Lossless data compression with polar codes
MetadataShow full item record
Please cite this item using this persistent URLhttp://hdl.handle.net/11693/15869
In this study, lossless polar compression schemes are proposed for finite source alphabets in the noiseless setting. In the first part, lossless polar source coding scheme for binary memoryless sources introduced by Arıkan is extended to general prime-size alphabets. In addition to the conventional successive cancellation decoding (SC-D), successive cancellation list decoding (SCL-D) is utilized for improved performance at practical block-lengths. For code construction, greedy approximation method for density evolution, proposed by Tal and Vardy, is adapted to non-binary alphabets. In the second part, a variable-length, zero-error polar compression scheme for prime-size alphabets based on the work of Cronie and Korada is developed. It is shown numerically that this scheme provides rates close to minimum source coding rate at practical block-lengths under SC-D, while achieving the minimum source coding rate asymptotically in the block-length. For improved performance at practical block-lengths, a scheme based on SCL-D is developed. The proposed schemes are generalized to arbitrary finite source alphabets by using a multi-level approach. For practical applications, robustness of the zero-error source coding scheme with respect to uncertainty in source distribution is investigated. Based on this robustness investigation, it is shown that a class of prebuilt information sets can be used at practical block-lengths instead of constructing a specific information set for every source distribution. Since the compression schemes proposed in this thesis are not universal, probability distribution of a source must be known at the receiver for reconstruction. In the presence of source uncertainty, this requires the transmitter to inform the receiver about the source distribution. As a solution to this problem, a sequential quantization with scaling algorithm is proposed to transmit the probability distribution of the source together with the compressed word in an efficient way.