Exercises

$d_\mathrm{ARQ}(r, L) = L \cdot d^{*}(r/L)$, where $d^{*}(\cdot)$ is the static Zheng-Tse DMT. Domain: $r \in [0, L \min(n_t, n_r)]$, $L \ge 1$.

$d_\mathrm{ARQ}(r, 1) = 1 \cdot d^{*}(r) = d^{*}(r)$ — the static DMT of Ch. 12. Consistent with "no ARQ = one round".

For any fixed $r$: $d^{*}(r/L) \to d^{*}(0) = n_t n_r$. Hence $d_\mathrm{ARQ}(r, L) = L \cdot d^{*}(r/L) \to L \cdot n_t n_r \to \infty$. This is the ergodic-capacity regime.

For $r \in \{0, 0.5, 1, 2, 3, 4\}$: $r/2 \in \{0, 0.25, 0.5, 1, 1.5, 2\}$. On segment $[0, 1]$: $d^{*}(r/2) = 4 - 3(r/2) = 4 - 1.5 r$. On segment $[1, 2]$: $d^{*}(r/2) = 1 - 1(r/2 - 1) = 2 - r/2$. Values: $\{4, 3.625, 3.25, 1, 0.5, 0\}$.

Wait — for $r = 2$, $r/2 = 1$, giving $d^{*}(1) = 1$. For $r = 3$, $r/2 = 1.5$, giving $d^{*}(1.5) = 0.5$. For $r = 4$, $r/2 = 2$, giving $d^{*}(2) = 0$.

$d_\mathrm{ARQ}(r, 2) = 2 \cdot d^{*}(r/2)$. Values: $\{8, 7.25, 6.5, 2, 1, 0\}$.

$d^{*}(r)$ at $r \in \{0, 0.5, 1, 2, 3, 4\}$: $\{4, 2.5, 1, 0, 0, 0\}$ (zero beyond $r_\mathrm{max} = 2$).

ARQ-DMT is strictly greater than the static DMT for all $r > 0$ where both are defined, and it extends the domain beyond $r_\mathrm{max}$ (the static DMT is $0$ for $r > \min(n_t, n_r)$; the ARQ-DMT extends to $r = L \cdot \min(n_t, n_r) = 4$).

CC sends the same rate-$r$ codeword at every round. Each round contributes exponent $d^{*}(r)$, and the $L$ rounds are independent, so the total exponent is $L \cdot d^{*}(r)$.

IR uses a mother code of rate $r/L$ (spread across $L$ rounds with fresh parity each round). The effective per-round rate seen by the decoder is $r/L$, contributing exponent $d^{*}(r/L)$ per round. Total: $L \cdot d^{*}(r/L)$.

$d^{*}$ is non-increasing: $d^{*}(r/L) \ge d^{*}(r)$ with equality iff $r/L = r$ (i.e., $L = 1$) or iff $d^{*}$ is constant on $[r/L, r]$ (only at $r = 0$). Hence IR's exponent exceeds CC's for all $r > 0$ with $L > 1$.

Operationally: IR lowers the per-round rate to a regime where the static DMT is steeper — more diversity per round — and still collects $L$ independent fading looks. CC keeps the per-round rate high (steep regime unused) and only collects the $L$ looks.

$d_\mathrm{ARQ}(r, L) = L \cdot d^{*}(r/L) = r \cdot \frac{d^{*}(r/L)}{r/L}$.

Claim: $r \mapsto d^{*}(r)/r$ is non-increasing on (0, r_\max] for the piecewise-linear Zheng-Tse DMT. At any interior point $r = k$ (integer), the right derivative of $d^{*}$ jumps down (the curve gets flatter going right), so $d^{*}(r)/r$ decreases as $r$ grows. More directly: $d^{*}(r) = (n_t - r)(n_r - r)$ at integer $r$ is a decreasing function of $r$; dividing by $r$ only accelerates the decrease.

As $L$ grows, $r/L$ decreases, so $d^{*}(r/L) / (r/L) = d^{*}(r/L) \cdot L / r$ non-decreases (by the step above). Multiplying by $r$: $L \cdot d^{*}(r/L)$ non-decreases in $L$. $\blacksquare$

$d^{*}(0) = 9$, $d^{*}(1) = 4$, $d^{*}(2) = 1$, $d^{*}(3) = 0$. Piecewise-linear: slope $-5$ on $[0, 1]$, slope $-3$ on $[1, 2]$, slope $-1$ on $[2, 3]$.

