Ferkans — Interactive Telecom Tutor

Bandwidth vs. Power: The Fundamental Engineering Tradeoff

A system designer has two knobs: bandwidth $W$ and transmit power $P$ . Bandwidth is regulated and expensive (spectrum auctions); power is limited by batteries, heat, and interference. The Gaussian channel capacity formula $C = W\log_2(1 + P/(N_0W))$ reveals how these two resources trade off. Understanding this tradeoff is essential for link budget design in every wireless standard from Wi-Fi to deep-space communication.

Definition:
Spectral Efficiency

The spectral efficiency of a channel is the capacity per unit bandwidth:

$\eta = \frac{C}{W} = \log_2\!\left(1 + \frac{P}{N_0W}\right) \quad \text{bits/s/Hz}.$

Notice that as $W \to \infty$ , the noise power $N = N_0W$ grows while the signal power $P$ remains fixed, so $\text{SNR} = P/(N_0W) \to 0$ . In this regime, $\log_2(1 + \text{SNR}) \approx \text{SNR}/\ln 2$ , and

$C \to \frac{P}{N_0\ln 2} \quad \text{bits/s},$

a finite limit independent of bandwidth.

Spectral efficiency

The ratio $\eta = C/W$ measured in bits/s/Hz. It quantifies how efficiently a system uses its allocated bandwidth.

Definition:
Energy per Bit $E_b/N_0$

The energy per information bit is $E_b = P/R_{b}$ , where $R_{b}$ is the data rate in bits/s. The ratio

$\frac{E_b}{N_0} = \frac{P}{R_{b} \cdot N_0} = \frac{\text{SNR}}{\eta}$

normalizes the SNR by the spectral efficiency $\eta = R_{b}/W$ . This is the natural metric for comparing systems operating at different spectral efficiencies.

Energy per bit to noise ratio ( $E_b/N_0$ )

The ratio of energy per information bit to noise spectral density. $E_b/N_0 = \text{SNR}/\eta$ where $\eta$ is the spectral efficiency. The Shannon limit is $E_b/N_0 \geq \ln 2 \approx -1.59$ dB.

Related: Signal-to-noise ratio (SNR)

Theorem: The Shannon Limit

For reliable communication over the AWGN channel, the minimum required energy per bit satisfies

$\frac{E_b}{N_0} \geq \ln 2 \approx -1.59 \text{ dB}.$

This bound is achieved in the limit of zero spectral efficiency ( $\eta \to 0$ , i.e., the wideband regime).

Even with infinite bandwidth, you cannot communicate reliably with less than $\ln 2$ joules of energy per nat of information (or equivalently $(\ln 2)/\ln 2 = 1$ energy unit per bit, normalized appropriately). This is the ultimate energy efficiency limit — it says that information has a minimum physical cost.

Proof

Express $E_b/\ntn{n0}$ in terms of $\eta$

At capacity, $\eta = \log_2(1 + \eta \cdot E_b/N_0)$ . Solving: $E_b/N_0 = (2^\eta - 1)/\eta$ .

Take the limit $\eta \to 0$

Using L'H^opital's rule or the Taylor expansion $2^\eta = 1 + \eta \ln 2 + O(\eta^2)$ :

$\lim_{\eta \to 0} \frac{2^\eta - 1}{\eta} = \ln 2.$

In dB: $10\log_{10}(\ln 2) = 10 \times (-0.1665) \approx -1.59$ dB.

Key Takeaway

The Shannon limit $E_b/N_0 = \ln 2 \approx -1.59$ dB is the ultimate energy efficiency bound. No communication scheme — no matter how clever — can achieve reliable transmission below this threshold. Deep-space probes (Voyager, New Horizons) operate within a few dB of this limit using turbo or LDPC codes with very low spectral efficiency.

