Ferkans — Interactive Telecom Tutor

The $1.53$ dB Shannon Tax

We now derive the number $\pi e / 6 \approx 1.533$ dB — the single most famous number in signal shaping. Where does it come from? Two observations, connected by one inequality:

Gaussian is optimal. The maximum-entropy distribution at fixed energy on $\mathbb{R}^n$ is the isotropic Gaussian. A finite, uniformly distributed constellation cannot match its entropy at the same energy.
Cubes are bad boundaries. A bounded-support distribution with the same entropy as a Gaussian has higher energy. For a cube the penalty is $\pi e / 6$ . For a sphere the penalty vanishes as $n \to \infty$ .

Combining these gives the claim: the best possible shaping gain (over cubic QAM) is $\pi e / 6$ , and it is achieved asymptotically by spherical constellations. This is the part of the gap to Shannon capacity that we can never close with coding alone — we have to shape the marginal distribution. But it is also upper bounded by a single universal dB number. Once you have your $1.53$ dB, you are done with shaping forever. The rest of the gap is coding.

,

Theorem: Shaping Gain Ceiling: $\gamma_s \le \pi e / 6$

For any lattice $\Lambda_s \subset \mathbb{R}^n$ , $G(\Lambda_s) \;\ge\; \frac{1}{2 \pi e},$ with equality only in the limit $n \to \infty$ for the $n$ -dimensional ball. Consequently the shaping gain satisfies $\gamma_s(\Lambda_s) \;=\; \frac{1}{12 G(\Lambda_s)} \;\le\; \frac{2 \pi e}{12} \;=\; \frac{\pi e}{6} \;\approx\; 1.4233,$ i.e.\ $\gamma_s \le 10 \log_{10}(\pi e / 6) \approx 1.533$ dB.

The proof is a one-liner combining two classical facts: the Gaussian maximum-entropy theorem (the most random-looking distribution at fixed energy is Gaussian) and the fact that the differential entropy of a bounded-support uniform distribution equals $\log V$ , where $V$ is the support volume. A uniform distribution on $\mathcal{V}(\Lambda_s)$ has the same entropy as the Gaussian if and only if their variances are related by $V^{2/n} / \sigma^2 = 2 \pi e$ , i.e.\ $G = 1/(2\pi e)$ .

Show Hint

Differential entropy of uniform on a region of volume $V$ equals $\ln V$ (in nats per dimension, dividing by $n$ ).

Differential entropy of a Gaussian with per-dimension variance $\sigma^2$ equals $\tfrac12 \ln(2 \pi e \sigma^2)$ .

Apply the Gaussian max-entropy bound.

Proof

Entropy of a uniform distribution on $\mathcal{V}(\Lambda_s)$

Let $\mathbf{X}$ be uniform on $\mathcal{V}(\Lambda_s)$ . Its differential entropy (per dimension, in nats) is $h(\mathbf{X})/n \;=\; \frac{1}{n} \ln V(\Lambda_s) \;=\; \frac{1}{n} \ln V(\Lambda_s).$ Its per-dimension energy is $\sigma^2(\Lambda_s) = G(\Lambda_s) V(\Lambda_s)^{2/n}$ .

Gaussian maximum-entropy

Among all random vectors with per-dimension variance $\sigma^2$ , the isotropic Gaussian maximises differential entropy, with value $\tfrac12 \ln(2 \pi e \sigma^2)$ per dimension. Hence $\frac{h(\mathbf{X})}{n} \;\le\; \frac{1}{2} \ln\!\bigl(2 \pi e\, \sigma^2(\Lambda_s)\bigr).$

Combine

Substituting $\sigma^2 = G V^{2/n}$ on the right and $h/n = (\ln V)/n$ on the left: $\frac{1}{n} \ln V \;\le\; \frac{1}{2} \ln\!\bigl(2 \pi e \, G V^{2/n}\bigr) \;=\; \frac{1}{2} \ln(2 \pi e G) + \frac{1}{n} \ln V.$ The $\tfrac1n \ln V$ terms cancel, leaving $0 \le \tfrac12 \ln(2 \pi e G)$ , i.e.\ $G \ge 1/(2\pi e)$ .

Equality and the shaping gain bound

Equality holds when the uniform distribution on $\mathcal{V}$ matches the Gaussian — which, by the entropy-power inequality, is possible only as $n \to \infty$ with $\mathcal{V}$ approaching a ball. Dividing into the constant $1/12$ gives $\gamma_s \le (2 \pi e)/12 = \pi e / 6 \approx 1.533$ dB. $\blacksquare$

, ,

Shaping Gain vs Dimension

The per-dimension shaping gain of the $n$ -ball (the infimum over lattices at dimension $n$ ) climbs from $0$ dB at $n = 1$ to the $\pi e / 6 \approx 1.533$ dB ceiling as $n \to \infty$ . The curve grows slowly: you need $n \sim 16$ to get the first dB, and another factor of $10$ in dimension to approach $1.4$ dB. This is why practical shaping stays in moderate dimensions (shell mapping uses $n = 16$ ) and why the last $0.1$ dB of shaping gain is almost impossibly expensive.

