Ferkans — Interactive Telecom Tutor

Two High-SNR Resources, One Channel

An $n_t \times n_r$ block-fading MIMO channel offers the designer two distinct asymptotic resources at high SNR, and they are not the same thing.

Resource 1 — Diversity. At a fixed target rate $R$ , the outage probability decays polynomially in SNR: $P_{\rm out}(R) \doteq \text{SNR}^{-d}$ . The exponent $d$ is the diversity order, and it measures the reliability slope of the communication link on a log-log BER-vs-SNR plot at high SNR. Alamouti on an $n_r$ -receive channel achieves the maximum $d = 2 n_r$ on a $2 \times n_r$ channel (Chapter 11); V-BLAST with zero-forcing achieves only $d = n_r - n_t + 1$ . Both operate at a fixed rate.

Resource 2 — Multiplexing. At scaling target rate $R(\text{SNR}) = r \log_2 \text{SNR} + o(\log \text{SNR})$ , the rate itself grows with the log-SNR at slope $r$ . The pre-log slope $r$ is the multiplexing gain, and it measures the capacity slope of the link. Telatar's MIMO capacity theorem (Chapter 10) says that the ergodic capacity of an $n_t \times n_r$ Rayleigh channel grows like $\min(n_t, n_r) \log_2 \text{SNR}$ , so the maximum multiplexing gain is $r_{\max} = \min(n_t, n_r)$ . V-BLAST with ML detection achieves this maximum; Alamouti, transmitting at fixed rate 1, achieves only $r = 0$ .

The point is that Alamouti and V-BLAST are not comparable on the same axis: Alamouti is a full-diversity-zero-multiplexing code, while V-BLAST is a full-multiplexing-low-diversity scheme. To compare them fairly we need a single performance curve that varies rate with SNR and measures both resources simultaneously. That curve is the Zheng-Tse diversity-multiplexing tradeoff $d^*(r)$ , which we define in this section and derive in §2.

The central insight: diversity and multiplexing are not independent design choices. Zheng and Tse (2003) showed that they trade off along a precise piecewise-linear curve $d^*(r) = (n_t - r)(n_r - r)$ — this is the fundamental constraint that every space-time code designer navigates.

,

Definition:
Exponential Equality $\doteq$

Let $f(\text{SNR})$ and $g(\text{SNR})$ be two positive functions of SNR. We write $f(\text{SNR}) \doteq g(\text{SNR})$ , and say that $f$ is exponentially equal to $g$ , if $\lim_{\text{SNR}\to\infty} \frac{\log f(\text{SNR})}{\log \text{SNR}} \;=\; \lim_{\text{SNR}\to\infty} \frac{\log g(\text{SNR})}{\log \text{SNR}}.$ Equivalently, $f(\text{SNR}) \doteq \text{SNR}^{a}$ iff $\lim_{\text{SNR}\to \infty} \log f(\text{SNR}) / \log \text{SNR} = a$ . The exponent $a$ is the only feature of $f$ that survives the limit.

What $\doteq$ ignores:

Multiplicative constants: $c \cdot \text{SNR}^{a} \doteq \text{SNR}^{a}$ for any constant $c > 0$ .
Polylogarithmic prefactors: $(\log \text{SNR})^k \cdot \text{SNR}^{a} \doteq \text{SNR}^{a}$ for any $k$ .
Lower-order polynomial corrections: $\text{SNR}^{a} + \text{SNR}^{a-1} \doteq \text{SNR}^{a}$ .

What $\doteq$ keeps: only the leading polynomial order. The relations $\dotle$ (" $\le$ " exponentially) and $\dotge$ are defined analogously.

The DMT is an asymptotic statement — it is exact only in the limit $\text{SNR} \to \infty$ . At finite SNR, the coding gain (the multiplicative constant in front of $\text{SNR}^{-d}$ ) matters just as much as the diversity exponent, and two codes with the same DMT can differ by 3–6 dB in actual performance. Chapter 13 constructs the CDA / Golden-code family which has both optimal DMT and good coding gain.

,

Definition:
Multiplexing Gain $r$

Consider a family of codes $\{\mathcal{C}_{\text{SNR}}\}$ indexed by SNR, with $\mathcal{C}_{\text{SNR}}$ operating at rate $R(\text{SNR})$ bits per channel use. The multiplexing gain of the family is $r \;=\; \lim_{\text{SNR}\to\infty} \frac{R(\text{SNR})}{\log_2 \text{SNR}}.$ Operationally, $r$ is the slope at which the code's rate grows with the log-SNR as SNR increases. A code with fixed rate $R(\text{SNR}) = R_0$ has multiplexing gain $r = 0$ ; a code with $R(\text{SNR}) = 2 \log_2 \text{SNR}$ has $r = 2$ .

