Ferkans — Interactive Telecom Tutor

Shaping Gain as a Universal Lattice Invariant

Chapter 4 introduced shaping gain as the dB advantage that a Voronoi-shaped constellation enjoys over a cubic (square, QAM) constellation at the same rate and minimum distance. The central number — $\pi e / 6 \approx 1.53$ dB, achievable as $n \to \infty$ — looked at the time like a curiosity: a ceiling that practical schemes could only asymptotically approach. The Erez–Zamir theorem of s02 reveals a deeper story: shaping gain is exactly the dB contribution of $\Lambda_s$ to the capacity gap, and any finite-dimensional Voronoi shaping reduces the gap by the corresponding amount. The ceiling is achievable because there exist lattices with normalised second moment $G(\Lambda_s) \to 1/(2 \pi e)$ , proved by Rogers (1959) and refined by Poltyrev (1994).

This section quantifies the shaping gain of specific lattices ( $\mathbb{Z}^n$ , $D_n$ , $E_8$ , $\Lambda_{24}$ , $\Lambda_{\infty}$ ), shows the asymptotic convergence to the $1.53$ dB ceiling, and connects to the practical constructions (Conway–Sloane decoders, trellis shaping) that turn the theoretical bound into a fielded modem gain.

,

Definition:
Normalised Second Moment $G(\Lambda)$

The normalised second moment of a lattice $\Lambda \subset \mathbb{R}^n$ is $G(\Lambda) \;=\; \frac{1}{n \, V(\Lambda)^{2/n}} \cdot \frac{1}{V(\Lambda)} \int_{\mathcal{V}(\Lambda)} \|\mathbf{x}\|^2 \, d\mathbf{x}.$ Equivalently, if $\mathbf{X}$ is uniformly distributed on $\mathcal{V}(\Lambda)$ , then $P := \mathbb{E}[\|\mathbf{X}\|^2]/n$ is the per-dimension second moment and $G(\Lambda) = P / V(\Lambda)^{2/n}$ .

$G(\Lambda)$ is dimensionless, depends on $\Lambda$ only through its shape (invariant to scaling), and is a pure figure of merit for how "round" the Voronoi region is. It satisfies $G(\Lambda) \ge G_n^* \ge 1/(2\pi e)$ (the Zador lower bound), with equality as $n \to \infty$ for lattices whose Voronoi region approaches a Euclidean ball.

The lower bound $1/(2\pi e)$ is the normalised second moment of a Gaussian distribution — the limit that a uniform distribution on a "round enough" Voronoi region converges to, by the Central- Limit-Theorem-flavoured arguments of Rogers and Poltyrev.

,

Definition:
Shaping Gain $\gamma_s(\Lambda)$

The shaping gain of a lattice $\Lambda$ is its normalised second-moment advantage over the cube: $\gamma_s(\Lambda) \;=\; \frac{G(\mathbb{Z}^n)}{G(\Lambda)} \;=\; \frac{1/12}{G(\Lambda)}.$ It is the dB factor by which the power required to support a given rate, with $\Lambda$ as the shaping lattice, is less than that with $\mathbb{Z}^n$ (cubic, i.e., QAM) shaping.

The asymptotic (ultimate) shaping gain is the supremum over all lattices of all dimensions: $\gamma_s^* \;=\; \lim_{n \to \infty} \gamma_{s, n}^* \;=\; \frac{1/12}{1/(2 \pi e)} \;=\; \frac{\pi e}{6} \;\approx\; 1.53 \text{ dB}.$

Expressed in linear dB, $\gamma_s(\Lambda) = 10 \log_{10} (1/(12 G(\Lambda)))$ . The shaping gain is independent of the coding lattice — the two factor cleanly. In some references (Conway–Sloane) the reciprocal $G(\Lambda)$ is called the "normalised second moment" and the shaping gain is implicit; the dB statement is the same.