Spectral Efficiency vs. $E_b/N_0$

The Shannon curve $\eta = \log_2(1 + \eta \cdot E_b/N_0)$ traces the boundary of the achievable region. Points below the curve are achievable; points above are not. Compare with practical modulation schemes (BPSK, QPSK, 16-QAM, 64-QAM, 256-QAM) at their respective operating points.

Parameters

Practical Modulation Schemes vs. Shannon Limit

Modulation	Spectral Efficiency (bits/s/Hz)	Required $E_b/N_0$ at BER $= 10^{-5}$ (dB)	Gap to Shannon (dB)
BPSK	1	9.6	11.2
QPSK	2	9.6	8.5
16-QAM	4	13.5	5.8
64-QAM	6	17.5	5.4
BPSK + Turbo	~1	0.7	2.3
LDPC (rate 1/2)	~1	0.2	1.8

The Gap to Capacity Shrinks with Coding

The comparison table reveals a striking fact: uncoded modulations operate 5–11 dB from the Shannon limit, while modern channel codes (turbo, LDPC) can close this gap to under 1 dB for long block lengths. The gap between uncoded modulation and the Shannon limit is split into a coding gain (from error-correcting codes) and a shaping gain (from using non-uniform input constellations to approximate the Gaussian distribution). Coding gain accounts for most of the gap; shaping gain contributes at most 1.53 dB (the so-called "shaping gap" of $\frac{\pi e}{6}$ ).

⚠️Engineering Note

Operating Point Selection in Real Systems

In practice, systems operate 1–3 dB from the AWGN Shannon limit. This gap comes from several sources: finite block length (especially for low-latency applications), imperfect channel estimation, implementation losses (finite-precision arithmetic, synchronization errors), and the requirement for a non-zero error rate target rather than the vanishing error rate assumed by Shannon.

The 5G NR standard uses adaptive modulation and coding with 29 MCS indices, each targeting a different point on the spectral efficiency curve. Link adaptation algorithms select the MCS that maximizes throughput at the current estimated SNR, typically targeting a block error rate of 10%.

Practical Constraints

•
Finite block length forces 0.5-1 dB gap even with optimal coding
•
Channel estimation overhead reduces effective SNR by 0.5-2 dB
•
Implementation losses (quantization, timing) add 0.2-0.5 dB

📋 Ref: 3GPP TS 38.214

Historical Note: Deep-Space Communication: Approaching the Shannon Limit

NASA's Voyager probes, launched in 1977, communicate from beyond the solar system using a transmit power of about 23 watts — less than a household light bulb. At a distance of over 23 billion km, the received SNR is extraordinarily low. The Voyager system uses a (7, 1/2) convolutional code concatenated with a Reed-Solomon outer code, operating at a spectral efficiency near zero and an $E_b/N_0$ of about 2.5 dB — roughly 4 dB from the Shannon limit.

More modern missions (Mars orbiters) use turbo codes operating within 1 dB of the limit. The James Webb Space Telescope uses LDPC codes. Shannon's 1948 formula predicted that such performance was possible; it took the engineering community 50 years to build codes that actually approach it.

Common Mistake: Infinite Bandwidth Does Not Mean Infinite Capacity

Mistake:

Assuming that taking $W \to \infty$ gives $C \to \infty$ because $C = W\log_2(1 + \text{SNR})$ .

Correction:

As $W \to \infty$ with fixed power $P$ , the per-Hz SNR $\text{SNR} = P/(N_0W)$ vanishes. The capacity converges to $C \to P/(N_0\ln 2)$ , a finite limit. Bandwidth is "free" only up to a point — beyond a certain bandwidth, additional spectrum yields diminishing returns because the noise power grows proportionally.

Quick Check

A system operates at $E_b/N_0 = 0$ dB (i.e., $E_b/N_0 = 1$ in linear). Is reliable communication possible?

Yes, because $0$ dB $> -1.59$ dB

No, because $E_b/N_0$ must exceed $0$ dB for reliable communication

It depends on the bandwidth

Only with an infinite block length