Channel Capacity

3 min read#information

The maximum rate at which information can cross a noisy channel with arbitrarily small error.

Contents

Channel Capacity

Channel capacity is the top speed of a communication channel — the highest rate, in bits per use, at which you can send information through noise and still recover it essentially perfectly. Shannon's noisy-channel coding theorem (1948) is the stunning result that this top speed C is not just a practical limit but a sharp threshold:

  • Transmit at any rate below C and, with a clever enough Error-Correcting Code, you can drive the error probability arbitrarily close to zero.
  • Transmit above C and reliable communication is impossible, no matter how much redundancy you add.

Before Shannon, engineers assumed noise forced a grim trade-off: go faster, get more errors. Shannon showed that below capacity you can have both speed and reliability — you just need long, well-designed codes.

Capacity is maximized mutual information

The capacity is defined through Mutual Information. The channel turns an input X into a noisy output Y; the information that actually gets through is I(X;Y). Capacity is the best you can do over all ways of choosing the input:

C = \max_{p(x)} I(X;Y).

How noise eats the rate

The plot shows C = 1 - H(p) for the binary symmetric channel. Capacity is highest for a quiet channel, plunges as the flip probability climbs toward \tfrac12, then — strikingly — recovers for p > \tfrac12, because a channel that almost always flips is just as informative as one that almost never does (you simply invert its output).

Capacity of a binary symmetric channel, C = 1 − H(p)
Reliable bits per channel use versus bit-flip probability p. Zero at p = 0.5, where input and output are independent.

Bandwidth, power, and the analog version

For continuous channels Shannon gave an equally famous formula. A channel of bandwidth B hertz with signal-to-noise ratio \mathrm{SNR} has capacity

C = B\,\log_2(1 + \mathrm{SNR}) \quad \text{bits/second.}

Capacity grows linearly with the slice of Spectrum you occupy and only logarithmically with power — which is why grabbing more bandwidth is the cheaper road to higher data rates, and why every modem, Wi-Fi link, and deep-space probe is engineered right up against this line.

A binary symmetric channel flips each bit with probability p = 0.5. What is its capacity?

See also