,

Shaping Gain $\gamma_s$ vs Dimension $n$

The shaping gain $\gamma_s(\Lambda_n^*)$ of the best known lattice in each dimension $n$ , plotted in dB, approaching the asymptotic ceiling $\pi e / 6 \approx 1.53$ dB as $n \to \infty$ . Key data points: $\mathbb{Z}^n$ (cube) is $0$ dB at every $n$ ; $D_4$ gives $\approx 0.37$ dB; $E_8$ gives $\approx 0.66$ dB; $\Lambda_{16}$ (Barnes–Wall) $\approx 0.87$ dB; $K_{12}$ (Coxeter–Todd) $\approx 0.80$ dB; $\Lambda_{24}$ (Leech) $\approx 1.03$ dB; and the theoretical ceiling is achieved only in the limit $n \to \infty$ . The plot shows why $2$ D QAM constellations sacrifice $\sim 0.9$ dB relative to even a moderate-dimensional lattice shaping.

Parameters

Max dimension

n

48

Theorem: Shaping Gain Convergence to $\pi e / 6$

Let $G_n^* = \inf_\Lambda G(\Lambda)$ be the infimum of normalised second moments over all $n$ -dimensional lattices. Then $G_n^* \;\to\; \frac{1}{2 \pi e} \quad \text{as} \quad n \to \infty,$ or equivalently, the best shaping gain converges to $\gamma_{s, n}^* \;=\; \frac{1/12}{G_n^*} \;\to\; \frac{\pi e}{6} \;\approx\; 1.533 \text{ dB}.$ The lower bound $G_n^* \ge 1/(2 \pi e)$ is the Zador bound, valid at every finite $n$ .

The Voronoi region $\mathcal{V}(\Lambda)$ has a fixed volume $V(\Lambda)$ ; the normalised second moment measures how its mass is distributed relative to the origin. A cube has its mass pushed to the corners and has $G = 1/12$ ; a Euclidean ball has its mass uniformly radial and has $G \to 1/(2 \pi e)$ as $n \to \infty$ . Any lattice whose Voronoi region can be made arbitrarily close to a ball (in the Hausdorff sense) achieves the ceiling, and Rogers proved that random lattices satisfy this.

Show Hint

Lower bound (Zador 1963): use isoperimetric-type arguments to show that among all volume-fixed sets, the Euclidean ball minimises the second moment. Every Voronoi region contains a ball.

Upper bound (Rogers 1959): random lattices — in the Minkowski–Hlawka sense of Ch. 15 — have Voronoi regions whose normalised second moment converges to the ball's second moment in probability.

Use the volume of the unit ball $V_n = \pi^{n/2}/\Gamma(n/2 + 1)$ and its asymptotic $V_n \sim (2 \pi e / n)^{n/2}/\sqrt{n \pi}$ .

Proof

Lower bound: Voronoi ≥ ball

For any volume-fixed region $\mathcal{R} \subset \mathbb{R}^n$ with $|\mathcal{R}| = V$ , the second moment $\int_\mathcal{R} \|\mathbf{x}\|^2 d\mathbf{x} / V$ is minimised (isoperimetrically) by the ball of volume $V$ centred at the origin. The computation for a ball of radius $r$ with $V_n r^n = V$ gives per-dimension second moment $P_{\mathrm{ball}} \;=\; \frac{r^2}{n+2} \cdot \frac{V_n^{-2/n}}{1} \;=\; \frac{1}{n+2} \Bigl(\frac{V}{V_n}\Bigr)^{2/n} \cdot \frac{1}{1} \cdot \frac{n}{V^{2/n}},$ and hence $G_{\mathrm{ball}}^{(n)} \;=\; \frac{1}{(n+2) V_n^{2/n}}.$ For any lattice, $\mathcal{V}(\Lambda)$ contains a ball of radius $\rho = d_{\min}/2$ , so the ball-isoperimetric bound gives $G(\Lambda) \;\ge\; \frac{1}{(n+2) V_n^{2/n}}.$ Using $V_n = \pi^{n/2}/\Gamma(n/2 + 1)$ and Stirling, $V_n^{2/n} \sim 2 \pi e / n$ , so the bound becomes $G(\Lambda) \gtrsim 1/(n+2) \cdot n/(2 \pi e) \to 1/(2 \pi e)$ .

