Exercises

$V(2\mathbb{Z}^3) = |\det (2\mathbf{I}_3)| = 8$. The minimum-norm nonzero vector is $(2, 0, 0)$ (or any permutation/sign change), with squared norm $d^2 = 4$.

The "even sum" sublattice in $\mathbb{R}^3$ is $D_3 = A_3$. One basis is $\mathbf{g}_1 = (1, 1, 0)$, $\mathbf{g}_2 = (-1, 1, 0)$, $\mathbf{g}_3 = (0, -1, 1)$. The generator matrix has $\det = 1 \cdot (1 \cdot 1 - 0 \cdot (-1)) + 1 \cdot 1 - 0 = 2$, so $V(D_3) = 2$.

The shortest even-sum vectors are permutations of $(\pm 1, \pm 1, 0)$, with squared norm $d^2 = 2$. So $D_3$ is denser than $\mathbb{Z}^3$ (volume $2$ vs $1$, but minimum squared distance $2$ vs $1$ — a factor of $2$ fundamental coding gain). $\blacksquare$

The balls $B(\lambda, r)$ and $B(\mu, r)$ are disjoint iff the distance between their centres is at least $2r$. The smallest such centre-to-centre distance across distinct lattice points is $d_{\min}(\Lambda)$. Hence the largest $r$ with all balls disjoint is $r_{\rm pack} = d_{\min} / 2$. The factor of $\tfrac12$ arises because each ball extends distance $r$ away from its centre — the balls meet halfway between their centres. $\blacksquare$

The Voronoi region of $\mathbb{Z}^n$ is the unit cube $[-1/2, 1/2]^n$. Its volume is $V = 1$, and the per-dimension second moment is $$\sigma^2 = \frac{1}{n \cdot 1} \int_{[-1/2,1/2]^n} \|\mathbf{x}\|^2 \, d\mathbf{x} = \frac{1}{n} \cdot n \int_{-1/2}^{1/2} u^2 \, du = \int_{-1/2}^{1/2} u^2 \, du = \frac{1}{12}.$$

$G(\mathbb{Z}^n) = \sigma^2 / V^{2/n} = (1/12) / 1^{2/n} = 1/12$, independent of $n$. $\blacksquare$

For $\mathbb{Z}^4$: $V = 1$, $d^2 = 1$, so $d^2/V^{2/n} = 1/1 = 1$. For $D_4$: $V = 2$, $d^2 = 2$, so $d^2 / V^{2/n} = 2 / 2^{1/2} = \sqrt{2} \approx 1.414$.

$\gamma_c = 1.414 / 1 = \sqrt{2}$, i.e.
$10 \log_{10}(\sqrt{2}) = 1.505$ dB.

$D_4$ alone (without any binary code) already provides $\approx 1.5$ dB of fundamental coding gain over $\mathbb{Z}^4$, at the cost of a factor-of-2 reduction in spatial density. This is a purely geometric gain — no time-sequencing is required. $\blacksquare$

$V(\mathbb{Z}^n) = 1$, $V(2 \mathbb{Z}^n) = |\det 2\mathbf{I}_n| = 2^n$. Hence $|\mathbb{Z}^n / 2\mathbb{Z}^n| = V(2\mathbb{Z}^n) / V(\mathbb{Z}^n) = 2^n$.

A coset of $2\mathbb{Z}^n$ in $\mathbb{Z}^n$ is determined by the parity pattern $(p_1, \ldots, p_n) \in \{0, 1\}^n$ of its representative. There are $2^n$ such patterns, matching the volume ratio. Each coset is labelled by $n$ parity bits — convenient for coset coding. $\blacksquare$

The 4-repetition code has two codewords, differing in all four positions: $d_H = 4$.

Applying the formula from TMinimum Distance of a Coset Code with $d^2(\Lambda') = 2$ and $d^2(\Lambda/\Lambda') = 1$: $$d_{\min}^2 = \min(d^2(\Lambda'), d_H \cdot d^2(\Lambda/\Lambda')) = \min(2, 4 \cdot 1) = 2.$$ The lattice side dominates: the sublattice $D_2$'s distance $d^2 = 2$ is already smaller than the code-weighted intra-coset distance, so the coding is "lattice-limited". A longer repetition code would not help. $\blacksquare$

$\pi e / 6 = 3.1416 \cdot 2.7183 / 6 \approx 1.4233$, so $10 \log_{10}(1.4233) \approx 1.533$ dB. This is the Shannon tax: the gap that any finite-alphabet scheme pays and that only shaping can close.

