Exponents of polar codes using algebraic geometric code kernels (Q398987)

Polar codes were proposed by \textit{E. Arikan} [IEEE Trans. Inf. Theory 55, No. 7, 3051--3073 (2009; \url{doi:10.1109/TIT.2009.2021379})] for binary discrete memoryless channels and later generalized to any \(q\)-ary (\(q\)\, a prime power) discrete memoryless channels. These codes take a channel with input alphabet the finite field \(\mathbb{F}_q\)\, and a \(l \times l\)\ square matrix \(G\)\, over \(\mathbb{F}_q\)\, (the kernel matrix) for encoding \(l^n\)\, channels (using the \(n^{th}\)\, Kronecker power of \(G\)). \textit{S. Korada} et al. [IEEE Trans. Inf. Theory 56, No. 12, 6253--6264 (2010; \url{doi:10.1109/TIT.2010.2080990})] introduced a quantity called the exponent or the rate of polarization of \(G\). The matrix considered by Arikan, \( G:= \left( \begin{matrix} 1 & 0 \\ 1 & 1 \end{matrix} \right) \), had exponent \(1/2\)\, but Korada et al. showed that a matrix obtained from a generator matrix for a shortened BCH code had an exponent greater than \(1/2\). That motivated the authors of the present paper to consider general algebraic geometric (AG) codes as source of kernels for polar codes. The paper studies the construction of kernels from AG codes and in particular from Hermitian and Suzuki codes. The main conclusion is that ``both Hermitian and Suzuki codes have larger exponents that Reed-Solomon codes over the same field, for \(q\geq 3\); however, the larger exponents are at the expense of larger kernel matrices. Comparing kernels of the same size, though over different fields, we see that Reed-Solomon kernels have larger exponents than both Hermitian and Suzuki kernels.''