For an $n_t \times n_r$ i.i.d. Rayleigh channel, Telatar's MIMO capacity theorem (Chapter 10) gives $\bar C(\text{SNR}) = \min(n_t, n_r) \log_2 \text{SNR} + O(1)$ at high SNR. A reliable family of codes must therefore satisfy $r \;\le\; \min(n_t, n_r).$ The upper bound $r_{\max} = \min(n_t, n_r)$ is the number of spatial degrees of freedom of the channel; it is achieved by any family whose rate matches the capacity slope.

,

Definition:
Diversity Gain $d^*$

Consider a family of codes $\{\mathcal{C}_{\text{SNR}}\}$ operating at rate $R(\text{SNR})$ with block error probability $P_e(\text{SNR})$ under ML decoding. The diversity gain of the family is $d^* \;=\; -\lim_{\text{SNR}\to\infty} \frac{\log P_e(\text{SNR})}{\log \text{SNR}}.$ Equivalently, $P_e(\text{SNR}) \doteq \text{SNR}^{-d^*}$ .

For a fixed-rate family ( $r = 0$ ), $d^*$ reduces to the classical diversity order of Chapters 10–11: the log-log slope of BER against SNR at high SNR. For a scaling-rate family ( $r > 0$ ), $d^*$ is the exponent at which error probability still decays while the rate is growing — a subtler quantity, because it measures the joint reliability-and-rate scaling.

Outage lower bound. For any code, the block error probability is lower-bounded by the outage probability at the target rate: $P_e(\text{SNR}) \;\ge\; P_{\rm out}(R(\text{SNR})).$ So the diversity gain of any family is bounded by the outage-diversity exponent, $d^* \;\le\; -\lim_{\text{SNR}\to\infty} \frac{\log P_{\rm out}(r \log_2 \text{SNR})}{\log \text{SNR}} \;=:\; d^*_{\rm out}(r).$ The Zheng-Tse theorem of §2 shows that this outage-diversity exponent is achievable — and hence equal to $d^*(r)$ — by a Gaussian random codebook, so $d^*(r) = d^*_{\rm out}(r)$ .

,

Alamouti at $(1,4)$ , V-BLAST at $(2, d)$ — Where Is the Fair Comparison?

On a $2 \times 2$ MIMO channel, Alamouti and V-BLAST produce very different $(r, d)$ operating points:

Alamouti at rate $1$ bit/use: transmits one symbol per channel use on average (two symbols over two channel uses). The rate is fixed in SNR, so $r_{\rm Alamouti} = 0$ . The diversity order is $d = n_t n_r = 4$ . Operating point: $(0, 4)$ .
V-BLAST with ML detection at rate $R(\text{SNR}) = 2 \log_2 \text{SNR}$ : transmits two independent streams, one per antenna; rate grows at slope $2$ . So $r_{\rm VBLAST} = 2$ . The diversity order at this rate, under ML detection, is $d = 1$ (the worst-case eigenvalue of $\mathbf{H}\mathbf{H}^{H}$ fades independently with exponent $1$ ). Operating point: $(2, 1)$ .
V-BLAST with zero-forcing at rate $R(\text{SNR}) = 2 \log_2 \text{SNR}$ : same $r = 2$ , but each spatial stream sees a post-ZF effective SNR that is limited by the weakest eigenvalue; the ZF diversity per stream is $n_r - n_t + 1 = 1$ . Operating point: $(2, 1)$ (same as V-BLAST-ML at $r = 2$ , but worse coding gain at finite SNR).

These two codes are NOT competitors on the same plot — they are sitting at two ends of an implicit curve. What is the curve? Is Alamouti actually the best achievable at $r = 0$ ? Is V-BLAST the best at $r = 2$ ? Or can a single code hit $(1, d)$ for some interior $d > 0$ ? The Zheng-Tse theorem of §2 answers all three questions with a single closed- form formula.