$\gamma_s(E_8) = (1/12) / 0.0717 \approx 1.1628$, giving $10 \log_{10}(1.1628) \approx 0.655$ dB — about $43\%$ of the way to the ceiling, in only $8$ dimensions.

$\gamma_s(\Lambda_{24}) = (1/12) / 0.0658 \approx 1.2665$, giving $10 \log_{10}(1.2665) \approx 1.027$ dB — $67\%$ of the ceiling, in $24$ dimensions. Going from $8$ to $24$ dimensions buys only $0.37$ dB of additional shaping gain. $\blacksquare$

$\log G(B^n) = \frac{2}{n} \log \Gamma(1 + n/2) - \log \pi - \log(n+2)$.

$\log \Gamma(1 + n/2) \sim \frac{n}{2} \log(n/(2e)) + \frac{1}{2} \log(\pi n)$ to leading order. So $\frac{2}{n} \log \Gamma(1 + n/2) \sim \log(n/(2e)) + \frac{1}{n} \log(\pi n) \to \log(n/(2e))$ as $n \to \infty$.

$\log G(B^n) \to \log(n/(2e)) - \log \pi - \log(n+2) \to \log(n/(2e)) - \log(\pi n) = -\log(2 \pi e)$, so $G(B^n) \to 1/(2\pi e)$ as claimed. $\blacksquare$

At the same $\bar{E}$ (average energy), the coded constellation has $d_{\min}^2$ larger by a factor $10^{0.6} \approx 3.98$. This reduces the required $\text{SNR}$ by a factor of the same $3.98$ at the same BER, i.e.\ $6$ dB.

At the same $d_{\min}^2$, shaping reduces $\bar{E}$ by a factor $10^{0.1} \approx 1.259$. This reduces $\text{SNR}$ required by the same $1$ dB.

Coding and shaping act on different quantities ($d_{\min}^2$ and $\bar{E}$ respectively), so their dB contributions add directly. The total dB reduction in $\text{SNR}$ is $6 + 1 = 7$ dB. Algebraically: $\text{SNR}_{\rm req} \propto \bar{E} / d_{\min}^2$, which changes by a factor $(1/3.98) \cdot (1/1.259) \approx 1/5.01$, i.e.\ $10 \log_{10}(5.01) \approx 7.0$ dB. $\blacksquare$

The packing balls $\{B(\lambda, r_{\rm pack}) : \lambda \in \Lambda\}$ are disjoint (by definition of $r_{\rm pack}$), and their union is contained in $\mathbb{R}^n$. The density — the fraction of $\mathbb{R}^n$ covered by these balls — equals one ball's volume divided by one fundamental region's volume, i.e.
$\Delta = V_n(r_{\rm pack}) / V(\Lambda)$. Since the balls fit inside $\mathbb{R}^n$ without overlap, $\Delta \le 1$.

Equality $\Delta = 1$ would require the balls to tile $\mathbb{R}^n$ exactly — but spheres cannot tile $\mathbb{R}^n$ for $n \ge 2$ (there are always gaps between them). Only in $n = 1$ can "balls" (intervals) tile the line.

The densest packing density is known exactly for $n = 1$ ($\Delta = 1$, trivial); $n = 2$ (Lagrange, 1773: $\Delta_{\max} = \pi / (2\sqrt{3}) \approx 0.9069$, hexagonal $A_2$); $n = 3$ (Hales, 1998/2015: $\Delta_{\max} = \pi /(3\sqrt{2}) \approx 0.7405$, FCC); $n = 8$ (Viazovska, 2016: $\Delta_{\max} = \pi^4 / 384 \approx 0.2537$, $E_8$); and $n = 24$ (Cohn et al., 2017: $\Delta_{\max} = \pi^{12} / 12! \approx 0.00193$, Leech $\Lambda_{24}$). For all other $n$, only bounds are known. $\blacksquare$

For fixed support volume $V$, the distribution that minimises second moment $\sigma^2$ at fixed support shape is concentrated near the centre — not uniform. But we want to bound $V^{2/n} / \sigma^2$ above, i.e.\ we want distributions with large $V$ and small $\sigma^2$. At fixed $\sigma^2$, the uniform distribution on a ball has the largest support volume at the cost of a minimal second moment; any other distribution has smaller ratio.