$d_\mathrm{ARQ}(r, 2) = 2 \cdot d^{*}(r/2)$. Evaluate at $r = 0, 0.5, 1, 2, 3, 4, 5, 6$: $d^{*}(r/2)$ at $\{0, 0.25, 0.5, 1, 1.5, 2, 2.5, 3\}$ $= \{9, 7.75, 6.5, 4, 2.5, 1, 0.5, 0\}$, multiplied by 2: $\{18, 15.5, 13, 8, 5, 2, 1, 0\}$.

$d^{*}(r)$ at $\{0.5, 1, 2, 3\}$: $\{6.5, 4, 1, 0\}$. $d_\mathrm{ARQ}(r, 2)$ at $\{0.5, 1, 2, 3\}$: $\{15.5, 13, 8, 5\}$. The ARQ-DMT is strictly above the static DMT for every $r$ in the common domain.

On a plot with $r$ on the horizontal axis and $d^{*}$ on the vertical axis: the static DMT goes from $(0, 9)$ down to $(3, 0)$. The ARQ-DMT goes from $(0, 18)$ down to $(6, 0)$ — it is a scaled-and-stretched copy of the static DMT with both axes scaled by $L = 2$.

$10^{-6} = \text{SNR}^{-9} \Rightarrow \log_{10} \text{SNR} = 6/9 \approx 0.67$ decades, i.e., $\text{SNR} \approx 10^{0.67} \approx 4.6$ (linear). In dB: $6.7$ dB.

$10^{-6} = \text{SNR}^{-3} \Rightarrow \log_{10} \text{SNR} = 6/3 = 2$ decades, i.e., $\text{SNR} = 10^2 = 100$ (linear). In dB: $20$ dB.

CC needs $20 - 6.7 = 13.3$ dB more SNR than IR to hit the same BLER. This is the ARQ-DMT payoff at the asymptote; at moderate SNR the gap shrinks because of finite-SNR coding-gain offsets and sub-asymptotic slopes (the empirical diversity is $\sim 0.7 d$ at 15 dB). But the order-of-magnitude is clear: IR provides a $\sim 10$ dB advantage over CC on $2 \times 2$ with $L = 3$ at $r = 1$.

$v = 3$ km/h $= 0.833$ m/s. $f_D = v f_c / c = 0.833 \cdot 3.5 \times 10^9 / (3 \times 10^8) = 9.7$ Hz.

$T_\mathrm{coh} \approx 1 / f_D \approx 103$ ms (loose estimate; tighter bounds give $T_\mathrm{coh} \approx 9 / (16 \pi f_D) \approx 18$ ms).

$T_\mathrm{rtt} = 4$ ms $\ll T_\mathrm{coh} \approx 20$ ms. Consecutive HARQ rounds fall inside the coherence time — they see nearly-identical channel realisations.

Without frequency hopping, the $L = 4$ rounds are correlated and the effective number of independent realisations is $L_\mathrm{eff} \approx T_\mathrm{rtt} \cdot L / T_\mathrm{coh} \cdot L \approx 1$. The ARQ-DMT gain is roughly $L_\mathrm{eff} \cdot d^{*}(r/L_\mathrm{eff}) \approx d^{*}(r)$ — essentially the static DMT with no ARQ benefit.

5G NR per-round frequency hopping (§5, Def. DPer-Round Frequency Hopping in 5G NR) deliberately hops the PRB allocation across retransmissions. If the frequency offset exceeds the coherence bandwidth $B_\mathrm{coh} \sim 1 / \tau_ \mathrm{rms}$ (typically a few MHz in indoor scenarios), the rounds become approximately independent again, restoring the full ARQ-DMT diversity. Frequency hopping is the de facto enabler of ARQ-DMT at low mobility.

$L_\mathrm{URLLC} \le T_\mathrm{budget} / T_\mathrm{rtt} = 1 / 0.5 = 2$ rounds at most.

Assuming $4 \times 4$ MIMO: $d_\mathrm{ARQ}(1, 2) = 2 \cdot d^{*}(0.5) = 2 \cdot (4 \cdot 3.5) \cdot (0.5 \to 0)$... let me redo this. On $4 \times 4$: $d^{*}(0) = 16$, $d^{*}(1) = 9$, $d^{*}(2) = 4$, $d^{*}(3) = 1$, $d^{*}(4) = 0$. Linear on $[0,1]$: slope $-7$; on $[1,2]$: slope $-5$; etc. $d^{*}(0.5) = 16 - 7 \cdot 0.5 = 12.5$. So $d_\mathrm{URLLC} = 2 \cdot 12.5 = 25$.

