Capacity of noisy, discrete memoryless channels under input constraints

In this thesis work, we examine the capacity of discrete memoryless channels under input constraints. We consider a certain class of input-restricted channels for which constrained sequences can be modeled as outputs of a finite-state machine(FSM). No efficient algorithm is known for computing the capacity of such a channel. For the noiseless case, i.e., when the channel input letter and the corresponding output letter are identical, it is shown that [1] the channel capacity is the logarithm of the largest eigenvalue of the adjacency matrix of the state-transition diagram of the FSM generating the allowed channel input sequences. Furthermore, the probability distribution on the input sequences achieving the channel capacity is first-order markovian. Here, we discuss the noisy case. For a specific input-restricted channel, we show that unlike the noiseless case, the capacity is no longer achieved by a first-order distribution. We derive upper and lower bounds on the maximum rate achievable by a K-th order markovian distribution on the allowed input sequences. The computational results show that the second-order distribution does strictly better than the first-order distribution for this particular channel. A sequence of upper bounds on the capacity of an input-restricted channel is also given. We show that this sequence converges to the channel capacity. The numerical results clarify that markovian distribution may achieve rates close to the capacity for the channel considered in this work.