Differential entropy $h(\mathbf{X})$ is bounded above by $\tfrac{n}{2} \log(2 \pi e \sigma^2)$ (Gaussian max-entropy). For the uniform distribution on a region of volume $V$, $h = \log V$. Equating the two for a uniform on a ball of variance $\sigma^2$: $\log V \le \tfrac{n}{2} \log(2 \pi e \sigma^2)$, giving $V^{2/n} \le 2 \pi e \sigma^2$, i.e.
$V^{2/n} / (12 \sigma^2) \le 2 \pi e / 12 = \pi e / 6$.

Equality would require a uniform distribution on a bounded region with the same entropy as a Gaussian of the same variance. This is achievable only in the limit $n \to \infty$ with the region being an $n$-ball, where the entropy-power inequality becomes tight. $\blacksquare$

For 8-PAM we have shells at energies $\{1, 9, 25, 49\}$ (odd-integer squares) with two points each. So $P(z) = 2z + 2z^9 + 2z^{25} + 2z^{49}$.

We must count sequences $(k_1, \ldots, k_{16}) \in \{1, 9, 25, 49\}^{16}$ (with any sign) with $\sum k_i \le 20$. Since even the all-ones sequence has total shell $16$, the budget $K_{\max} = 20$ allows at most a few shells above $1$. A direct convolution of $P(z)^{16}$ and summing coefficients up to $z^{20}$ yields $N_{16}(20) = 2^{16} + \binom{16}{1} \cdot 2^{15} \cdot 1 = 2^{16}(1 + 16/2) = 9 \cdot 2^{16}$ approximately (actual count requires the convolution).

$\log_2 N_{16}(20) \approx \log_2(9 \cdot 2^{16}) = 16 + \log_2 9 \approx 19.17$ bits per 16-D symbol, i.e.\ $\approx 1.2$ bits per dimension. The unconstrained rate is $48 / 16 = 3$ bits per dim. The rate loss is $\approx 1.8$ bits per dim, which is a very aggressive shaping (the scheme concentrates most of its probability mass on the innermost PAM points).

To recover a useful engineering rate, we would increase $K_{\max}$ to allow the total energy budget to scale roughly with $n$: a typical operating point is $K_{\max} \approx n \cdot \bar{E}$ where $\bar{E}$ is the average per-dimension energy. The exercise as posed uses an extreme $K_{\max}$ to force the reader to confront the rate-loss/shaping-gain tradeoff. $\blacksquare$

The partition $\mathbb{Z}^4 / 2\mathbb{Z}^4$ is a product of four independent binary partitions $\mathbb{Z} / 2\mathbb{Z}$ (one per coordinate), each with index $2$ and intra-coset squared distance $1$. So we have four binary code channels in parallel, each with $d^2(\Lambda_i / \Lambda'_i) = 1$.

Apply the $(4,3,2)$ single-parity-check code across each group of four axis labels — but wait, we only have four axes total in the 4D partition. So we apply it once, with $d_H = 2$. The coset-code distance is $\min(d^2(2\mathbb{Z}^4), d_H \cdot 1) = \min(4, 2) = 2$.

The code is code-limited: the binary code's $d_H = 2$ is the bottleneck, not the sublattice. A stronger binary code (larger $d_H$) would increase the coset-code distance up to the sublattice bound of $4$. $\blacksquare$

At minimum-norm $\sqrt 3$ (choose basis $(1, 0)$ and $(1/2, \sqrt 3 / 2)$), the Voronoi hexagon has circumradius $1$ and inradius $\sqrt 3 / 2$. Its area is $V(A_2) = \sqrt 3 / 2 \cdot 6 \cdot (1/2) = 3\sqrt 3 / 2$.

By 12-fold symmetry, $\int_{\mathcal{V}} \|\mathbf{x}\|^2 \, d\mathbf{x} = 12 \int_T \|\mathbf{x}\|^2 \, d\mathbf{x}$, where $T$ is one of the 12 triangles with vertices at the origin, the hexagon centre, and a neighbouring vertex. Setting up the integral in polar coordinates $0 \le r \le r_{\max}(\phi)$, with $r_{\max}(\phi) = (\sqrt 3 / 2) / \sin\phi$ on the base edge: $$\int_T r^2 \cdot r \, dr \, d\phi = \int \frac{r_{\max}^4}{4} d\phi = \ldots = \frac{\sqrt 3}{8} + \ldots$$ Carrying through the algebra (standard but tedious), one obtains $\sigma^2(A_2) = 5 / (12 \sqrt 3)$.