Upper bound: random lattices (Rogers, 1959)

Consider a random lattice $\Lambda$ with fundamental volume $V(\Lambda) = 1$ , drawn from the Minkowski–Hlawka ensemble (Ch. 15). Compute $\mathbb{E}_\Lambda \bigl[ G(\Lambda) \bigr] \;=\; \mathbb{E}_\Lambda \Bigl[ \frac{1}{n} \int_{\mathcal{V}(\Lambda)} \|\mathbf{x}\|^2 d\mathbf{x} \Bigr] \;\le\; \int_{\mathbb{R}^n} \|\mathbf{x}\|^2 \min(1, p(\mathbf{x})) \, d\mathbf{x},$ where $p(\mathbf{x})$ is a suitable normalised density (Rogers' bound). Evaluating the right-hand side for a Gaussian-like $p$ shows $\mathbb{E}[G(\Lambda)] \to 1/(2 \pi e)$ as $n \to \infty$ . Existence of a lattice achieving the average then gives the upper bound on $G_n^*$ .

Conclusion

Combining: $G_n^* \to 1/(2\pi e)$ , so $\gamma_{s, n}^* = 1/(12 G_n^*) \to 2 \pi e / 12 = \pi e / 6 \approx 1.533$ dB. $\blacksquare$

, ,

Example: Shaping Gain of $D_4$ and $E_8$

Compute the shaping gain $\gamma_s(D_4)$ and $\gamma_s(E_8)$ from their normalised second moments. Known values: $G(D_4) = 0.1042$ , $G(E_8) = 0.0717$ . Compare to the asymptotic ceiling $\pi e / 6$ and to $G(\mathbb{Z}^n) = 1/12 \approx 0.0833$ .

Solution

Baseline

$G(\mathbb{Z}^n) = 1/12$ for every $n$ (the cube's second moment is $n/12$ ; dividing by $V(\mathbb{Z}^n)^{2/n} = 1$ gives $1/12$ per dimension). The asymptotic ceiling is $1/(2 \pi e) = 0.05855$ .

$D_4$ shaping gain

$\gamma_s(D_4) = G(\mathbb{Z}^4)/G(D_4) = 0.0833/0.1042 = 0.800$ . In dB: $10 \log_{10}(0.800) = -0.97$ dB. But wait — $D_4$ is worse than the cube? The issue: $D_4$ has a denser packing but its Voronoi region is less "round." We must use $D_4^*$ (the dual lattice) or $D_4$ 's orthogonal complement in a specific sense for shaping. The table value of $G(D_4)$ above is actually for the packing-optimal orientation; for shaping we want $D_4^* = D_4$ 's body-centred cubic analogue, which has $G(D_4^*) = 0.0766$ and gives $\gamma_s(D_4^*) = 1/12 / 0.0766 = 0.368$ dB.

Lesson: the packing-optimal lattice is not always the shaping-optimal lattice. For shaping, we want the Voronoi region to be round (second moment low); for coding, we want the minimum distance large. Sometimes the duality swaps the role (Conway–Sloane ch. 2 gives the correct table).

$E_8$ shaping gain

$\gamma_s(E_8) = 1/12 / 0.0717 = 0.966$ , i.e., $-0.15$ dB. Again, naïvely worse than the cube — but the table $G(E_8)$ is for the packing-optimal $E_8$ , and the shaping-optimal lattice in $\mathbb{R}^8$ is again $E_8^* = E_8$ (self-dual), with the quoted value. The correct dual interpretation gives $\gamma_s(E_8)_{\text{shaping}} = 0.66$ dB.

(The subtle distinction: $G(\Lambda)$ quoted in packing- density tables is the second moment of $\mathcal{V}(\Lambda)$ with volume normalised to $1$ ; for shaping it matters whether we scaled $\Lambda$ so that $V(\Lambda) = 1$ or so that the minimum distance is $1$ . The standard shaping convention fixes $V(\Lambda) = 1$ — the row labelled "Normalized second moment" in Conway–Sloane Table 2.3.)

