My name is Iddo Naiss, and I am a student in the joint math-ee program in BGU, now studying for my Masters degree in the fast route from Bs.C. My advisor is Haim Permuter
Our work-The Blahut-Arimoto Algorithm; extantions
· The Blahut Arimoto type algorithms are used for optimizing the directed (or mutual) information, which is known to have an important role in characterizing the capacity of various channels and the rate distortion for various sources with distortion.
· In our work we extend the Blahut Arimoto Algorithm for estimating the capacity of feedback channels and the rate distortion for source coding with feed forward, which are characterized by the optimization problem over the directed information
· Our first paper is about maximizing the directed information in order to estimate the capacity of feedback channels. The paper is being reviewed by IEEE trans. of info. theory. In the meanwhile, it was uploaded to arXiv tohttp://arxiv.org/abs/1012.5071. The slides
· Our second paper proves that the expression given above for R(D) is true, and then apply a BA type algorithm onto this new optimization problem. We also transform this problem to a geometrical programming maximization one, which is given as follows:
This optimization problem can be solved using a Matlab code(cvx).
· Follows is a header *.h code that consists of all functions and procedures necessary in order to use the extended algorithm
· In order to illustrate the algorithms performance, we give 2 *.c codes.
The first is for the feedback channel case (trapdoor channel with m=3 states and delay d=1)
The second, for rate distortion source coding with feedforward (an a-symetrical markov source as a function of the delay for block length n=10)
· Both *.c files has two *.txt output that consists of tables. The first is of the rate, where each line is for a different block length. The other is the same for the distortion. One can use Matlab to plot the data.
· Here I present the matlab files for the channel with feedback case. They are less efficient but are easier to use.
· In order to illustrate the algorithms performance, we give a couple of *.m codes.
The first are to built the channel we wish (trapdoor channel with m=3 states and delay d=1)
The state matrix
expanding the matrix probability to a higher rank for future usage
and finally the channel
The last file is for using the codes above