The Spectral-Efficiency Plane

A Map of All Communication Systems

The point is that every bandwidth-limited coded-modulation scheme — BICM in LTE, turbo codes in 3G, orthogonal signaling for deep-space links, modem standards like V.34 — sits at a single point on a two-dimensional plane whose axes are spectral efficiency η\eta and required Eb/N0E_b/N_0. This plane is bounded above by the Shannon limit, and every practical scheme is at some distance below it. The distance is what we call the gap to capacity, and the structure of this plane organizes more or less every design decision in digital communications.

Now here is the key observation: the Shannon curve partitions the plane into "achievable" and "not achievable" regions, but it is agnostic about what the scheme looks like. A point (η,Eb/N0)(\eta, E_b/N_0) could be realized by countless schemes — what matters is only whether it lies below the curve.

,

Definition:

The Shannon Limit on the Spectral-Efficiency Plane

Consider the real AWGN channel at spectral efficiency η\eta bits per 2D dimension. Writing SNR=Es/N0\text{SNR} = E_s/N_0 and using Es=η⋅EbE_s = \eta \cdot E_b, the capacity formula η≤log⁡2(1+SNR)\eta \le \log_2(1 + \text{SNR}) inverts to

(EbN0)min⁡(η)  =  2η−1η.\left(\frac{E_b}{N_0}\right)_{\min}(\eta) \;=\; \frac{2^{\eta} - 1}{\eta}.

This is the Shannon limit curve on the spectral-efficiency plane. A coded modulation scheme at efficiency η\eta requires an Eb/N0E_b/N_0 strictly greater than this value to achieve an arbitrarily small error probability.

Two asymptotic regimes:

  1. Power-limited. As η→0\eta \to 0, the limit approaches (2η−1)/η→ln⁡2≈−1.59(2^{\eta} - 1)/\eta \to \ln 2 \approx -1.59 dB.
  2. Bandwidth-limited. As η→∞\eta \to \infty, the limit scales as 2η/η2^{\eta}/\eta, i.e., 3 dB of additional Eb/N0E_b/N_0 per extra bit/2D for large η\eta.
,

The Spectral-Efficiency Plane with Shannon Bound

The yellow curve marks the Shannon limit Eb/N0=(2η−1)/ηE_b/N_0 = (2^{\eta} - 1)/\eta. Choose a target rate η\eta and an operating Eb/N0E_b/N_0, and the marker shows where the scheme sits. Points to the left of the curve are impossible for any scheme of that rate; points to the right are achievable but leave gap to be closed by better coding. Toggle the 1 dB margin band to see the region where modern LDPC/turbo systems typically operate.

Parameters
2

Target spectral efficiency

6

Operating SNR

Theorem: The −1.59-1.59 dB Limit

For any code achieving arbitrarily small error probability on the AWGN channel at spectral efficiency η>0\eta > 0,

EbN0  >  2η−1η.\frac{E_b}{N_0} \;>\; \frac{2^{\eta} - 1}{\eta}.

Taking η→0+\eta \to 0^+, the right-hand side approaches ln⁡2\ln 2, which in decibels is 10log⁡10(ln⁡2)≈−1.5910 \log_{10}(\ln 2) \approx -1.59 dB. Hence no reliable communication is possible below −1.59-1.59 dB of Eb/N0E_b/N_0, at any spectral efficiency.

Think of this as saying that each bit of information requires at least ln⁡2⋅N0\ln 2 \cdot N_0 joules of signal energy, no matter how much bandwidth you are willing to spend. This is the "ultimate currency exchange rate" between energy and information on a Gaussian channel, and it is why orthogonal signaling with repetition codes was the first coding scheme to come within 1-2 dB of capacity — it operates near η≈0\eta \approx 0.

Show Hint

Start from the capacity formula C=log⁡2(1+SNR)C = \log_2(1 + \text{SNR}) and substitute SNR=η⋅Eb/N0\text{SNR} = \eta \cdot E_b/N_0.

Invert to solve for Eb/N0E_b/N_0 as a function of η\eta.

Take the limit η→0\eta \to 0 using ln⁡(1+x)≈x\ln(1+x) \approx x (or equivalently L'Hôpital's rule).

Proof
Start from the AWGN capacity formula

The capacity per channel use is C=log⁡2(1+SNR)C = \log_2(1 + \text{SNR}) where SNR=Es/N0\text{SNR} = E_s/N_0. For reliable communication at rate η\eta bits per use we need η<C\eta < C, equivalently SNR>2η−1\text{SNR} > 2^{\eta} - 1.

