Binary Linear Codes with Optimal Scaling and Quasi-Linear Complexity

We present the first family of binary codes that attains optimal scaling and quasi-linear complexity, at least for the binary erasure channel (BEC). In other words, for any fixed δ > 0, we provide codes that ensure reliable communication at rates within ε > 0 of the Shannon capacity with block length n = O(1/ε2+δ), construction complexity Θ(n), and encoding/decoding complexity Θ(n log n). Furthermore, this scaling between the gap to capacity and the block length is optimal in an information-theoretic sense. Our proof is based on the construction and analysis of binary polar codes obtained from large kernels. It was recently shown that, for all binary-input symmetric memoryless channels, conventional polar codes (based on a 2× 2 kernel) allow reliable communication at rates within ε > 0 of the Shannon capacity with block length, construction, encoding and decoding complexity all bounded by a polynomial in 1/ε. In particular, this means that the block length n scales as O(1/εμ), where μ is referred to as the scaling exponent. It is furthermore known that the optimal scaling exponent is μ = 2, and it is achieved by random linear codes. However, for general channels, the decoding complexity of random linear codes is exponential in the block length. As far as conventional polar codes, their scaling exponent depends on the channel, and for the BEC it is given by μ = 3.63. This falls far short of the optimal scaling guaranteed by random codes. Our main contribution is a rigorous proof of the following result: there exist `× ` binary kernels, such that polar codes constructed from these kernels achieve scaling exponent μ(`) that tends to the optimal value of 2 as ` grows. We furthermore characterize precisely how large ` needs to be as a function of the gap between μ(`) and 2. The resulting binary codes maintain the beautiful recursive structure of conventional polar codes, and thereby achieve construction complexity Θ(n) and encoding/decoding complexity Θ(n log n). This implies that block length, construction, encoding, and decoding complexity are all linear or quasi-linear in 1/ε2, which meets the information-theoretic lower bound.