$d_\mathrm{eMBB} = 4 \cdot d^{*}(0.25) = 4 \cdot (16 - 7 \cdot 0.25) = 4 \cdot 14.25 = 57$.

URLLC: $d = 25$; eMBB: $d = 57$. The URLLC latency budget costs $\sim 32$ units of diversity exponent — a $\sim 32 / 57 = 56\%$ reduction. In SNR terms at target BLER $10^{-5}$: URLLC needs $\log_{10}(10^5) / 25 \cdot 10 = 2$ dB of margin above the outage threshold; eMBB needs $5/57 \cdot 10 = 0.88$ dB. The $\sim 1$ dB difference is the "URLLC latency tax". In practice URLLC compensates with low-rate coding ($R_m = 1/5$ instead of $1/3$), larger antenna arrays, and PDCP duplication — all of which restore the reliability budget at the cost of air-time / equipment.

For any ARQ protocol carrying message $W$ with information $H(W) = L r \log_2 \text{SNR}$ across $L$ rounds, Fano's inequality says $P_e \ge 1 - (I_L + 1) / H(W)$, where $I_L$ is the $L$-round cumulative mutual information. Rearranging, $P_e \ge P_\mathrm{out}(L r \log_2 \text{SNR}, L) - o(1)$ at large block length.

The $L$-round outage is $$P_\mathrm{out}(L r \log_2 \text{SNR}, L) = \Pr\!\left[ \sum_{\ell=1}^L I(\mathbf{H}_\ell) < L r \log_2 \text{SNR}\right],$$ where $I(\mathbf{H}_\ell) = \log_2 \det(\mathbf{I} + \tfrac{\text{SNR}}{n_t}\mathbf{H}_\ell\mathbf{H}_\ell^H)$. By the Wishart large-deviations argument of Ch. 12 applied to each independent round, the outage event factorises into $L$ independent eigenvalue-exponent LPs.

Reparametrise $\lambda_i^{(\ell)} = \text{SNR}^{-\alpha_i^{(\ell)}}$ per round $\ell$. The outage constraint becomes $\sum_\ell \sum_i (1 - \alpha_i^{(\ell)})^+ < L r$. By symmetry / convexity of the static LP's objective (weight vector $(2i - 1 + n_r - n_t)$), the joint LP attains its minimum when all $L$ rounds share the same per-round multiplexing $r_\ell = r$, i.e., $\sum_i (1 - \alpha_i^{(\ell)})^+ = r$ for each $\ell$. The per-round LP value is $d^{*}(r)$.

The above gives $\sum_\ell d^{*}(r) = L \cdot d^{*}(r)$, but the theorem claims $L \cdot d^{*}(r/L)$. The discrepancy is the rate convention (see ⚠The Rate Convention: Per-Round vs Long-Term). If $r$ denotes the per-round multiplexing gain, then $d_\mathrm{ARQ} = L \cdot d^{*}(r)$. If $r$ denotes the long-term effective multiplexing (the convention of El Gamal-Caire-Damen), then the per-round rate is $r / L$ and the exponent is $L \cdot d^{*}(r/L)$. Both statements are correct; only the parametrisation differs.

With the long-term convention: $P_e \ge \text{SNR}^{-L \cdot d^{*}(r/L)} \cdot \text{poly}(\log\text{SNR})$. This matches the upper bound achieved by random Gaussian codes (Thm. TARQ-DMT (El Gamal-Caire-Damen 2006) achievability), so the ARQ-DMT is tight. $\blacksquare$

Using the CDA-generator construction, the codeword-difference matrix in round $\ell$ satisfies $\det(\boldsymbol{\Delta}_\ell \boldsymbol{\Delta}_\ell^H) \ge c(r) > 0$ uniformly in SNR. Substituting into the Chernoff PEP bound gives per-round exponent $d^{*}(r/L)$ in the ARQ-DMT rate convention.

By Thm. 2 of El Gamal-Caire-Damen 2004 (Ch. 17), the MMSE-GDFE lattice decoder achieves the same DMT exponent as the joint ML decoder. Hence the per-round IR-LAST exponent matches the Gaussian-random-code exponent $d^{*}(r/L)$.

