Ferkans — Interactive Telecom Tutor

From Formula to Engineering: What the DMT Actually Says

Armed with the Zheng-Tse formula $d^*(r) = (n_t - r)(n_r - r)$ , we can now read the curve as an engineer. Three operational messages emerge.

Message 1 — the tradeoff is quadratic at $r = 0$ and linear at $r = r_{\max}$ . The initial slope $-dd^*/dr|_{r = 0} = n_t + n_r$ is large: sacrificing the first unit of diversity buys you a lot of multiplexing. The final slope $-dd^*/dr|_{r = r_{\max}^-} = 1$ is small: sacrificing the last unit of multiplexing gets you only one unit of diversity back. The curve is concave on each linear segment, with shrinking segment slopes: the first $r = 0 \to 1$ step costs $n_t + n_r - 1$ units of $d^*$ , the second $r = 1 \to 2$ step costs $n_t + n_r - 3$ units, …, the last $r = m-1 \to m$ step costs $n_t + n_r - (2m - 1)$ units (= $|n_t - n_r| + 1$ ).

Message 2 — the curve is symmetric in $(n_t, n_r)$ . A $4 \times 2$ and a $2 \times 4$ channel have identical DMT curves, even though their physical characteristics differ. Nonetheless, the tradeoff between reliability and rate is the same: you cannot tell by looking at an $(r, d^*)$ plot whether the transmitter or the receiver has more antennas. This is a deep consequence of the outage-exponent computation: the Wishart density of $\mathbf{H}\mathbf{H}^{H}$ vs $\mathbf{H}^{H}\mathbf{H}$ differ only in the normalising constant, not in the eigenvalue-exponent rate function.

Message 3 — rank adaptation = DMT climbing. LTE and 5G NR contain an explicit rank indicator that the UE reports to the gNB, specifying the number of spatial streams $k \in \{1, \ldots, \min(n_t, n_r)\}$ to be used for the next slot. Conceptually, $k$ is exactly the operating multiplexing gain — and the choice of $k$ navigates the DMT curve in response to channel conditions. The engineering note below makes this precise.

,

Definition:
DMT Slope and Segment Structure

The DMT curve has $\min(n_t, n_r)$ linear segments; on the $k$ -th segment $r \in [k, k+1]$ the slope is $\left.\frac{dd^*}{dr}\right|_{r \in (k, k+1)} = (n_t - k - 1)(n_r - k - 1) - (n_t - k)(n_r - k) = -(n_t + n_r - 2k - 1).$ The initial slope (at $r = 0^+$ ) is $-(n_t + n_r - 1)$ ; the final slope (at $r = r_{\max}^-$ ) is $-(|n_t - n_r| + 1)$ .

Consequence. The DMT curve is concave and piecewise linear, not the continuous quadratic. It lies strictly above the continuous $(n_t - r)(n_r - r)$ between corner points. The gap is achieved by time-sharing between two adjacent integer-rate codes: a code at $r = k$ run a fraction $1 - (r - k)$ of the time, and a code at $r = k+1$ run a fraction $r - k$ of the time, achieves multiplexing gain $r$ and diversity gain equal to the linear interpolation. This is the operational meaning of the "piecewise-linear interpolation".

,

Theorem: DMT Symmetry in $(n_t, n_r)$

Under the assumptions of Thm. TZheng-Tse Diversity-Multiplexing Tradeoff, $d^*(r) \text{ for an } n_t \times n_r \text{ channel} \;=\; d^*(r) \text{ for an } n_r \times n_t \text{ channel}$ for all $r \in [0, \min(n_t, n_r)]$ . In particular, the tradeoff curves of an $(n_t, n_r)$ channel and its transpose $(n_r, n_t)$ are identical, independent of which side has more antennas.

The DMT only depends on the eigenvalues of the Wishart matrix. The Wishart matrices $\mathbf{H}\mathbf{H}^{H}$ (size $n_r \times n_r$ ) and $\mathbf{H}^{H}\mathbf{H}$ (size $n_t \times n_t$ ) have identical nonzero eigenvalues. The DMT exponent sees only those shared eigenvalues — the transmitter-vs-receiver asymmetry washes out.

Show Hint

The nonzero eigenvalues of $\mathbf{H}\mathbf{H}^{H}$ equal those of $\mathbf{H}^{H}\mathbf{H}$ .

The Zheng-Tse proof only uses the eigenvalue distribution — transpose the channel, transpose nothing in the proof.

Verify directly from the corner-point formula $(n_t - k)(n_r - k)$ , which is symmetric in $(n_t, n_r)$ .

Proof

Invariance of nonzero eigenvalues under transposition

For any matrix $\mathbf{H} \in \mathbb{C}^{n_r \times n_t}$ , the matrices $\mathbf{H}\mathbf{H}^{H} \in \mathbb{C}^{n_r \times n_r}$ and $\mathbf{H}^{H}\mathbf{H} \in \mathbb{C}^{n_t \times n_t}$ have the same set of nonzero eigenvalues — a standard fact (the "push- through" or "rank identity"). In particular the ordered nonzero eigenvalues $\lambda_1 \ge \lambda_2 \ge \cdots \ge \lambda_m$ with $m = \min(n_t, n_r)$ are identical for the two matrices.