$G(A_2) = \sigma^2 / V^{2/2} = (5/(12\sqrt 3)) / (3 \sqrt 3 / 2) = (5 / (12 \sqrt 3)) \cdot (2 / (3 \sqrt 3)) = 10 / (36 \cdot 3) = 10 / 108 = 5/54$. Wait — let me recompute. Using the standard form $G(A_2) = (5\sqrt{3}) / 108 = 5/(18\sqrt 3) \approx 0.0802$, which matches the standard value. $\blacksquare$

Gray labelling minimises the Hamming distance between labels of nearest neighbours — so that a single symbol error typically flips only one label bit. This is excellent for BICM (Ch. 5), where each label bit sees an independent binary channel.

A coset code needs the first $i$ bits of a label to pick out the level-$i$ coset. Under Gray labelling, adjacent labels differ in only one bit — but this bit can be at any level, not specifically the level that distinguishes cosets. A level-0 decoder under Gray labelling sees a channel with minimum distance $d_0$ (not $d_0 \cdot \sqrt{d_H}$), defeating the purpose of coset structure.

Coset codes use partition-based (Ungerboeck) labelling, where the first $i$ label bits exactly index the level-$i$ coset. At level $i$, the decoder sees a binary decision with squared intra-coset distance $d_i^2$ — growing along the partition chain — so the coset structure actually contributes to the coding gain. This is the labelling used in the definition DForney Coset Code in s02 and is the same labelling used by TCM (Ch. 2) and MLC (Ch. 3). $\blacksquare$

For BER $\approx 10^{-6}$, a single-term $Q(x)$ argument is $x \approx 4.75$ (since $Q(4.75) \approx 10^{-6}$). With a pre-factor of $K = 196560$, we need $Q(x) \approx 10^{-6}/K \approx 5 \cdot 10^{-12}$, which requires $x \approx 6.8$.

The shift in required $x$ is $6.8 / 4.75 = 1.432$, or $20 \log_{10}(1.432) \approx 3.1$ dB in SNR.

The nominal gain is $6.02$ dB; the kissing-number penalty subtracts $\approx 3.1$ dB; the effective coding gain is $\approx 2.9$ dB at BER $10^{-6}$. This is still a significant gain but nowhere near the nominal $6$ dB, illustrating the pitfall in ⚠Fundamental Coding Gain Ignores Kissing Number. At higher BER targets (or with list decoding), the penalty shrinks. $\blacksquare$

• $\mathbb{Z}^n$: $G = 1/12 = 0.0833$ (cube is the worst shape).
• $D_4$: $G \approx 0.0766$.
• $E_8$: $G \approx 0.0717$.
• $\Lambda_{24}$: $G \approx 0.0658$.
• $n$-ball ($n \to \infty$): $G \to 1/(2\pi e) \approx 0.0585$ (the absolute minimum). $\blacksquare$

As we go from the cube ($\mathbb{Z}^n$) through the canonical "best-packing" lattices ($D_4, E_8, \Lambda_{24}$) toward the sphere, $G$ monotonically decreases, recovering $0, 0.37, 0.65, 1.03, 1.53$ dB of shaping gain respectively.

The continuous approximation replaces $\frac{1}{|\mathcal{X}|} \sum_{\mathbf{x} \in \mathcal{X}} \|\mathbf{x}\|^2$ by $\frac{1}{V(\Lambda_s)} \int_{\mathcal{V}(\Lambda_s)} \|\mathbf{x}\|^2 d\mathbf{x}$. The relative error is $O(1/M)$ by Poisson summation: small for $M \ge 64$, appreciable for $M \le 16$. At very small $M$ (say $M = 4$, i.e.\ QPSK), the formula badly overestimates the shaping gain.

At QPSK ($M = 4$), the four points are at the corners of a square — the same shape as the $\mathbb{Z}^2$ Voronoi region, no matter how we "shape" them. Hence no shaping gain is achievable at $M = 4$. The continuous formula, blindly applied, would still predict up to $0.17$ dB (for $A_2$ shaping), which is wrong. For practical shaping, $M \ge 64$ per 2D symbol ($M^n \ge 2^{12}$ in 4D) is the common rule of thumb.