With common random dither, the $L$-round PEP averages over the dither and factorises into $L$ independent per-round PEPs (via independence of $\mathbf{H}_{1}, \ldots, \mathbf{H}_{L}$). The joint exponent is the sum: $L \cdot d^{*}(r/L) = d_\mathrm{ARQ}(r, L)$.

$P_e^\mathrm{IR\text{-}LAST} \doteq \text{SNR}^{-L \cdot d^{*}(r/L)}$ matches the ARQ-DMT upper bound, so IR-LAST achieves it. $\blacksquare$

$d^{*}(0.25)$ on $2 \times 2$: interpolate $(0, 4)$ and $(1, 1)$ at $r = 0.25$: $d^{*}(0.25) = 4 - 3 \cdot 0.25 = 3.25$. $d_\mathrm{ARQ}(1, 4) = 4 \cdot 3.25 = 13$.

Empirical slope $\approx 7$; theoretical slope $= 13$. The ratio is $7/13 \approx 0.54$.

The empirical slope is $\sim 54\%$ of the asymptotic slope — a sizable finite-SNR gap. Typical for BLER in the $10^{-4}$ range on $2 \times 2$ at moderate SNR. The remaining 46% of asymptotic diversity materialises only above $\sim 20$–$25$ dB (the BLER curve steepens as SNR grows, approaching slope 13 in the deep tail).

For a BLER target of $10^{-5}$ (URLLC) at this operating point, the system needs $\sim 15$–$20$ dB of SNR margin, but the empirical slope tells us we must budget on the shallower slope at the target BLER, not the asymptotic slope. Link-level simulation (not just the ARQ-DMT formula) is essential for design sizing.

$T_\ell =$ "first success at round $\ell$" $= \{I_{\ell-1} < L R\} \cap \{I_\ell \ge L R\}$, where $I_\ell = \sum_{k \le \ell} I(\mathbf{H}_{k})$. $\Pr(T_\ell) = P_\mathrm{out}(L R, \ell - 1) - P_\mathrm{out}(L R, \ell)$.

$\Pr(\text{success}) = 1 - P_\mathrm{out}(L R, L)$ (success occurs iff the $L$-round outage does not).

$\mathbb{E}[\text{rounds}] = \sum_{\ell=1}^L \ell \cdot \Pr(T_\ell) + L \cdot P_\mathrm{out}(L R, L)$ (if all $L$ rounds fail, still $L$ rounds used). Using the identity $\sum_\ell \ell (P_{\ell-1} - P_\ell) = L P_L + \sum_\ell P_\ell$ (Abel summation): $$\mathbb{E}[\text{rounds}] = \sum_{\ell=0}^{L-1} P_\mathrm{out} (L R, \ell) + 1 = 1 + \sum_{\ell=1}^{L-1} P_\mathrm{out} (L R, \ell).$$

Information bits per block: $L R N$ (if decoded). Channel uses per block: $\mathbb{E}[\text{rounds}] \cdot N$. Effective rate: $$\eta_\mathrm{eff} = \frac{L R \cdot (1 - P_\mathrm{out}(L R, L))} {1 + \sum_{\ell=1}^{L-1} P_\mathrm{out}(L R, \ell)}.$$ At high SNR, all $P_\mathrm{out}(\cdot, \ell) \to 0$ so $\eta_\mathrm{eff} \to L R$. At low SNR, $P_\mathrm{out} \to 1$ so $\eta_\mathrm{eff} \to 0$.

Each ARQ round uses $N$ channel uses. The per-round outage exponent is the static DMT $d^{*}(r/L)$, which requires the per-round block length $N \ge n_t + n_r - 1$ (by Thm. 9.1 of Tse-Viswanath 2005, equivalently the Zheng-Tse theorem's block- length condition).

The per-round outage exponent truncates: $d^{*}_{\rm trunc}(r, N) = (n_t - r)(N - r)$ at integer corners (Ch. 12 §5). The ARQ-DMT formula becomes $d_\mathrm{ARQ}(r, L; N) = L \cdot d^{*}_{\rm trunc}(r/L, N)$, which is strictly less than $L \cdot d^{*}(r/L)$ whenever $N < n_t + n_r - 1$ at the operating $r/L$.

In high-mobility mmWave scenarios, per-round block length $N$ can be as short as 1–2 OFDM symbols. This truncates both the static and ARQ-DMT curves. A system running $L = 4$ HARQ rounds with $N = 1$ on $4 \times 4$ MIMO achieves much less than $4 \cdot d^{*}(r/4) = 4 \cdot 16 = 64$ at $r \to 0$; more like $4 \cdot (4 \cdot 1) = 16$ — a factor of 4 loss from block-length truncation.

