Enhancements in Score-based Channel Estimation for Real-Time Wireless Systems

Florian Strasser, Marion Bäro, Wolfgang Utschick · 2025

Read Enhancements in Score-based Channel Estimation for Real-Time Wireless Systems by Florian Strasser, Marion Bäro, Wolfgang Utschick (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-07839.txt.

Introduction

Faster wireless channel estimation is the goal of a 2025 paper by Florian Strasser and 2 co-authors. Before a MIMO system, which uses many antennas at once, can send data well, it must estimate the radio channel from pilot signals. Score-based generative models do this job accurately but can be too slow for real-time use.

The team redesigns the noise schedule and skips steps during sampling, cutting the number of denoising steps the model must run. An extreme version uses a single denoising step guided by signal-to-noise information. On synthetic city-like channel data, latency drops sharply without hurting accuracy. Speed of this kind decides whether a clever estimator can leave the lab and run inside live networks.

Catalog record

CREATORFlorian Strasser, Marion Bäro, Wolfgang Utschick
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-07839

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Enhancements in Score-based Channel Estimation for Real-Time Wireless Systems Thanks: This work was funded by Huawei Technologies Sweden AB, Stockholm. Florian Strasser, Marion Bäro, Wolfgang Utschick Abstract We propose enhancements to score-based generative modeling techniques for low-latency pilot-based channel estimation in a point-to-point single-carrier multiple-input multiple-output (MIMO) wireless system. Building on recent advances in score-based models, we investigate a specific noise schedule design and sampling acceleration by step-skipping to reduce the number of denoising steps during inference. We additionally propose a single-step signal-to-noise ratio informed denoiser as an extreme case of the step-skipping approach. Our methods achieve significant latency reductions without performance degradation, as demonstrated on a synthetic channel dataset representing an urban macrocell MIMO communications scenario. I Introduction Accurate channel state information (CSI) acquisition is a key factor when scaling up wireless communication systems [ 1 ] . However, conventional estimators based on maximum likelihood, minimum mean square error (MMSE) , and Bayesian techniques are reaching their limits. Therefore, machine learning (ML) and in particular generative artificial intelligence (GenAI) based methods have been proposed due to their great potential in learning the underlying prior distribution of a wireless channel [ 2 ] . This distribution covers the site-specific channel characteristics of a base station (BS) ’s surrounding. In accordingly designed algorithms, exploiting this prior knowledge can significantly improve the channel estimation quality. Hence, various GenAI -based channel estimation algorithms using, e.g, Gaussian mixture models (GMMs) [ 3 ] , generative adversarial networks (GANs) [ 4 ] , and variational autoencoders (VAEs) [ 5 ] have recently been proposed. The latest advancements derive channel estimators based on diffusion models (DMs) [ 6 ] and the closely related score-based generative models (SBMs) [ 7 ] , showcasing great potential. Especially, their prevailing generative abilities compared to, e.g., GANs , together with a stable training process and design flexibility, make them interesting candidates for wireless algorithm development. The major drawback of DMs and SBMs is their high latency during inference due to the iterative sampling procedure. To deploy algorithms that utilize DMs / SBMs in real-time wireless communications systems, we thus have to overcome this limitation. I-A Related Work The authors of [ 8 ] introduced a SBM -based framework, incorporating the pilot information in a guidance term during posterior sampling. However, their implementation exhibits high computational complexity. The channel estimator proposed in [ 9 ] features a deterministic adaptation of the denoising diffusion probabilistic model (DDPM) [ 6 ] and a signal-to-noise ratio (SNR) -dependent denoising procedure. Also, by exploiting structural properties of the multiple-input multiple-output (MIMO) system model, the authors can utilize a lightweight neural network (NN) implementation, reducing computational complexity. Building on this work, [ 10 ] extends the channel estimator to a broader set of problems, specifically under-determined and quantized channel estimation. Additionally, they propose a framework solely using noisy pilots for training, altogether avoiding the necessity of ground-truth channels. In [ 11 ] , the SBM - and DDPM -based methods are compared to a channel estimator based on a denoising diffusion implicit model (DDIM) [ 12 ] , which exhibits better performance and also reduces the number of denoising steps. Another recently published work [ 13 ] , incorporating many ideas of [ 9 ] , improves the estimator’s efficiency even more by using the DPM-Solver framework [ 14 ] . The authors of [ 15 ] propose to use a Brownian-Bridge DM in an integrated sensing and communications system to estimate the sensing channel. Apart from channel estimation, DMs