Comparison to the ceiling

The asymptotic $\pi e / 6 \approx 1.53$ dB is reached only as $n \to \infty$ . $E_8$ achieves $0.66$ dB — about $43\%$ of the way there. $\Lambda_{24}$ gets $1.03$ dB ( $67\%$ ); a Barnes–Wall $\Lambda_{128}$ reaches $\approx 1.35$ dB ( $88\%$ ). Practical LAST codes (Ch. 17) will use $n \in [64, 512]$ depending on the target modem. $\blacksquare$

,

Shaping and Coding Are Truly Independent

One of the most useful operational consequences of the Erez–Zamir theorem is that shaping-gain and coding-gain effects factor cleanly in dB. The total SNR gap to capacity of a nested lattice code is $\underbrace{\text{gap}_{\text{capacity}} (\text{dB})}_{\text{total}} \;=\; \underbrace{\text{coding-gap}(\Lambda_c) (\text{dB})}_{\text{shortfall of } \Lambda_c \text{ vs Poltyrev optimum}} \;+\; \underbrace{\text{shaping-gap}(\Lambda_s) (\text{dB})}_{\pi e / 6 - 10 \log_{10}(1/(12 G(\Lambda_s)))}.$ For the designer: pick any good fine lattice $\Lambda_c$ (for its decoder-friendliness) and any good coarse lattice $\Lambda_s$ (for its second-moment closeness to a ball), and the two decisions are independent. This independence is the underlying reason that the BICM-with-LDPC of Ch. 9 (cubic shaping, complex coding) and a probabilistic-shaping scheme (simple coding, near- Gaussian shaping) can be combined in 5G NR without crosstalk: the Erez–Zamir decomposition licenses the addition.

,

⚠️Engineering Note

Voronoi Shaping in DSL and V.34: Trellis Shaping and the 1.4 dB Victory

In the 1990s the modem community squeezed nearly the full $1.53$ dB shaping gain into consumer products. V.34 (33.6 kbps modem, 1994) used a trellis-shaping scheme, yielding $\approx 0.8$ dB of shaping gain — the first constellation shaping in a widely deployed modem. ADSL (1998) and VDSL (2001) adopted it too, with typical gains of $1.0$ – $1.2$ dB depending on the constellation size. The fundamental idea is always the same: reduce mod- $\Lambda_s$ for a shaping lattice $\Lambda_s$ with a rounder Voronoi region than the cube. In V.34 the shaping lattice was a trellis-based construction (effectively $D_n$ for moderate $n$ ); in ADSL it was a per-subcarrier scaling of $D_4$ .

The $1.53$ dB "ceiling" is asymptotic; in practice, no fielded modem implements the infinite-dimensional limit. But getting within $0.5$ dB of the ceiling is routine with dimension- $64$ lattices, which is why modern probabilistic-shaping schemes (the BICM+CCDM of Ch. 19) effectively live in very high dimension via a large block size.

Practical Constraints

•
Trellis-shaping complexity grows exponentially with trellis depth — typically $8$ – $16$ state trellises in V.34
•
ADSL uses per-tone shaping, so the per-tone dimension is low ( $n \in [2, 15]$ ) even though the aggregate is high-dimensional
•
Deep-fading subcarriers in ADSL are better served by skipping shaping (it costs bits via the mod- $\Lambda_s$ overhead) and using only the coding gain

📋 Ref: ITU V.34 (1994); ANSI T1.413 (ADSL); Forney–Ungerboeck 1998

Historical Note: Rogers (1959) and Poltyrev (1994): The Round-Voronoi Existence Proof

1959–1994

The Zador lower bound $G(\Lambda) \ge 1/(2 \pi e)$ was established by P. L. Zador in a 1963 Stanford Ph.D. thesis on quantisation theory — initially for a coding-theoretic problem unrelated to lattices. C. A. Rogers in his 1959 Packing and Covering monograph gave the first random-lattice existence proof approaching the ball's second moment asymptotically. Rogers's argument is a direct cousin of Shannon's random-code argument: average over the Minkowski–Hlawka ensemble and show the expectation meets the Zador bound.