Convert to $E_b/\ntn{n0}$

Using Es=η⋅EbE_s = \eta \cdot E_b, the condition SNR>2η−1\text{SNR} > 2^{\eta} - 1 becomes

EbN0  >  2η−1η.\frac{E_b}{N_0} \;>\; \frac{2^{\eta} - 1}{\eta}.

Take the low-$\eta$ limit

As η→0+\eta \to 0^+, expand 2η−1=ηln⁡2+12(ηln⁡2)2+⋯2^{\eta} - 1 = \eta \ln 2 + \tfrac{1}{2}(\eta \ln 2)^2 + \cdots. Dividing by η\eta gives

lim⁡η→0+2η−1η  =  ln⁡2.\lim_{\eta \to 0^+} \frac{2^{\eta} - 1}{\eta} \;=\; \ln 2.

In dB: 10log⁡10(ln⁡2)≈−1.59210 \log_{10}(\ln 2) \approx -1.592 dB. For η>0\eta > 0, the limit is strictly larger than ln⁡2\ln 2, and grows without bound as η→∞\eta \to \infty. ■\blacksquare

,

Example: Locating Classic Schemes on the Plane

Compute the required Shannon Eb/N0E_b/N_0 for an AWGN channel at each of the following spectral efficiencies, and compare with a common target for modern wireless: Pb=10−5P_b = 10^{-5} uncoded, or modestly coded.

(a) η=1\eta = 1 bit/2D (e.g., uncoded BPSK benchmark). (b) η=2\eta = 2 bits/2D (uncoded QPSK, or rate-1/21/2 code on 16-QAM). (c) η=4\eta = 4 bits/2D (uncoded 16-QAM, or rate-1/21/2 code on 256-QAM). (d) η=6\eta = 6 bits/2D (uncoded 64-QAM).

Solution
Evaluate the Shannon limit at each $\eta$

Using (Eb/N0)min⁡(η)=(2η−1)/η(E_b/N_0)_{\min}(\eta) = (2^\eta - 1)/\eta:

η\eta (bits/2D) (2η−1)/η(2^\eta - 1)/\eta dB
1 1 0.00 dB
2 1.5 1.76 dB
4 3.75 5.74 dB
6 10.5 10.21 dB
Compare to uncoded performance

Uncoded BPSK needs ≈9.6\approx 9.6 dB of Eb/N0E_b/N_0 for Pb=10−5P_b = 10^{-5}, uncoded QPSK the same, uncoded 16-QAM ≈13.4\approx 13.4 dB, and uncoded 64-QAM ≈17.7\approx 17.7 dB. The gap to capacity is therefore roughly 9.6,11.6,7.7,7.59.6, 11.6, 7.7, 7.5 dB respectively — this is the budget that coding and shaping will have to recover. Notice that the gap is largest around QPSK and smallest at very high-rate constellations.

Read off the operational takeaway

In the bandwidth-limited regime (η≥4\eta \ge 4), uncoded QAM sits at 7-8 dB above capacity. A strong LDPC plus a well-chosen QAM order can close this gap to about 1 dB; the remaining 1 dB requires shaping. In the power-limited regime (η≤1\eta \le 1), the uncoded gap is 10+ dB and is reduced by powerful binary codes that expand bandwidth — a qualitatively different design.

Locations of Historical Coding Milestones on the Spectral-Efficiency Plane

Locations of Historical Coding Milestones on the Spectral-Efficiency Plane
Illustrative placement (not to scale) of Voyager-era concatenated codes (low η\eta, ≈2.5\approx 2.5 dB from capacity), V.34 modem TCM (η=8.4\eta = 8.4 bits/s/Hz, ≈4\approx 4 dB from capacity), turbo and LDPC codes approaching the Shannon limit to within 1 dB across a wide range of η\eta, and modern 5G NR and Wi-Fi 7 LDPC+QAM+shaping at η\eta up to ≈10\approx 10 bits/s/Hz, within 1-2 dB of capacity.