,

Empirical DMT: $-\log P_{\rm out}/\log \text{SNR}$ Converges to $d^*(r)$

This plot shows the empirical outage-diversity exponent $d_{\rm empirical}(\text{SNR}) = -\log P_{\rm out}(r \log_2 \text{SNR}) / \log \text{SNR}$ as a function of SNR, for a chosen multiplexing gain $r$ and MIMO dimensions $(n_t, n_r)$ . As $\text{SNR} \to \infty$ , the empirical exponent converges to the Zheng-Tse exponent $d^*(r) = (n_t - r)(n_r - r)$ . At moderate SNR the curve deviates — sometimes significantly — from the asymptote, which is the practical caveat of all DMT statements: the DMT is a high-SNR statement, and even "high SNR" in simulations typically means $\text{SNR} \ge 20$ dB. The convergence rate itself depends on $(n_t, n_r, r)$ — the $r = 0$ corner converges fastest, the $r = \min(n_t, n_r)$ corner slowest.

Parameters

n_t

2

n_r

2

Multiplexing gain

r

1

Example: DMT of the Scalar Rayleigh Channel ( $n_t = n_r = 1$ )

Compute the DMT curve $d^*(r)$ for the scalar $n_t = n_r = 1$ Rayleigh channel with coherent detection and a single complex symbol per channel use. Verify the general formula $d^*(r) = (n_t - r)(n_r - r)$ .

Solution

Outage probability at scaling rate

The scalar Rayleigh channel has instantaneous capacity $C(|h|^2) = \log_2(1 + \text{SNR} |h|^2)$ with $|h|^2 \sim \exp(1)$ . At target rate $R(\text{SNR}) = r \log_2 \text{SNR}$ , the outage event is $\mathcal{O} = \{C(|h|^2) < r \log_2 \text{SNR}\} = \{|h|^2 < (\text{SNR}^{r} - 1)/\text{SNR}\} \doteq \{|h|^2 < \text{SNR}^{r-1}\}.$

Outage probability exponent

Since $|h|^2 \sim \exp(1)$ , $\mathbb{P}(|h|^2 < \epsilon) = 1 - e^{-\epsilon} \doteq \epsilon$ for small $\epsilon$ . Therefore $P_{\rm out}(R(\text{SNR})) \doteq \text{SNR}^{r-1} = \text{SNR}^{-(1-r)}.$ The diversity exponent is $d^*(r) = 1 - r$ for $r \in [0, 1]$ , and $d^*(r) = 0$ beyond $r = 1$ (zero reliability once rate exceeds capacity slope).

Verify Zheng-Tse formula

With $n_t = n_r = 1$ , the Zheng-Tse formula predicts $d^*(r) = (1-r)(1-r) = (1-r)^2$ at integer $r \in \{0, 1\}$ — i.e., the corner points are $(0, 1)$ and $(1, 0)$ . Linearly interpolating, $d^*(r) = 1 - r$ for $r \in [0, 1]$ , matching our direct computation.

So the scalar Rayleigh DMT is the straight line from $(0, 1)$ to $(1, 0)$ — the simplest possible DMT curve, and the one that sets the pattern for all higher-dimensional MIMO channels: each unit of $r$ costs exactly $1$ unit of $d^*$ , at least in the scalar case. For multi-antenna channels the cost per unit is higher than 1 near the $r = 0$ corner and exactly 1 near the $r = r_{\max}$ corner — the next section formalises this.

,

Historical Note: Foschini 1996: V-BLAST and the Birth of Spatial Multiplexing

1996

Gerard Foschini's 1996 Bell Labs Technical Journal paper, "Layered space-time architecture for wireless communication in a fading environment when using multi-element antennas", introduced what is today called V-BLAST (Vertical Bell Labs Layered Space-Time): transmit $n_t$ independent data streams over $n_t$ antennas and recover them at the receiver via successive interference cancellation.

Foschini's revolutionary claim — contrary to the then-prevailing wisdom of Alamouti-style repetition — was that MIMO channels support rate growth proportional to $\min(n_t, n_r)$ , not merely improved reliability at fixed rate. This is the spatial multiplexing concept: the channel has multiple independent spatial dimensions, and a clever transmitter should use them all to carry independent data rather than repeat the same data with diversity.

The 1998 Wolniansky-Foschini-Golden-Valenzuela prototype (the "V-BLAST demonstration" at Bell Labs, Holmdel) experimentally confirmed the theoretical predictions on an indoor $4 \times 4$ MIMO channel at 36 bps/Hz — an order of magnitude above single-antenna Shannon capacity at the same SNR.