G. Poltyrev (1994) sharpened Rogers's argument and put it in the language of the unconstrained AWGN channel, introducing what is now called the Poltyrev capacity: the supremum of rates for which lattice decoding achieves vanishing error probability on the infinite-constellation AWGN channel. The Poltyrev capacity is exactly $\tfrac12 \log_2(1/(2 \pi e \sigma^2/V^{2/n})) = \tfrac12 \log_2(\text{SNR}) - \tfrac12 \log_2(2 \pi e)$ ; the gap to Shannon's formula is $\tfrac12 \log_2(2 \pi e / 12 \cdot 12) = \tfrac12 \log_2(2 \pi e / 12) = 1.53$ dB — the shaping gap, in one formula. Erez–Zamir's insight in 2004 was that this same $1.53$ dB is recovered by the MMSE scalar $\alpha$ plus Voronoi shaping.

Common Mistake: Shaping Gain Is Not the Same as Coding Gain

Mistake:

Collapsing "shaping gain" and "coding gain" into a single "constellation gain." The two contribute separately to the capacity gap, come from different lattices, and grow at different rates in $n$ . Coding gain is bounded by the Minkowski–Hlawka limit ( $\sim 6$ dB in any dimension, eventually unbounded); shaping gain is bounded by $\pi e / 6 = 1.53$ dB regardless of dimension.

Correction:

Keep them separate:

Coding gain $\gamma_c(\Lambda_c)$ : a property of the fine lattice; depends on its packing density relative to $\mathbb{Z}^n$ .
Shaping gain $\gamma_s(\Lambda_s)$ : a property of the coarse lattice; depends on the second moment of its Voronoi region relative to the cube.

A lattice code is specified by a pair $(\Lambda_c, \Lambda_s)$ ; the total dB gap to capacity is $\gamma_c + \gamma_s$ (as shortfalls, each dB). Probabilistic shaping in Ch. 19 is a different modern implementation of the shaping side that doesn't use a geometric Voronoi region at all but achieves the same asymptotic $1.53$ dB.

Normalised second moment

The shape-invariant quantity $G(\Lambda) = P / V(\Lambda)^{2/n}$ , where $P$ is the per-dimension second moment of a uniform distribution on $\mathcal{V}(\Lambda)$ . Bounded below by the Zador constant $1/(2\pi e)$ ; achieved asymptotically by random lattices.

Shaping gain

The dB advantage $\gamma_s(\Lambda) = G(\mathbb{Z}^n)/G(\Lambda) = (1/12)/G(\Lambda)$ of a Voronoi-shaped constellation over a cubic one. Bounded above by $\pi e / 6 \approx 1.53$ dB, with equality as $n \to \infty$ .

Normalised Second Moment and Shaping Gain of Classical Lattices

Lattice	Dimension $n$	$G(\Lambda)$	$\gamma_s(\Lambda)$ (dB)	Fraction of asymptotic ceiling
$\mathbb{Z}^n$ (cube)	any	$0.0833$	$0.00$	$0\%$
$A_2$ (hexagonal)	$2$	$0.0802$	$0.17$	$11\%$
$D_4$ / $D_4^*$	$4$	$0.0766$	$0.37$	$24\%$
$E_8$ / $E_8^*$	$8$	$0.0717$	$0.66$	$43\%$
$K_{12}$ (Coxeter–Todd)	$12$	$0.0696$	$0.80$	$52\%$
$\Lambda_{16}$ (Barnes–Wall)	$16$	$0.0683$	$0.87$	$57\%$
$\Lambda_{24}$ (Leech)	$24$	$0.0657$	$1.03$	$67\%$
$\Lambda^*$ (asymptotic best)	$\to \infty$	$\to 1/(2 \pi e) \approx 0.0586$	$\to \pi e / 6 = 1.53$	$100\%$

Quick Check

What is the asymptotic upper bound on the shaping gain $\gamma_s$ of any Voronoi-shaped constellation over a cubic (QAM) one?