After $L$ rounds of CC, the effective SNR is $L \cdot \text{SNR}$ (SNR combining on a single symbol replicated $L$ times).

The single-round capacity at effective SNR $L \text{SNR}$ is $\log_2(L \text{SNR} + 1) \approx \log_2 \text{SNR} + \log_2 L$ at high $\text{SNR}$. The per-round rate $R$ is typically chosen at $\log_2 \text{SNR}$, so the "extra" $\log_2 L$ margin is wasted — the decoder could already decode at round 1 with high probability.

At very high SNR, $P_\mathrm{out}(R, 1) \to 0$ and the ACK is emitted on round 1. CC never transmits further rounds; it degenerates to a single-shot scheme with rate $R$. The diversity benefit of CC over single-shot is thus only at moderate SNR where round 1 fails with non-negligible probability.

IR is different: its per-round rate is $r/L$, so even at very high SNR the cumulative rate after $L$ rounds is $r$ — exactly the target. IR "uses up" the available SNR productively across rounds; CC doesn't.

$P_\mathrm{out} = \{0.6, 0.35, 0.15, 0.05\}$ at $\ell = 1, 2, 3, 4$. Expected rounds $= 1 + 0.6 + 0.35 + 0.15 = 2.10$.

Each $P_\mathrm{out}(\ell r, \ell) \doteq \text{SNR}^{-\ell d^{*} (r/\ell)}$. On $2 \times 2$: $d^{*}(1) = 1$, $d^{*}(0.5) = 2.5$, $d^{*}(1/3) = 3$, $d^{*}(0.25) = 3.25$. Diversity exponents per round: $\{1, 5, 9, 13\}$. Per-5-dB decay factors: $10^{-d/2} \approx 10^{-0.5, -2.5, -4.5, -6.5}$, i.e., $\{0.32, 3.2e-3, 3.2e-5, 3.2e-7\}$. Scaled: $\{0.6, 0.35, 0.15, 0.05\} \times \{0.32, 3.2e-3, 3.2e-5, 3.2e-7\}$ $\approx \{0.19, 1.1e-3, 4.7e-6, 1.6e-8\}$. Expected rounds $= 1 + 0.19 + 1.1e-3 + 4.7e-6 \approx 1.19$.

At 15 dB (10 dB above baseline): outages decay further by factor $\{10^{-1}, 10^{-5}, 10^{-9}, 10^{-13}\}$. Round 1 dominates at these SNRs; expected rounds $\approx 1.02$. At 20 dB: round 1 outage $\approx 6 \cdot 10^{-3}$; expected rounds $\approx 1.006$.

SNR: 5 dB $\to$ 2.1 rounds (16 ms); 10 dB $\to$ 1.19 rounds (10 ms); 15 dB $\to$ 1.02 rounds (8.2 ms); 20 dB $\to$ 1.006 rounds (~8 ms). The expected latency collapses rapidly above the outage threshold; this is exactly why the scheme is called "hybrid ARQ" — at high SNR it behaves like a one-shot code, at low SNR like an $L$-shot code.

Without dithering, a CDA lattice code transmits a deterministic codeword given the information bits. The PEP for a specific codeword pair depends on the algebraic determinant of their difference. For the worst-case pair, this determinant is the NVD constant $c(r)$ — a lower bound, but not tight at finite SNR.

The Gaussian random-code ensemble has a distribution over codewords, and the ensemble-average PEP matches the Gallager-exponent formula. This is the achievability argument of §2.

The common random dither $\mathbf{d} \sim \mathrm{Unif}(\mathcal{V}(\Lambda_c))$ makes the transmitted codeword uniformly distributed over the Voronoi region of the shaping lattice. The average PEP over the dither matches (asymptotically) the Gaussian-ensemble PEP, restoring the random-coding achievability.

Without the dither, the IR-LAST PEP is bounded by the worst-case codeword-pair's bound, which does not quite match the Gaussian-random-code bound at every SNR. The asymptotic exponent is still correct (the NVD constant dominates), but the coding gain is worse and the finite-SNR performance degrades. Moreover, the joint PEP across rounds no longer factorises cleanly, so the $L \cdot d^{*}(r/L)$ accounting can fail at non-integer rates.