Parameters

Max dimension

n

64

Show

\pi e / 6

asymptote

Theorem: Normalised Second Moment of the $n$ -Ball

The $n$ -dimensional ball of volume $V$ has normalised second moment $G(B^n) \;=\; \frac{1}{(n+2)} \cdot \frac{1}{\pi} \cdot \Gamma(1 + n/2)^{2/n}.$ As $n \to \infty$ , $G(B^n) \to 1/(2 \pi e)$ by Stirling's approximation.

The ball is the "roundest" possible region at a given volume, so it achieves the smallest possible second moment per unit volume. The fact that $G(B^n)$ converges to the Gaussian limit says that a high-dimensional uniform distribution on a ball is indistinguishable from a Gaussian in its moments — the entropy gap closes.

Show Hint

Integrate $\|\mathbf{x}\|^2$ in spherical coordinates.

Use the volume $V_n = \pi^{n/2} R^n / \Gamma(1 + n/2)$ of an $R$ -radius ball.

Stirling: $\Gamma(1 + n/2) \sim (n/(2e))^{n/2}$ gives the limit.

Proof

Second moment of a ball of radius $R$

By symmetry, $\int_{B(R)} \|\mathbf{x}\|^2 \, d\mathbf{x} = n \omega_{n-1} \int_0^R r^{n+1} \, dr = \frac{n \omega_{n-1} R^{n+2}}{n+2},$ where $\omega_{n-1}$ is the surface area of the unit $(n-1)$ -sphere. Dividing by $n V_n = n \cdot \omega_{n-1} R^n / n$ gives the per-dimension second moment $\sigma^2(B(R)) = R^2 / (n+2)$ .

Normalisation

Using $V_n(R) = \pi^{n/2} R^n / \Gamma(1 + n/2)$ , we have $V_n^{2/n} = \pi R^2 / \Gamma(1 + n/2)^{2/n}$ . Therefore $G(B^n) = \frac{\sigma^2}{V^{2/n}} = \frac{R^2/(n+2)}{\pi R^2 / \Gamma(1 + n/2)^{2/n}} = \frac{\Gamma(1 + n/2)^{2/n}}{\pi (n+2)}.$

Asymptotic limit

Using Stirling's approximation $\Gamma(1 + n/2) \sim \sqrt{\pi n} (n/(2e))^{n/2}$ , one obtains $\Gamma(1+n/2)^{2/n} \sim n/(2e)$ to leading order, so $G(B^n) \sim n / (2e \pi (n+2)) \to 1/(2\pi e)$ as $n \to \infty$ . $\blacksquare$

,

Reading the Ultimate Gap

A canonical way to read Shannon's capacity formula $C = \tfrac12 \log_2(1 + \text{SNR})$ is as a prescription: if you are $\Delta$ dB below capacity at some rate $R$ , then $\Delta$ decomposes as

$\Delta = \underbrace{\gamma_{c,\rm uncoded}}_{\text{coding gap}} \;-\; \underbrace{\gamma_c}_{\text{coding gain}} \;-\; \underbrace{\gamma_s}_{\text{shaping gain}}.$

The three components are independent. An uncoded QAM constellation at moderate rate sits at $\Delta \approx 9$ dB from capacity. A good code closes about $8$ dB of that (the coding side); the last $1.53$ dB is the shaping tax, which no code — however long or cleverly designed — can close. Shaping is the necessary complement to coding, not an optional refinement.

Example: Shaping Gain of $E_8$

Compute the shaping gain $\gamma_s(E_8)$ and express it as a fraction of the ultimate ceiling $\pi e / 6$ .

Solution

Second moment of $E_8$

Conway–Sloane (Ch. 2) report $G(E_8) \approx 0.07167$ , based on exact integration over the $E_8$ Voronoi cell.

Shaping gain

$\gamma_s(E_8) = (1/12) / 0.07167 \approx 1.1628$ , i.e.
$10 \log_{10}(1.1628) \approx 0.655$ dB.

Fraction of the ceiling

The ceiling is $\pi e / 6 \approx 1.4233$ (ratio), i.e.\ $1.533$ dB. $E_8$ gets to $0.655 / 1.533 \approx 43\%$ of the available shaping gain in just $8$ dimensions. For comparison, the Leech lattice $\Lambda_{24}$ gets to $\approx 67\%$ , and $\mathbb{Z}^n$ (standard QAM) gets to $0\%$ . $\blacksquare$

Common Mistake: Shaping and Coding Contributions Are Orthogonal, Not Interchangeable

Mistake:

Claiming that a stronger code can compensate for missing shaping, or that a better shaping strategy reduces the need for coding.