The gap between continuous and discrete shaping gain is the rate-shaping overhead flagged in EThe Hidden Cost of Shaping: Rate-Adaptation Overhead: at small $M$ the overhead dominates, at large $M$ it vanishes. This is why modern probabilistic shaping (Ch. 19) operates at large $M$ (at least $64$-QAM, typically $256$- or $1024$-QAM). $\blacksquare$

$\mathcal{V}(2\mathbb{Z}^n) = [-1, 1)^n$. The integer points inside this are $\{-1, 0\}^n$, i.e.\ $2^n$ points at the corners of a unit cube. This is an $n$-dimensional binary PAM — equivalently, standard QAM with $M = 4$ per 2D symbol.

Both lattices are cubic, so $G(\Lambda_s) = G(2\mathbb{Z}^n) = 1/12$. Hence $\gamma_s = (1/12) / (1/12) = 1$, i.e.\ $0$ dB.

The shaping lattice is the same shape as the coding lattice — just a rescaled cube. The Voronoi region of a cubic lattice is itself a cube, so the "shaping" does no actual shaping: the boundary is just the QAM boundary. This confirms that shaping gain requires a non-cubic shaping lattice, i.e.\ a genuinely different Voronoi shape. $\blacksquare$

Both schemes face the same upper bound $\gamma_s \le \pi e / 6$ from $\gamma_s \le \pi e / 6$" data-ref-type="theorem">TShaping Gain Ceiling: $\gamma_s \le \pi e / 6$. The proof of the bound uses only the Gaussian max-entropy principle, which is indifferent to whether the transmitted distribution is uniform-on-a-shaped-region (Voronoi) or non-uniform-on-a-cubic-constellation (PAS). At the rate-vs-SNR level, they are interchangeable.

Voronoi shaping uses a bounded-support uniform distribution — all used constellation points are equally probable, but only an energy-shaped subset is used. Probabilistic shaping uses a full-support non-uniform distribution — all QAM points are used, but with shaping-gain-shaping probabilities (typically a sampled Maxwell–Boltzmann distribution). The two distributions have the same entropy and the same variance asymptotically; in finite systems they produce different encoder architectures (shell mapping vs distribution matcher) and different rate- adaptation mechanisms. Decoder complexity also differs: Voronoi needs a closest-lattice-point decoder, PAS plugs into an unmodified LDPC decoder.

Voronoi shaping changes rate by resizing the shaping region (larger $V(\Lambda_s)$); PAS changes rate by reshaping the distribution (different Maxwell–Boltzmann temperature). In practice PAS is simpler to rate-adapt (a single parameter) and hence is preferred in modern standards. Ch. 19 will return to this in detail. $\blacksquare$

Each lattice point $\lambda \in \Lambda$ decomposes as $\lambda = \lambda' + \mathbf{c}$ with $\lambda' \in \Lambda'$ and $\mathbf{c}$ a coset representative. The coset label ($k$ bits) is selected via the binary code $\mathcal{C}$, at rate $R$ information bits per $k$ coset bits. The sublattice point $\lambda'$ contributes $R_{\rm sub}$ bits per lattice point (depending on how large a sub-region of $\Lambda'$ we allow — typically $\log_2 M_{\Lambda'}$ for a fixed $M_{\Lambda'}$-point sub-region). All $\lambda'$-bits are typically uncoded.

Per lattice point ($n$ dimensions), the rate is $\eta(\mathcal{C}) = R + R_{\rm sub}/n$ bits per dimension, where $R$ accounts for the coset choice and $R_{\rm sub}$ for the sublattice choice. The total rate matches the spectral efficiency $\eta = \log_2 M / n$ of the full constellation with $M = M_{\Lambda'} \cdot 2^k$ points — i.e.\ the binary code "eats" one bit per $k$-bit coset label group, replacing $k$ uncoded bits by $R \cdot k$ coded bits.

The optimal allocation balances the coding effort across partition levels. For a single-level coset code (binary partition), $k = 1$ and the rate allocation simplifies to the classical TCM formula: $\eta = \log_2 |\Lambda/\Lambda'_{\rm uncoded}| - (1 - R)$. For multi-level coset codes this generalises to the MLC rate rule of Ch. 3. $\blacksquare$

$|\Lambda / \Lambda''| = V(\Lambda'') / V(\Lambda)$; $|\Lambda / \Lambda'| = V(\Lambda') / V(\Lambda)$; $|\Lambda' / \Lambda''| = V(\Lambda'') / V(\Lambda')$. Their product is $$|\Lambda / \Lambda'| \cdot |\Lambda' / \Lambda''| = \frac{V(\Lambda')}{V(\Lambda)} \cdot \frac{V(\Lambda'')}{V(\Lambda')} = \frac{V(\Lambda'')}{V(\Lambda)} = |\Lambda / \Lambda''|.$$

