonal STBC (QOSTBC)

Theorem: Jafarkhani QOSTBC: Rate 1, Diversity $2 n_r$ , Pair-Wise ML Decoder

BER: OSTBC vs QOSTBC vs V-BLAST at $n_t = 4$