Back to the Bits: Parametric Symbol Compression for Semantic O-RAN Fronthaul

Bib

@ARTICLE{yoo2026bits,
  author={Yoo, Hanju and Choi, Dongha and Kim, Songkuk and Chae, Chan-Byoung and {Heath, Jr.}, Robert W.},
  journal={Submitted},
  title={Back to the Bits: Parametric Symbol Compression for Semantic O-RAN Fronthaul},
  year={2026}
}

Hanju Yoo^†, Dongha Choi^†, Songkuk Kim^†, Chan-Byoung Chae^†, and Robert W. Heath, Jr.^‡

^† School of Integrated Technology, Yonsei University, Seoul 03722, South Korea
^‡ Department of Electrical and Computer Engineering, University of California San Diego, La Jolla, CA 92093, USA

Scenario illustration for semantic O-RAN fronthaul compression — The semantic encoder may sit at the CU/DU, but the RU still needs a compact bitstream. The proposed codec compresses semantic I/Q symbols before fronthaul transport and keeps RU-side decoding neural-free.

Abstract

Semantic communication systems such as DeepJSCC convert input data into complex-valued I/Q symbols for efficient end-to-end transmission. In many practical deployments, these symbols must be further compressed into bits for fronthaul transport to radio units (RUs), which often lack neural network capabilities and have limited hardware resources. Yet, compression of semantic symbols for fronthaul has been understudied. In this paper, we present a lightweight parametric symbol compression scheme that models the symbol distribution using a (generalized) Student-t distribution and applies entropy coding based on this probability modeling. The receiver performs only probability table reconstruction and entropy decoding, making the method suitable for deployment on non-neural, low-power devices. Experimental results show that the proposed coder reduces rate by 19-19.2% vs. zlib and 7.2-7.6% vs. LZMA. It also provides an additional 19.1-20.2% rate reduction over a block floating-point baseline used in O-RAN fronthaul compression profiles. In runtime, it yields end-to-end codec speedup gains of up to ~2.9x over zlib and ~4.2-14.1x over LZMA.

Why This Matters

Most semantic communication papers stop at the learned encoder output, but O-RAN deployment does not. If the RU is simple, the fronthaul cannot assume another neural transform just to compress semantic latents. A useful codec must reduce rate, preserve semantic quality, and add only small side information.

This paper uses the distributional structure of semantic symbols directly. Since semantic latents are continuous-valued and heavy-tailed, generic byte-wise compressors such as zlib and LZMA miss symbol-level probability information. Block floating-point compression is also limited because it scales blocks but does not entropy-code according to the actual symbol law.

Key Idea

The method extends the spirit of modulation compression from finite constellations to neural semantic latents. Instead of sending a QAM constellation index, it sends quantized semantic-symbol indices under a compact fitted distribution. The probability model is small enough to describe with a few header bits per pack, but accurate enough to keep the cross-entropy close to the empirical source entropy.

19-19.2%Rate reduction versus `zlib`.

7.2-7.6%Rate reduction versus `LZMA`.

19.1-20.2%Additional rate reduction over BFP fronthaul compression.

up to 14.1xTotal codec speedup versus `LZMA`.

Key Results

Parametric symbol coding beats byte-wise dictionary compressors because it models the quantized semantic-symbol PMF directly.
Student-t gives the best rate-latency balance; generalized Student-t is more robust when the empirical distribution deviates from a pure Student-t.
Global fitting is preferable when latency dominates, while adaptive per-pack fitting is preferable when rate is the main constraint.
Compared with BFP, the proposed codec gains rate by using entropy coding rather than only block-wise scaling.
The method keeps inference-time neural complexity out of the RU.

Rate distortion curve for semantic symbol compression — Parametric codecs reach higher SQNR at the same bits/symbol than dictionary baselines, with the gap widening when the quantization alphabet becomes finer.

Rate latency tradeoff for semantic symbol compression — The useful region is lower-left: lower bits/symbol and lower runtime per Msymbol.

Bits breakdown for semantic symbol compression — Rate savings come primarily from reducing model mismatch while keeping side information small.