DMT depends only on the nonzero eigenvalues

The outage event at rate $r \log_2 \text{SNR}$ is $\sum_i \log(1 + \text{SNR} \lambda_i / n_t) < r \log_2 \text{SNR}$ — a function of the $m$ nonzero eigenvalues only. The outage density depends on the joint distribution of these eigenvalues, which is symmetric in $(n_t, n_r)$ up to a normalising constant that falls out under $\doteq$ .

Corner-point formula is symmetric

The formula $(n_t - k)(n_r - k)$ is symmetric in $(n_t, n_r)$ by inspection; so is the piecewise-linear interpolation. The DMT curve is therefore identical for $(n_t, n_r)$ and $(n_r, n_t)$ . $\blacksquare$

,

Example: Slope Cost per Unit of $r$ : $4 \times 4$ vs $2 \times 2$

For $2 \times 2$ and $4 \times 4$ channels, tabulate the DMT segment slopes and interpret the change in tradeoff structure as antennas scale.

Solution

$2 \times 2$ segments

Corners: $(0, 4), (1, 1), (2, 0)$ .

Segment $[0, 1]$ : slope $= (1 - 4)/1 = -3$ . Cost per unit $r$ = 3.
Segment $[1, 2]$ : slope $= (0 - 1)/1 = -1$ . Cost per unit $r$ = 1.

The first $r$ -unit costs 3x what the last unit costs.

$4 \times 4$ segments

Corners: $(0, 16), (1, 9), (2, 4), (3, 1), (4, 0)$ .

Segment $[0, 1]$ : slope $= (9 - 16)/1 = -7$ . Cost $= 7$ .
Segment $[1, 2]$ : slope $= (4 - 9)/1 = -5$ . Cost $= 5$ .
Segment $[2, 3]$ : slope $= (1 - 4)/1 = -3$ . Cost $= 3$ .
Segment $[3, 4]$ : slope $= (0 - 1)/1 = -1$ . Cost $= 1$ .

Successive odd numbers $7, 5, 3, 1$ — the cost per unit of $r$ decreases linearly as we climb the curve.

Interpretation

On a $4 \times 4$ channel, the cheap multiplexing gains are at the top of the curve (near $r = r_{\max}$ ): the last unit of $r$ costs only $1$ unit of diversity. The expensive gains are near $r = 0$ : the first unit of $r$ costs $n_t + n_r - 1 = 7$ units of diversity.

Operational reading. A system that is heavily reliability- constrained (URLLC, safety messages) should avoid climbing past $r = 0$ at all — each unit of $r$ costs dearly. A system that is heavily rate-constrained (eMBB) should climb to $r = r_{\max} - 1$ or $r_{\max}$ — the last units are nearly free. The crossover is around the mid-point of the curve, where the slope equals $-n_r$ (receive-diversity-only baseline).

Why the structure is universal. The slope decrement of $2$ per segment is a combinatorial consequence of the Wishart Vandermonde factor $\prod_{i<j}(\lambda_i - \lambda_j)^2$ . Each additional eigenvalue brought into outage contributes $2$ less to the exponent than the previous one — the "marginal diversity" of the $k$ -th eigenvalue shrinks linearly in $k$ . Section 12.2 of Tse-Viswanath draws out this combinatorial structure in detail.

,

DMT Corner Points for Common MIMO Configurations

Configuration $(n_t, n_r)$	$r_{\max}$	$d^*(0)$	Corner points	Initial slope $-(n_t + n_r - 1)$
$(1, 1)$	$1$	$1$	$(0, 1), (1, 0)$	$-1$
$(1, 2)$	$1$	$2$	$(0, 2), (1, 0)$	$-2$
$(2, 2)$	$2$	$4$	$(0, 4), (1, 1), (2, 0)$	$-3$
$(2, 4)$ or $(4, 2)$	$2$	$8$	$(0, 8), (1, 3), (2, 0)$	$-5$
$(3, 3)$	$3$	$9$	$(0, 9), (1, 4), (2, 1), (3, 0)$	$-5$
$(4, 4)$	$4$	$16$	$(0, 16), (1, 9), (2, 4), (3, 1), (4, 0)$	$-7$
$(8, 8)$	$8$	$64$	$(0, 64), (1, 49), \ldots, (7, 1), (8, 0)$	$-15$

🚨Critical Engineering Note

Rank Adaptation in LTE and 5G NR: Walking the DMT Curve

Every LTE and 5G NR receiver computes and reports a rank indicator (RI) — an integer $k \in \{1, 2, \ldots, \min(n_t, n_r)\}$ telling the base station how many spatial streams to transmit on the next slot. The RI is computed by the UE from measured channel coefficients and SNR: roughly, the UE picks the largest $k$ such that the expected throughput at rank $k$ exceeds the expected throughput at rank $k - 1$ .

