A noisy channel can carry information reliably even when individual transmitted symbols are sometimes confused.

The capacity $C$ of a channel is the boundary between information rates that can, in principle, be made arbitrarily reliable with suitable coding and rates that cannot.

Capacity depends on the channel model and on any constraints placed on its inputs, such as allowed signal power or symbol alphabet.

Information theory connects this operational limit with mutual information: the transmitter should use input statistics that make the channel output reveal as much information as possible about the input.

Capacity is therefore not the raw symbol rate, the bit rate of one modulation, or the performance of one particular code. It is a property of the communication channel together with its allowed input strategy.

Different channel models lead to different capacity formulas. Discrete memoryless channels and band-limited Gaussian-noise channels are two important cases.