The tension between Alamouti (full diversity, rate 1) and V-BLAST (full multiplexing, rate $\min(n_t, n_r)$ ) persisted for five years until Zheng and Tse's 2003 paper resolved it by showing that both are corner points of a single tradeoff curve.

,

Quick Check

A family of codes operates at rate $R(\text{SNR}) = 2 \log_2 \text{SNR}$ bits per channel use on a $4 \times 4$ i.i.d. Rayleigh MIMO channel. Its multiplexing gain $r$ is:

$r = 0$ (because the family is fixed-rate)

$r = 2$ (the pre-log slope)

$r = 4$ (the full spatial degrees of freedom)

$r$ depends on the family's error probability, not the rate

Correction:

r = 2

(the pre-log slope)

By Definition $r$ $r$ " data-ref-type="definition">DMultiplexing Gain $r$ , $r = \lim R(\text{SNR}) / \log_2 \text{SNR} = 2$ . This is the slope at which rate grows with log-SNR.

Common Mistake: $\doteq$ Is Not $=$ — Constants and Polylogs Are Invisible

Mistake:

Reading $P_{\rm out}(R) \doteq \text{SNR}^{-d}$ as $P_{\rm out}(R) = C \cdot \text{SNR}^{-d}$ for some explicit constant $C$ (so e.g. concluding that two codes with the same DMT have identical outage probabilities at every SNR).

Correction:

The exponential-equality relation $\doteq$ only captures the leading polynomial order. Two codes with the same DMT — e.g., Alamouti and the Golden code at $r = 1$ on $2 \times 2$ — generally have different multiplicative constants in front of $\text{SNR}^{-d^*}$ . Those constants translate into dB offsets of the BER-vs-SNR curve: the Golden code is typically $\sim 3$ – $4$ dB better than Alamouti at rate $2$ bits/use on $2 \times 2$ , even though both have diversity order $2$ at $r = 1$ . The coding-gain gap is invisible to $\doteq$ .

Rule of thumb. Use the DMT as a first-order design criterion to choose the code's asymptotic operating point. Use finite-SNR Monte Carlo (or the determinant criterion of Chapter 11) to choose between codes with the same DMT. Both analyses are necessary; neither is sufficient alone.

Diversity Gain

The exponent $d^*$ at which a code family's block error probability decays polynomially in SNR at high SNR: $P_e(\text{SNR}) \doteq \text{SNR}^{-d^*}$ . For fixed-rate families, $d^*$ coincides with the classical diversity order of Chapters 10–11. For scaling-rate families with multiplexing gain $r$ , $d^*$ is a function of $r$ called the DMT.

Multiplexing Gain

The pre-log slope $r = \lim R(\text{SNR}) / \log_2 \text{SNR}$ at which a code family's rate grows with log-SNR. The maximum on an $n_t \times n_r$ i.i.d. Rayleigh channel is $r_{\max} = \min(n_t, n_r)$ , equal to the spatial degrees of freedom.

Exponential Equality ( $\doteq$ )

Two positive functions of SNR are exponentially equal, $f \doteq g$ , if $\lim \log f / \log \text{SNR} = \lim \log g / \log \text{SNR}$ . Only the leading polynomial order is kept; multiplicative constants, polylog prefactors, and lower-order terms are ignored. Central to the DMT framework.

Why This Matters: Why Link Adaptation Slides Along the DMT Curve

In a real 5G or Wi-Fi system the SNR at the receiver varies continuously with distance, beam alignment, and interference. The scheduler cannot commit to "diversity mode" or "multiplexing mode" once and for all — it must adapt. As SNR increases, the scheduler picks higher modulation order and higher MIMO rank, increasing the effective $r$ at the cost of $d^*$ . As SNR decreases, the scheduler falls back to robust single- stream transmission with strong coding, increasing $d^*$ at the cost of $r$ . The scheduler is literally walking along the DMT curve as channel quality changes. Section 3 makes this operational reading precise via rank adaptation in LTE/5G.

Motivating the Tradeoff: Diversity vs Multiplexing

Two High-SNR Resources, One Channel

Definition: Exponential Equality ≐\doteq≐

Definition: Multiplexing Gain rrr

Definition: Diversity Gain d∗d^*d∗

Alamouti at (1,4)(1,4)(1,4), V-BLAST at (2,d)(2, d)(2,d) — Where Is the Fair Comparison?

Empirical DMT: −log⁡Pout/log⁡SNR-\log P_{\rm out}/\log \text{SNR}−logPout​/logSNR Converges to d∗(r)d^*(r)d∗(r)