Ferkans — Interactive Telecom Tutor

Beyond the Asymptotic Regime

The point is that Shannon's 1948 capacity theorem is ASYMPTOTIC: it says "for blocklength $n \to \infty$ , rate $R < C$ is achievable with vanishing error probability." Real systems have finite $n$ and target BLER levels. Ultra-Reliable Low-Latency Communication (URLLC) in 5G NR targets BLER $10^{-5}$ at 1 ms latency — totally outside Shannon's regime. Polyanskiy-Poor-Verdú (2010) gave the modern finite-blocklength framework. The theory is elegant; the design consequences for coded modulation are still being worked out.

Definition:
Finite-Blocklength Achievable Rate

The finite-blocklength achievable rate $R^*(n, \epsilon)$ is the largest rate at which there exists a code of blocklength $n$ with block-error probability at most $\epsilon$ . Formally, $R^*(n, \epsilon) = \frac{1}{n}\log_2 M^*(n, \epsilon),$ where $M^*(n, \epsilon)$ is the maximum codebook size. For $n \to \infty$ , $R^*(n, \epsilon) \to C$ — Shannon's theorem. The question is: what IS $R^*$ at finite $n$ ?

Definition:
Channel Dispersion $V$

The channel dispersion $V$ of a DMC or AWGN channel is the variance of the information density per channel use at capacity- achieving input. For the real AWGN channel with SNR $\rho$ : $V(\rho) = \frac{\rho(\rho + 2)}{2(\rho + 1)^2} \cdot (\log_2 e)^2.$ $V$ measures how much the achievable rate varies when one averages mutual-information densities — the second-order cost of finite blocklength.

Theorem: Polyanskiy-Poor-Verdú Normal Approximation

For the AWGN channel with SNR $\rho$ , blocklength $n$ , and target BLER $\epsilon \in (0, 1/2)$ : $R^*(n, \epsilon) = C(\rho) - \sqrt{\frac{V(\rho)}{n}}\,Q^{-1}(\epsilon) + O\!\left(\frac{\log n}{n}\right)$ where $C(\rho) = \frac{1}{2}\log_2(1 + \rho)$ is the AWGN capacity, $V(\rho)$ is the channel dispersion, and $Q^{-1}$ is the inverse Q-function.

Proof

Achievability: random Gaussian codebook + Feinstein's maximal coding

Start from Shannon's random-coding argument with a Gaussian input. The error probability under maximum-likelihood decoding is bounded by a CLT-like expression involving the information density random variable $i_n(X^n; Y^n)$ .

Information density CLT

$i_n(X^n; Y^n) = n C + \sqrt{n V}\,Z + o(\sqrt{n})$ where $Z \to \mathcal{N}(0, 1)$ by the Central Limit Theorem applied to the i.i.d. per-use information densities.

Invert the tail

Choosing the rate threshold so that the tail probability equals $\epsilon$ : $R \le C - \sqrt{V/n}\,Q^{-1}(\epsilon)$ . This matches the expression above up to logarithmic corrections.

Converse

The meta-converse bound (Polyanskiy's "Theorem 27") matches the achievability up to $O(\log n / n)$ , establishing tightness of the leading-order terms. $\blacksquare$

Finite-Blocklength Rate for URLLC

Maximum rate vs blocklength $n$ at fixed SNR and target BLER. The gap to Shannon's capacity widens as $n$ shrinks — this is the URLLC design constraint.

Parameters

SNR [dB]5

Target BLER

\epsilon

0.00001

Example: 5G NR URLLC: Rate at $n = 100$ , $\epsilon = 10^{-5}$

At SNR 5 dB (linear $\rho = 10^{0.5} \approx 3.16$ ), compute the Polyanskiy approximation for $n = 100$ , $\epsilon = 10^{-5}$ . How does this compare to Shannon's $C$ ?

Solution

Capacity

$C = \frac{1}{2}\log_2(1 + 3.16) = 1.05$ bits/use.

Dispersion

$V = 3.16 \cdot 5.16 / (2 \cdot 4.16^2) \cdot (\log_2 e)^2 = 0.470 \cdot 2.081 = 0.978$ .

Q inverse

$Q^{-1}(10^{-5}) \approx 4.265$ .

Finite-blocklength rate

$R^* \approx 1.05 - \sqrt{0.978/100} \cdot 4.265 = 1.05 - 0.42 = 0.63$ bits/use. A 40% HIT relative to Shannon — the URLLC penalty.

Engineering implication

URLLC requires either larger SNR, shorter packets (with accepted rate loss), or LOWER target BLER flexibility. The asymptotic capacity is a misleading design target for URLLC.

,

Design Consequences for URLLC

The finite-blocklength framework reshapes URLLC design:

MCS tables must include LOW-rate, HIGH-reliability modes (5G NR MCS Table 3 starts at rate 30/1024 ≈ 3%).
Repetition / HARQ become essential — not for throughput but for additional diversity copies that tighten the CLT approximation.
Pilot / DMRS overhead dominates at very short blocklengths — pilotless "blind" transmission is an active research area.

⚠️Engineering Note

URLLC in 5G NR

5G NR URLLC (Rel-15, frozen 2018) targets:

BLER $10^{-5}$ at 1 ms latency.
Short TTIs (2 or 7 OFDM symbols per slot, vs 14 in eMBB).
MCS Table 3 (mandatory for URLLC PDSCH): starts at QPSK rate 30/1024 (extremely robust), caps at 64-QAM rate 772/1024.
Enhanced DMRS density and mini-slot scheduling.
Pre-emption indication for URLLC-in-eMBB frame. Rel-16 extensions added AI-based link adaptation and sidelink-URLLC.

Common Mistake: Shannon Is Not a URLLC Design Target

Mistake:

"URLLC targets 80% of Shannon capacity with a 1 dB margin."

Correction:

Shannon's $C$ is an ASYMPTOTIC limit and is NOT achievable at URLLC blocklengths. The real limit is the Polyanskiy bound $C - \sqrt{V/n}\,Q^{-1}(\epsilon)$ , which at $n = 100$ , $\epsilon = 10^{-5}$ , and typical SNRs is 30-50% BELOW Shannon. Using Shannon as the design target overestimates achievable rate by a large factor.

Key Takeaway

Shannon's capacity is asymptotic. At URLLC blocklengths ( $n \sim 100$ ) and BLERs ( $\epsilon \sim 10^{-5}$ ), the Polyanskiy-Poor- Verdú bound $C - \sqrt{V/n}\,Q^{-1}(\epsilon)$ is the correct design target. Its gap to Shannon is the fundamental cost of short packets + high reliability.

Coded Modulation for URLLC and Short Packets