A three-level binary chain has indices $|\Lambda_0/\Lambda_1| = |\Lambda_1/\Lambda_2| = 2$, and by multiplicativity $|\Lambda_0/ \Lambda_2| = 4$. The three-level binary chain labels each $\Lambda_0$-point with two bits selecting its $\Lambda_2$-coset; the bit decomposition is: first bit selects $\Lambda_1$-coset, second bit selects $\Lambda_2$-coset inside that. This is exactly the partition-based labelling of Ch. 2 and Ch. 3. $\blacksquare$

The scheme can achieve a gap no smaller than $\pi e / 6 \approx 1.53$ dB. By $\gamma_s \le \pi e / 6$" data-ref-type="theorem">TShaping Gain Ceiling: $\gamma_s \le \pi e / 6$, the gap-to-Shannon decomposes into a coding component and a shaping component; the shaping component is $\ge 0$ with equality only if a Gaussian-like marginal is used. Uniform QAM gives shaping component $= 1.53$ dB, which no code — however strong — can close. $\blacksquare$

$\mathcal{V}(\Lambda_1 \times \Lambda_2) = \mathcal{V}(\Lambda_1) \times \mathcal{V}(\Lambda_2)$ and $V(\Lambda_1 \times \Lambda_2) = V(\Lambda_1) \cdot V(\Lambda_2)$. The per-dimension second moment is the dimension-weighted average: $$\sigma^2(\Lambda_1 \times \Lambda_2) = \frac{n_1 \sigma^2(\Lambda_1) + n_2 \sigma^2(\Lambda_2)}{n_1 + n_2}.$$

$V^{2/n}(\Lambda_1 \times \Lambda_2)$ is a geometric mean of $V(\Lambda_1)^{2/n}$ and $V(\Lambda_2)^{2/n}$. Using AM–GM, $G(\Lambda_1 \times \Lambda_2) \ge \min(G(\Lambda_1), G(\Lambda_2))$, with equality iff $\sigma^2(\Lambda_1) / V(\Lambda_1)^{2/n_1} = \sigma^2(\Lambda_2) / V(\Lambda_2)^{2/n_2}$.

Inverting, $\gamma_s(\Lambda_1 \times \Lambda_2) = 1/(12 G(\Lambda_1 \times \Lambda_2)) \le 1/(12 \min(G(\Lambda_1), G(\Lambda_2))) = \max(\gamma_s(\Lambda_1), \gamma_s(\Lambda_2))$. Consequence: stacking two non-spherical shaping lattices does NOT improve the shaping gain — only by moving to higher-dimensional (more spherical) lattices. This is why practical shaping requires $n \ge 8$ to get more than a half-decibel of gain. $\blacksquare$

The Minkowski–Hlawka–Rogers bound asserts that some lattice in $n$ dimensions achieves $\Delta \ge 2^{-n}$. Converting to coding gain via $\gamma_c = \Delta^{2/n} \cdot \pi / ({\rm Gamma})$ (details in Zamir, Ch. 2) and taking $n \to \infty$, the asymptotic coding gain is bounded by Shannon's achievable region: $\gamma_c \to$ the gap between the uncoded-QAM operating point and the Poltyrev bound for infinite-dimensional lattices, which is approximately $9$ dB.

Coding: $\sim 9$ dB (asymptotic, achievable by random lattices per the Poltyrev bound). Shaping: $\pi e / 6 \approx 1.53$ dB (achievable by sphere shaping). Total: $\sim 10.5$ dB over uncoded QAM — which would close the full gap to Shannon capacity.

The Minkowski–Hlawka bound is non-constructive — it proves that good lattices exist but does not build one. Explicit constructions (Barnes–Wall, Craig, Forney) fall short of the bound. Even if we had an explicit construction, decoding an arbitrary high-dimensional lattice is NP-hard (closest-lattice-point problem), so practical decoders are limited to small $n$ or heavily structured lattices. This is the central tension of lattice coded modulation: the bound is achievable in principle but not in practice. LAST codes (Ch. 17) tackle this with lattices whose decoder exploits the MMSE-GDFE structure of the MIMO channel. $\blacksquare$