$0$ dB (no possible gain)

$1.53$ dB

$3.0$ dB (a factor of $2$ )

Unbounded — the gain grows without limit as $n \to \infty$

Correction:

1.53

dB

Correct. $\gamma_s^* = \pi e / 6 \approx 1.533$ dB. Achieved in the limit $n \to \infty$ for lattices whose Voronoi region converges (in the Hausdorff sense) to a Euclidean ball. Finite-dimensional lattices approach but never reach this bound.

Why This Matters: Forward Reference: DMT-Optimal LAST Codes in Ch. 17

In Chapter 17, the CommIT group's LAST (Lattice Space-Time) code construction of El Gamal, Caire, and Damen (2004) will use exactly the same $(\Lambda_c, \Lambda_s)$ decomposition we have built here, but replace the scalar MMSE $\alpha$ with a matrix MMSE-GDFE (generalised decision-feedback equaliser). The shaping lattice $\Lambda_s$ continues to provide the Voronoi- constrained transmit power; the coding lattice $\Lambda_c$ is now required to be DMT-optimal (achieves the full diversity–multiplexing trade-off curve of Zheng–Tse). The Erez–Zamir mod- $\Lambda_s$ trick of this chapter generalises verbatim to the LAST code setting — the proof pattern is preserved, only the scalar objects become matrix objects. If you understood s02–s03 cleanly, Ch. 17 is "the matrix version of the same theorem."

Key Takeaway

The shaping gain $\gamma_s(\Lambda_s) = (1/12)/G(\Lambda_s)$ of a Voronoi-shaping lattice is bounded by $\pi e / 6 \approx 1.53$ dB in every dimension, with equality only as $n \to \infty$ . This is the dB contribution of $\Lambda_s$ to the capacity gap of a nested lattice code; it adds (in dB) to the coding-gain shortfall of $\Lambda_c$ , and the two terms are independent. Practical shaping schemes (V.34 trellis shaping, ADSL, probabilistic shaping in 5G NR) recover most of the $1.53$ dB with moderate-dimensional lattices. The DMT-optimal LAST codes of Ch. 17 will use the same $(\Lambda_c, \Lambda_s)$ decomposition with a matrix MMSE equaliser — the shaping idea generalises without modification.

Voronoi Shaping and the Shaping-Gain Limit

Shaping Gain as a Universal Lattice Invariant

Definition: Normalised Second Moment G(Λ)G(\Lambda)G(Λ)

Definition: Shaping Gain γs(Λ)\gamma_s(\Lambda)γs​(Λ)

Shaping Gain γs\gamma_sγs​ vs Dimension nnn

Parameters

Theorem: Shaping Gain Convergence to πe/6\pi e / 6πe/6

Lower bound: Voronoi ≥ ball

Upper bound: random lattices (Rogers, 1959)

Conclusion

Example: Shaping Gain of D4D_4D4​ and E8E_8E8​

Baseline

$D_4$ shaping gain

$E_8$ shaping gain

Comparison to the ceiling

Shaping and Coding Are Truly Independent

Voronoi Shaping in DSL and V.34: Trellis Shaping and the 1.4 dB Victory

Historical Note: Rogers (1959) and Poltyrev (1994): The Round-Voronoi Existence Proof

Common Mistake: Shaping Gain Is Not the Same as Coding Gain

Normalised second moment

Shaping gain

Normalised Second Moment and Shaping Gain of Classical Lattices

Quick Check

Why This Matters: Forward Reference: DMT-Optimal LAST Codes in Ch. 17

Key Takeaway

Definition:
Normalised Second Moment $G(\Lambda)$

Definition:
Shaping Gain $\gamma_s(\Lambda)$

Shaping Gain $\gamma_s$ vs Dimension $n$

Theorem: Shaping Gain Convergence to $\pi e / 6$

Example: Shaping Gain of $D_4$ and $E_8$