Fast Gao-like Decoding of Horizontally Interleaved Linearized Reed-Solomon Codes

TL;DR

This paper introduces a fast Gao-like decoding algorithm for horizontally interleaved linearized Reed-Solomon codes, achieving subquadratic complexity and enabling decoding beyond the unique-decoding radius, which improves efficiency in code-based cryptography.

Contribution

It presents a novel Gao-like decoder for HILRS codes with subquadratic complexity, extending decoding capabilities beyond the traditional radius and improving over existing syndrome-based methods.

Findings

Achieves subquadratic decoding complexity in code length.

Decodes errors beyond the unique-decoding radius.

Provides tight bounds on failure probability.

Abstract

Both horizontal interleaving as well as the sum-rank metric are currently attractive topics in the field of code-based cryptography, as they could mitigate the problem of large key sizes. In contrast to vertical interleaving, where codewords are stacked vertically, each codeword of a horizontally $s$-interleaved code is the horizontal concatenation of $s$ codewords of $s$ component codes. In the case of horizontally interleaved linearized Reed-Solomon (HILRS) codes, these component codes are chosen to be linearized Reed-Solomon (LRS) codes. We provide a Gao-like decoder for HILRS codes that is inspired by the respective works for non-interleaved Reed-Solomon and Gabidulin codes. By applying techniques from the theory of minimal approximant bases, we achieve a complexity of $\tilde{\mathcal{O}}(s^{2.373} n^{1.635})$ operations in $\mathbb{F}_{q^m}$, where $\tilde{\mathcal{O}}(\cdot)$ neglects logarithmic factors, $s$ is the interleaving order and $n$ denotes the length of the component codes. For reasonably small interleaving order $s \ll n$, this is subquadratic in the component-code length $n$ and improves over the only known syndrome-based decoder for HILRS codes with quadratic complexity. Moreover, it closes the performance gap to vertically interleaved LRS codes for which a decoder of complexity $\tilde{\mathcal{O}}(s^{2.373} n^{1.635})$ is already known. We can decode beyond the unique-decoding radius and handle errors of sum-rank weight up to $\frac{s}{s + 1} (n - k)$ for component-code dimension $k$. We also give an upper bound on the failure probability in the zero-derivation setting and validate its tightness via Monte Carlo simulations.