and SBMs have also been proposed for semantic communications [ 16 , 17 , 18 ] , CSI data augmentation [ 19 , 20 ] , joint channel estimation and data detection [ 21 ] and end-to-end learning [ 22 ] . I-B Contributions The main focus of this work is to further reduce the complexity of a SBM -based MIMO channel estimation algorithm. We examine improving the iterative denoising procedure of the SBM rather than any improvements in the NN architecture itself. Our contributions can be summarized as follows: • We adapt the channel estimation algorithm proposed in [ 9 ] using formulations for a SBM with a variance-exploding (VE) Gaussian kernel. • We propose two different techniques to accelerate the denoising procedure, i.e., a specific noise schedule design and a step-skipping approach. • We derive a single-step SNR -informed channel estimator, which is the extreme case of the step-skipping idea. • We validate the effectiveness of our proposed SBM enhancements with numerical simulations conducted on synthetic channel data. The rest of this work is structured as follows. We introduce the channel estimation problem and the system model in Section II . Section III shortly discusses the SBM framework and explains its usage in the channel estimation algorithm. Enhancements of the SBM -based channel estimator are proposed in Section IV and supported with simulation results presented in Section V . Section VI concludes this work. I-C Notation Bold-faced lower-case and upper-case letters denote vectors and matrices, respectively. The matrix {\bm{F}}_{N} describes an N -dimensional discrete Fourier transformation (DFT) matrix and \mathbf{I}_{M} is the M -dimensional identity matrix. The vectorization (stacking of columns) of a matrix {\bm{A}}\in\mathbb{C}^{M\times N} is denoted as \vect({{\bm{A}}})\in\mathbb{C}^{MN} , and {\bm{A}}\otimes{\bm{B}}\in\mathbb{C}^{MP\times NQ} is the Kronecker product of {\bm{A}} and {\bm{B}}\in\mathbb{C}^{P\times Q} . A complex-valued multivariate Gaussian distribution with mean {\bm{\mu}} and covariance matrix {\bm{C}} is described by \mathcal{N}_{\mathbb{C}}({\bm{\mu}},{\bm{C}}) . The gradient of a function f w.r.t. a variable {\bm{x}} is written as \nabla_{{\bm{x}}}f . II Problem Formulation We consider a point-to-point single-carrier MIMO system with an {N_{\mathrm{tx}}} -antenna transmitter and an {N_{\mathrm{rx}}} -antenna receiver. The pilot-based transmission scheme is described by {\bm{Y}}={\bm{H}}{\bm{P}}+{\bm{N}}, (1) where {\bm{Y}}\in\mathbb{C}^{{N_{\mathrm{rx}}}\times{N_{\mathrm{p}}}} comprises the received signals, {\bm{H}}\in\mathbb{C}^{{N_{\mathrm{rx}}}\times{N_{\mathrm{tx}}}} describes the unknown channel matrix, {\bm{P}}\in\mathbb{C}^{{N_{\mathrm{tx}}}\times{N_{\mathrm{p}}}} is the unitary pilot matrix satisfying {\bm{P}}{\bm{P}}^{\mathrm{H}}=\mathbf{I}_{N_{\mathrm{tx}}} and {\bm{N}}\in\mathbb{C}^{{N_{\mathrm{rx}}}\times{N_{\mathrm{p}}}} describes Gaussian noise with columns {\bm{n}}_{i}\sim\mathcal{N}_{\mathbb{C}}({\bm{0}},\eta^{2}\mathbf{I}) . The objective is to recover the unknown channel matrix {\bm{H}} from the observed signal {\bm{Y}} . We assume {N_{\mathrm{p}}}={N_{\mathrm{tx}}} , resulting in a fully-determined channel estimation scenario. Rewriting ( 1 ) as {\bm{y}}=({\bm{P}}^{\mathrm{T}}\otimes\mathbf{I}_{{N_{\mathrm{rx}}}}){\bm{h}}+{\bm{n}}={\bm{A}}{\bm{h}}+{\bm{n}}, (2) with {\bm{y}}=\vect({\bm{Y}}) , {\bm{h}}=\vect({\bm{H}}) , {\bm{n}}=\vect({\bm{N}}) and {\bm{A}}=({\bm{P}}^{\mathrm{T}}\otimes\mathbf{I}_{{N_{\mathrm{rx}}}}) enables us to work with vectors instead of matrices. Note that {\bm{A}} is also unitary since it is derived by the Kronecker product of two unitary matrices. The channels are assumed to be distributed according to an unknown prior p({\bm{h}}) . Assuming \E_{p}[{\bm{h}}]={\bm{0}} and \E_{p}[\|{\bm{h}}\|^{2}]={N_{\mathrm{tx}}}{N_{\mathrm{rx}}} leads to the SNR definition \mathrm{SNR}({\bm{y}})=1/\eta^{2} . III Preliminaries III-A Score-based Generative Modeling The SBM considered in this work is based on the VE stochastic differential equation (SDE) framework introduced in [ 7 ] , which is briefly reviewed in this section. We use the forward SDE \operatorname{d}\!{\bm{h}}(t)=\sqrt{\frac{\operatorname{d}\![\sigma^{2}(t)]}{\operatorname{d}\!t}}\operatorname{d}\!{\bm{w}}(t) (3) for t\in[0,T] with {\bm{h}}(0)\coloneqq{\bm{h}}\sim p({\bm{h}}) and a standard Wiener process {\bm{w}}(t) to continuously diffuse a ground-truth channel sample into Gaussian noise. The noise schedule \sigma(t) controls the diffusion process and must satisfy the condition \lim_{t\to 0}\sigma(t)=0 . Alternatively, a noisy sample can directly be produced by sampling from the VE Gaussian kernel q_{\sigma}({\bm{h}}(t)|{\bm{h}}(0))=\mathcal{N}_{\mathbb{C}}({\bm{h}}(t);{\bm{h}}(0),\sigma^{2}(t)\mathbf{I}) , such that {\bm{h}}(t)={\bm{h}}(0)+\sigma(t){\bm{\epsilon}}(t) (4) using the reparameterization trick with {\bm{\epsilon}}(t)\sim\mathcal{N}_{\mathbb{C}}({\bm{0}},\mathbf{I}) . This leads to the SBM ’s SNR definition \mathrm{SNR}^{\mathrm{SBM}}(t)=1/\sigma^{2}(t) . Input: Training dataset in beamspace domain \mathcal{H}_{\textrm{train}}^{\textrm{beam}} , noise variances \{\sigma_{k}\}_{k=1}^{K} for epoch \leftarrow 1,\ldots,n_{\textnormal{epochs}} do 1 for batch

Read the complete text