2026/06/04 by Marcin Zukowski · 1 voice
Computer Science · #cs.DS #cs.PF
Huffman encoding has been an enduring technique for 70+ years, ubiquitous in compression algorithms since its invention. In this paper we propose a new approach to Huffman coding, based on a data structure from wavelet trees. The resulting pivot-coded Huffman (PivCo-Huffman) enables high-performance SIMD-friendly encoding and decoding operations. In our tests PivCo-Huffman consistently outperforms state-of-the-art Huffman codecs in decoding throughput. Additionally, we show how ANS-coding can be selectively applied to skewed nodes in this structure, yielding compression ratios approaching those of ANS-based codecs while preserving very high decompression speeds.