Chapter Summary

Chapter Summary

Key Points

• 1.

  A nested lattice code is a pair $(\Lambda_c, \Lambda_s)$ with $\Lambda_s \subset \Lambda_c$, giving a codebook $\mathcal{C} = \Lambda_c \cap \mathcal{V}(\Lambda_s)$ of $|\Lambda_c / \Lambda_s| = V(\Lambda_s)/V(\Lambda_c)$ codewords at rate $R = (1/n) \log_2 |\Lambda_c / \Lambda_s|$ bits per real dimension. The fine lattice $\Lambda_c$ carries the coding gain, the coarse lattice $\Lambda_s$ carries the shaping gain; the two decouple cleanly into independent design problems.

• 2.

  Erez–Zamir (2004): nested lattice codes with MMSE scaling $\alpha = \text{SNR}/(1 + \text{SNR})$ and a dither $\mathbf{d}$ uniform on $\mathcal{V}(\Lambda_s)$ achieve the AWGN capacity $\tfrac12 \log_2(1 + \text{SNR})$ exactly, provided $\Lambda_c$ is Poltyrev-good (Loeliger random lattice) and $\Lambda_s$ is Rogers-good ($G(\Lambda_s) \to 1/(2 \pi e)$). The crypto-lemma (dither + modulo spread the transmit signal to uniform on $\mathcal{V}(\Lambda_s)$) is the single trick behind the proof and recurs throughout lattice-network theory.

• 3.

  Shaping gain is bounded by $\pi e / 6 \approx 1.53$ dB in every dimension, with the bound achieved only as $n \to \infty$. Concrete lattices: $D_4^*$ gives $0.37$ dB, $E_8$ gives $0.66$ dB, $\Lambda_{24}$ gives $1.03$ dB, and dimension-$128$ Barnes–Wall reaches $1.35$ dB. V.34 trellis shaping and ADSL harvested most of the $1.53$ dB in the 1990s; 5G NR's probabilistic shaping (Ch. 19) is the modern equivalent.

• 4.

  Costa (1983): the dirty-paper channel $\mathbf{y} = \mathbf{x} + \mathbf{s} + \mathbf{w}$ with non-causal transmitter knowledge of $\mathbf{s}$ has the same capacity as the clean channel, independent of $\|\mathbf{s}\|^2$. The lattice realisation (Erez–Shamai–Zamir 2005) adds one mod-$\Lambda_s$ at the transmitter and reuses the Erez–Zamir MMSE + dither construction verbatim. Tomlinson–Harashima precoding is the scalar cousin, fielded in V.34, ADSL, VDSL, G.fast, and MU-MIMO vector perturbation.

• 5.

  The closest-lattice-point problem is NP-hard worst-case but polynomial $O(n^3)$ average-case at high SNR, via the Pohst / Schnorr–Euchner sphere decoder (Viterbo–Boutros 1999, Damen–El Gamal–Caire 2003). QR decomposition reduces the problem to triangular layers; Schnorr–Euchner zigzag + norm pruning solves it in polynomial average time. For poorly-conditioned lattices, LLL reduction ($O(n^6)$ one-time) restores the $O(n^3)$ per-decoding behaviour.

• 6.

  The "proof pattern" of this chapter — crypto-lemma + MMSE $\alpha$ + mod-$\Lambda_s$ — is the backbone of lattice-coding theory. It achieves capacity on the AWGN channel (s02), cancels interference on the dirty-paper channel (s04), and generalises without modification to the matrix case (LAST codes, Ch. 17). Master it here, and the rest of Part IV falls into place.

• 7.

  The total capacity gap of any lattice code is additive (in dB) in coding gap and shaping gap, which factor cleanly through the two sublattices. For the designer: pick any good $\Lambda_c$ and any good $\Lambda_s$ independently, and the total gap to Shannon is the sum of the two individual shortfalls. This clean decoupling is why BICM+LDPC+CCDM in 5G NR and the DMT-optimal LAST codes of Ch. 17 can both be cast in the nested-lattice framework without crosstalk between the design problems.