Correction:

The total gap to Shannon capacity equals (in dB) the sum of the coding-gap and shaping-gap terms. A code improves the coding term but leaves the shaping term untouched; a shaping scheme improves the shaping term but leaves the coding term untouched. The two cannot substitute for each other. Even a capacity-achieving binary code on a uniform QAM constellation is $\pi e / 6 \approx 1.53$ dB short of Shannon — the shaping tax is unavoidable without shaping.

Common Mistake: The $\pi e / 6$ Ceiling Is Asymptotic

Mistake:

Assuming a practical shaping scheme at $n = 8$ or $n = 16$ dimensions can recover the full $1.53$ dB.

Correction:

The ceiling $\gamma_s = \pi e / 6$ is attained only in the limit $n \to \infty$ . At $n = 8$ (Gosset lattice $E_8$ ) the best shaping gain is $\approx 0.65$ dB; at $n = 24$ (Leech lattice $\Lambda_{24}$ ) it is $\approx 1.03$ dB. The last half-decibel of shaping gain is prohibitively expensive in complexity, which is why all practical systems accept $\gamma_s \approx 1$ dB and move on.

Why This Matters: Probabilistic Shaping: Recovering $1$ dB with a Different Mechanism

Modern standards (DVB-S2X, 5G, optical coherent) recover the same shaping gain through a different mechanism: probabilistic shaping. Instead of confining lattice points to a bounded region, the constellation is left fixed and the input distribution is made non-uniform — inner points appear more frequently than outer points. Böcherer, Steiner, and Schulte (2015) showed that this is equivalent to spherical shaping in the high-rate regime. Chapter 19 will return to this in detail, with the probabilistic-amplitude-shaping (PAS) architecture of modern coherent optical links as the main example. The connection to this chapter is direct: the information-theoretic ceiling $\pi e / 6$ applies to probabilistic shaping as well, and from the same max-entropy argument.

Quick Check

Why is $\pi e / 6$ a universal upper bound on shaping gain?

Because the Gaussian is the maximum-entropy distribution at fixed variance, and the $n$ -ball achieves this limit as $n \to \infty$ .

Because the AWGN channel capacity is $\tfrac12 \log_2(1 + \text{SNR})$ .

Because $\pi e$ is the Euler number times $\pi$ .

Because the $n$ -ball has volume $\pi^{n/2} / \Gamma(n/2 + 1)$ .

Correction:

Because the Gaussian is the maximum-entropy distribution at fixed variance, and the

n

-ball achieves this limit as

n \to \infty

.

The shaping gain is proportional to the ratio of "support volume" to "energy". The Gaussian distribution — which maximises entropy at fixed variance — provides the tightest such ratio in infinite dimensions. Bounded-support distributions (like uniform on a lattice's Voronoi region) cannot exceed this ratio, and the $n$ -ball approaches it asymptotically. The numerical value $\pi e / 6$ is the dB-ratio between the Gaussian's optimum and the cube's $1/12$ .

Key Takeaway

The shaping gain ceiling is $\pi e / 6 \approx 1.53$ dB. This number comes from a single inequality: the differential entropy of a uniform distribution on a bounded region is at most the differential entropy of a Gaussian at the same variance. The ceiling is attained in the limit $n \to \infty$ by the $n$ -ball; all finite-dimensional constellations fall short. In practice, $1$ dB is readily attainable at $n \sim 16$ ; the last half-decibel costs enormous complexity. The next section surveys the two practical shaping schemes — shell mapping and trellis shaping — that attempt to approach the ceiling.

The Shaping Gain and the πe/6\pi e / 6πe/6 Ceiling

The 1.531.531.53 dB Shannon Tax

Theorem: Shaping Gain Ceiling: γs≤πe/6\gamma_s \le \pi e / 6γs​≤πe/6

Entropy of a uniform distribution on $\mathcal{V}(\Lambda_s)$

Gaussian maximum-entropy

Combine

Equality and the shaping gain bound

Shaping Gain vs Dimension

Parameters

Theorem: Normalised Second Moment of the nnn-Ball

Second moment of a ball of radius $R$

Normalisation

Asymptotic limit

Reading the Ultimate Gap

Example: Shaping Gain of E8E_8E8​

Second moment of $E_8$

Shaping gain

Fraction of the ceiling

Common Mistake: Shaping and Coding Contributions Are Orthogonal, Not Interchangeable

Common Mistake: The πe/6\pi e / 6πe/6 Ceiling Is Asymptotic

Why This Matters: Probabilistic Shaping: Recovering 111 dB with a Different Mechanism

Quick Check

Key Takeaway

The Shaping Gain and the $\pi e / 6$ Ceiling

The $1.53$ dB Shannon Tax

Theorem: Shaping Gain Ceiling: $\gamma_s \le \pi e / 6$

Theorem: Normalised Second Moment of the $n$ -Ball

Example: Shaping Gain of $E_8$

Common Mistake: The $\pi e / 6$ Ceiling Is Asymptotic

Why This Matters: Probabilistic Shaping: Recovering $1$ dB with a Different Mechanism