Historical Note: Shannon's 1948 Paper and the Curve That Changed Communications

1948

Claude Shannon's A Mathematical Theory of Communication (1948) established the capacity formula C=Wlog⁡2(1+SNR)C = W \log_2(1 + \text{SNR}) for the bandlimited AWGN channel as part of a broader program that also introduced entropy, mutual information, and the notion of a channel code. For two decades after its publication, coding theorists worked primarily in the power-limited regime — the regime in which Shannon's limit is most dramatic, because the gap between uncoded and capacity is more than 10 dB. It took until Ungerboeck's 1982 paper on TCM before the bandwidth-limited regime received a coding-theoretic treatment at the same depth, and until the 1993 invention of turbo codes before within-1-dB-of-capacity became routine. The spectral-efficiency plane as a diagnostic tool crystallized in the Forney-Ungerboeck 1998 review, which remains the canonical reference.

Common Mistake: Capacity per channel use vs. capacity per second

Mistake:

Confusing CC in bits per channel use with capacity in bits per second, or conflating the two notions of "spectral efficiency" (bits/2D vs. bits/s/Hz).

Correction:

The expressions C=log⁡2(1+SNR)C = \log_2(1 + \text{SNR}) bits per channel use and C=Wlog⁡2(1+SNR)C = W \log_2(1 + \text{SNR}) bits/s describe the same channel; the second is just the first multiplied by the symbol rate. For a system using two real dimensions per complex symbol at one symbol per 1/W1/W Hz, bits/2D equals bits/s/Hz numerically. For real-baseband systems or systems with fractional pulse-shaping rolloff, be careful: the conversion between "per dimension" and "per hertz" depends on how the spectrum is counted.

⚠️Engineering Note

How Standards Populate the Spectral-Efficiency Plane

An LTE/NR MCS table is a discrete approximation of a small region of the spectral-efficiency plane. Each MCS index selects a constellation and an LDPC code rate, yielding an operating (η,Eb/N0req)(\eta, E_b/{N_0}_{\rm req}) pair. Good table design produces approximately equal 0.5-1 dB vertical spacing between adjacent MCS entries so that link adaptation can track channel variation in finite steps without wasting too much of the fast fading budget. The table also encodes a latency-vs-reliability tradeoff: lower MCS entries carry extra redundancy for HARQ retransmissions, while higher entries are thinner codes used when the channel is known to be good.

Practical Constraints
  • •

    5G NR Rel-15 Table 5.1.3.1-1 covers MCS indices 0-27 spanning η≈0.23\eta \approx 0.23 to 5.555.55 bits/s/Hz

  • •

    Each MCS corresponds to a spectral-efficiency-plane point; adaptive MCS selection is a quantized linkage between channel estimate and this plane

  • •

    Outer HARQ and CQI reporting control the vertical granularity of the linkage

📋 Ref: 3GPP TS 38.214, Table 5.1.3.1-1

Key Takeaway

The plane (spectral efficiency, Eb/N0E_b/N_0) is the dashboard of coded modulation. The Shannon curve cuts it in two; the target is to sit as close as possible to the curve at your chosen η\eta. Power-limited systems live on the flat part near −1.59-1.59 dB; bandwidth-limited systems live on the steep part where each extra bit of η\eta costs 3 dB. Coding gain measures how much of the 7-10 dB uncoded gap you can recover.

Shannon limit

The minimum Eb/N0E_b/N_0 that any scheme transmitting reliably at spectral efficiency η\eta bits per 2D dimension must satisfy: (Eb/N0)min⁡=(2η−1)/η(E_b/N_0)_{\min} = (2^\eta - 1)/\eta. Tends to ln⁡2≈−1.59\ln 2 \approx -1.59 dB as η→0\eta \to 0 and grows as 2η/η2^\eta/\eta for large η\eta.

Related: Capacity Gap, Coding Gain, Spectral Efficiency

Quick Check

A turbo code operating at η=1\eta = 1 bit/2D achieves Pb=10−5P_b = 10^{-5} at Eb/N0=1.0E_b/N_0 = 1.0 dB. How far is this from the Shannon limit at the same η\eta?

about −0.6-0.6 dB, i.e., below capacity

about 1.01.0 dB from capacity

about 2.52.5 dB from capacity

about 55 dB from capacity

Correction:
about 1.01.0 dB from capacity

At η=1\eta = 1, (2η−1)/η=1(2^\eta - 1)/\eta = 1, which is 00 dB. Operating at 1.01.0 dB of Eb/N0E_b/N_0 puts the scheme exactly 1 dB above the Shannon bound — a textbook near-capacity result.

Coding Gain vs. Bandwidth EfficiencyThe Gap to Capacity: Coding and Shaping