$\mathcal{O}_L = \{\sum_{\ell=1}^L \sum_{i=1}^m (1 - \alpha_i^{(\ell)})^+ < L r\}$, i.e., the sublevel set of the convex function $\Phi(\{\boldsymbol{\alpha}^{(\ell)}\}) = \sum_\ell \sum_i (1 - \alpha_i^{(\ell)})^+$ below level $L r$.

$(1 - \alpha)^+ = \max(1 - \alpha, 0)$ is convex in $\alpha$ (max of two linear functions). The sum over $(i, \ell)$ is convex.

The sublevel set of a convex function is convex. Hence $\mathcal{O}_L$ is a convex subset of $(\mathbb{R}_+^m)^L$.

The LP $\inf_{(\alpha) \in \mathcal{O}_L} \sum_{\ell, i} (2i - 1 + n_r - n_t) \alpha_i^{(\ell)}$ over a convex set with linear objective has a vertex-optimal solution, which is attained at $(\alpha_i^{(\ell)} = 0 \text{ or } 1)$ corners — the per-round equal-allocation vertices. This is exactly the achievability argument we used to derive $L \cdot d^{*}(r/L)$.

$L \le 0.5 / 0.25 = 2$ rounds.

$d^{*}(0.25)$ on $4 \times 4$: $(4 - 0.25)(4 - 0.25) = 14.0625$ (continuous formula); actually, piecewise interpolation between $(0, 16)$ and $(1, 9)$ gives $d^{*}(0.25) = 16 - 7 \cdot 0.25 = 14.25$. $d_\mathrm{ARQ}(0.5, 2) = 2 \cdot 14.25 = 28.5$.

$\text{SNR}^{-28.5} = 10^{-6} \Rightarrow \log_{10} \text{SNR} = 6/28.5 \approx 0.21$ decades $= 2.1$ dB.

On a $4 \times 4$ URLLC link at $r = 0.5$ with $L = 2$ rounds, the system needs only $\sim 2$ dB of SNR margin above the outage threshold to hit $10^{-6}$ reliability — a very modest requirement. The ARQ-DMT formula gives a slope of 28.5, which is very steep; even a half-round HARQ budget is enormously valuable. This is why 5G NR URLLC relies heavily on spatial diversity (large antenna arrays) — to get the static DMT exponent as large as possible, so that even $L = 1$ or $L = 2$ yields adequate reliability.

$P_\mathrm{out}(\cdot, \ell) \to 0$ for all $\ell \ge 1$. Numerator: $L R \cdot 1 = L R$. Denominator: $1 + 0 + \cdots + 0 = 1$. Hence $\eta_\mathrm{eff} \to L R$ — the system behaves as if it transmits successfully at the first round every time, and the per-block throughput is the cumulative rate $L R$ across the full block.

$P_\mathrm{out}(\cdot, \ell) \to 1$ for all $\ell$. Numerator: $L R \cdot (1 - 1) = 0$. Denominator: $1 + (L-1) \cdot 1 = L$. Hence $\eta_\mathrm{eff} \to 0$ — at low SNR, the system spends all $L$ rounds per block and still fails, so the throughput collapses.

Between the two limits, $\eta_\mathrm{eff}$ has a sigmoid-like rise vs SNR. The steepness of the rise is governed by the ARQ-DMT exponent: high $L$ and high $n_t n_r$ give steeper transitions. Interactive plot 📊5G NR HARQ-IR Throughput Envelope illustrates this.

Zheng-Tse (2003) gave the static DMT $d^{*}(r)$ for a single- shot MIMO channel. Retransmission / ARQ was studied as a throughput mechanism (Wozencraft-Jacobs tradition, Caire-Tuninetti 2001 for AWGN) but not as a resource that reshapes the fundamental tradeoff curve.

El Gamal, Caire, and Damen introduced the delay dimension $L$ as a third axis of the DMT. Their theorem $d_\mathrm{ARQ}( r, L) = L \cdot d^{*}(r/L)$ showed that ARQ is not just a throughput trick but a genuine information-theoretic resource that multiplies the diversity exponent at fixed rate. They also constructed the first explicit ARQ-DMT-optimal codes (IR-LAST).

The ARQ-DMT theorem underpins HARQ design in every modern cellular standard. LTE Rel-8 (2009) and 5G NR Rel-15 (2018) both use LDPC + circular-buffer rate matching + redundancy versions — the engineering realisation of the information- theoretic IR-HARQ analysed by El Gamal-Caire-Damen. The paper is a textbook example of how a clean theorem shapes standards a decade later.