This is literally walking the DMT curve. Each rank $k$ corresponds to a target multiplexing gain $r = k$ ; rank adaptation selects the corner point $(k, (n_t - k)(n_r - k))$ on the DMT curve that maximises throughput at the current SNR. At low SNR the optimal rank is $1$ — the corner $(1, (n_t - 1)(n_r - 1))$ gives maximum diversity consistent with a rate that grows. At high SNR the optimal rank is $\min(n_t, n_r)$ — the corner $(\min(n_t, n_r), 0)$ gives maximum rate but no outage protection (the latter comes from HARQ retransmissions, which is how 5G NR closes the reliability gap at high rank).

5G NR specifics. The RI is reported over PUCCH / PUSCH with a periodicity of $5$ – $80$ ms. The reported RI is fed to the scheduler, which allocates precoded resource elements. This is the adaptation loop that keeps the system operating on the DMT curve in real time.

Practical Constraints

•
RI feedback periodicity: $5$ – $80$ ms in 5G NR (type-I codebook).
•
Rank $\in \{1, 2, \ldots, \min(n_t, n_r, 8)\}$ for NR Rel-17 (up to $8$ layers in PDSCH).
•
Rank adaptation is per-UE, per-TTI; jointly optimised with modulation-coding-scheme (MCS) index.
•
Massive MIMO ( $n_t \ge 32$ ) uses rank adaptation with $k \ll n_t$ ; the diversity buffer $(n_t - k)(n_r - k)$ stays large.

📋 Ref: 3GPP TS 38.214 (physical layer procedures for data), §5.2.1.4 — CSI reporting

,

Common Mistake: The $\min$ -Corner Is $(\min(n_t, n_r), 0)$ , Not $(n_t, 0)$

Mistake:

On an asymmetric MIMO channel (e.g., $4 \times 2$ or $2 \times 4$ ), locating the right endpoint of the DMT curve at $r = n_t$ — i.e., assuming one can multiplex up to the number of transmit antennas.

Correction:

The maximum multiplexing gain is $r_{\max} = \min(n_t, n_r)$ , not $n_t$ or $n_r$ individually. On a $4 \times 2$ channel one can only support $2$ parallel streams, because the receiver has only $2$ antennas to disentangle them. The fourth-antenna excess at the transmitter provides diversity (available through precoding) but not multiplexing (which would require receive-side spatial resolution that doesn't exist).

Correspondingly, the DMT right endpoint is always at $(\min(n_t, n_r), 0)$ : for $4 \times 2$ it is $(2, 0)$ , not $(4, 0)$ . The left endpoint at $d^*(0) = n_t n_r$ uses the full product — excess transmit antennas do contribute diversity, just not multiplexing.

Mnemonic. Multiplexing is receiver-limited as well as transmitter- limited; diversity is only limited by the product $n_t \cdot n_r$ . The asymmetry enters through the Wishart rank $\min(n_t, n_r)$ — the number of nonzero eigenvalues of $\mathbf{H}\mathbf{H}^{H}$ .

Quick Check

What is the DMT curve of a $6 \times 2$ i.i.d. Rayleigh MIMO channel?

$d^*(r)$ has corners at $(0, 12), (1, 5), (2, 0)$

$d^*(r)$ has corners at $(0, 12), (1, 5), (2, 0), (3, 0), \ldots, (6, 0)$

$d^*(r)$ has corners at $(0, 8), (1, 3), (2, 0)$

$d^*(r)$ is the same as for $2 \times 2$

Correction:

d^*(r)

has corners at

(0, 12), (1, 5), (2, 0)

$\min(n_t, n_r) = 2$ , so there are $3$ corner points at $r = 0, 1, 2$ . $d^*(0) = 6 \cdot 2 = 12$ , $d^*(1) = 5 \cdot 1 = 5$ , $d^*(2) = 4 \cdot 0 = 0$ . The extra transmit antennas contribute to $d^*(0)$ (full $n_t n_r = 12$ diversity) but not to $r_{\max} = 2$ .

Three Ways to Read the DMT Curve

The DMT curve has three equivalent operational readings:

Information-theoretic. $d^*(r)$ is the outage-diversity exponent at rate-scaling $r \log_2 \text{SNR}$ . Every code's error probability decays no faster than $\text{SNR}^{-d^*(r)}$ at that rate. Tight via Zheng-Tse 2003.
Code-design. $d^*(r)$ is the diversity order achieved by the best possible space-time code at multiplexing gain $r$ . Families sitting at $(r, d)$ with $d < d^*(r)$ are sub-optimal; families on the curve are DMT-optimal (§4).
Scheduler. $d^*(r)$ is the reliability-vs-throughput budget that a rank-adaptive scheduler navigates in real time. The rank $k$ is the operating corner point; rank adaptation corresponds to jumping between adjacent corners.

All three readings are compatible and refer to the same asymptotic curve. The distinction matters when you are interpreting measurements: a code's actual BER at finite SNR depends on both its DMT operating point and its coding gain, which the DMT does not constrain.

,

Interpreting the DMT Curve