Improved GPR-Based CSI Acquisition via Spatial-Correlation Kernel
Syed Luqman Shah, Graduate Student Member , IEEE , Nurul Huda Mahmood, Member , IEEE , and Italo Atzeni, Senior Member , IEEE · 2026
Read Improved GPR-Based CSI Acquisition via Spatial-Correlation Kernel by Syed Luqman Shah, Graduate Student Member , IEEE , Nurul Huda Mahmood, Member , IEEE , and Italo Atzeni, Senior Member , IEEE (2026) 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-2601-14759.txt.
Introduction
Syed Luqman Shah and co-authors improve channel estimation for wireless systems in this 2026 paper. Accurate channel estimation with low pilot overhead and computational complexity is key to efficiently utilizing multi-antenna wireless systems. Pilots are known signals sent so receivers can measure the channel. Fewer pilots mean more room for real data.
The field is evolving from purely statistical descriptions toward physics- and geometry-aware propagation models. This work focuses on incorporating channel information in that newer spirit. Their method improves GPR-based CSI acquisition via a spatial-correlation kernel. GPR is Gaussian process regression, and CSI is channel state information. The gain is room. Each saved pilot is a slot that can carry pay data in its place. Readers get sharper channel knowledge at lower pilot cost.
Catalog record
| CREATOR | Syed Luqman Shah, Graduate Student Member , IEEE , Nurul Huda Mahmood, Member , IEEE , and Italo Atzeni, Senior Member , IEEE |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 33 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-14759
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Improved GPR-Based CSI Acquisition via Spatial-Correlation Kernel Syed Luqman Shah, Graduate Student Member , IEEE , Nurul Huda Mahmood, Member , IEEE , and Italo Atzeni, Senior Member , IEEE Abstract Accurate channel estimation with low pilot overhead and computational complexity is key to efficiently utilizing multi-antenna wireless systems. Motivated by the evolution from purely statistical descriptions toward physics- and geometry-aware propagation models, this work focuses on incorporating channel information into a Gaussian process regression (GPR) framework for improving the channel estimation accuracy. In this work, we propose a GPR-based channel estimation framework along with a novel Spatial-correlation (SC) kernel that explicitly captures the channel’s second-order statistics. We derive a closed-form expression of the proposed SC-based GPR estimator and prove that its posterior mean is optimal in terms of minimum mean-square error (MMSE) under the same second-order statistics, without requiring the underlying channel distribution to be Gaussian. Our analysis reveals that, with up to 50 % 50\% pilot overhead reduction, the proposed method achieves the lowest normalized mean-square error, the highest empirical 95 % 95\% credible-interval coverage, and superior preservation of spectral efficiency compared to benchmark estimators, while maintaining lower computational complexity than the conventional MMSE estimator. I Introduction Gaussian process regression (GPR) has emerged as a promising approach for channel state information (CSI) acquisition under tight pilot budgets [ 13 , 7 ] , which is critical for 6G-and-beyond wireless networks [ 6 ] . In contrast, conventional least-squares (LS) and minimum mean-square error (MMSE) estimators require pilot sequences whose length scales with the number of transmit antennas, resulting in substantial overhead for large antenna arrays [ 9 ] . Though techniques based on compressed sensing mitigate this overhead by exploiting sparsity in the angular or delay domains [ 2 , 9 ] , their performance is sensitive to noise and violations of low-rank assumptions in rich-scattering environments [ 10 ] . These limitations motivate the use of GPR-based CSI acquisition, which offers improved estimation accuracy [ 13 ] while reducing the pilot overhead [ 7 , 9 ] . GPR provides a non-parametric Bayesian framework for channel estimation, encoding prior knowledge through a covariance function (kernel). The posterior mean yields the channel estimate, while the posterior covariance quantifies its uncertainty [ 12 , 9 ] . GPR-based models have been successfully applied to interference prediction and proactive resource allocation [ 8 ] , hardware distortion modeling [ 1 ] , and CSI prediction for future channel forecasting [ 5 ] . More recently, [ 9 , 7 ] modeled wireless propagation over a two-dimensional antenna grid and showed that observing only a subset of transmit antennas channel response enables reconstruction of unobserved channel entries. Despite these advances, existing GPR-based channel estimation methods [ 13 , 9 , 7 ] suffer from three fundamental limitations: i) marginal-likelihood-based hyperparameter learning incurs rapidly increasing computational complexity with antenna dimension [ 12 ] ; ii) commonly used distance-based covariance kernels implicitly assume stationarity and weak anisotropy, leading to model mismatch in channels with direction-dependent scattering and non-separable transmit-receive coupling [ 3 ] ; and iii) prior second-order channel statistics are not explicitly incorporated under sparse training, limiting systematic exploitation of known covariance structures [ 7 ] . This work builds on the reduced-pilot GPR framework proposed in [ 7 ] and addresses the above limitations by introducing a novel Spatial-correlation (SC) kernel that is induced directly by the channel covariance. Unlike distance-based kernels, the proposed SC kernel derives its covariance from established channel models. This eliminates marginal-likelihood hyperparameter learning and reduces computational complexity while improving estimation accuracy under sparse training. We show that the proposed SC kernel is positive semidefinite (PSD). Based on the SC kernel, we propose an SC-based GPR estimator (SC-GPR). We derive a closed-form expression for the proposed SC-GPR estimator whose posterior mean is MMSE-optimal under the same second-order statistics. The resulting SC-GPR further exploits the structure of standard channel models to enable scalable implementation. In addition to achieving the MMSE mean, the proposed SC-GPR offers the following advantages: i) reduced pilot overhead; ii) uncertainty-aware channel estimates via credible intervals; and iii) lower computational cost. We validate the proposed approach using two closed-form channel covariance families under three reduced-overhead schemes that decrease the pilot load by 50\% , 67\% , and 75\% , respectively. Numerical results show that, even with a 50\% pilot reduction, the proposed method achieves lower normalized mean-square error (NMSE) than the benchmark schemes. With a 75\% reduction, it further outperforms the benchmarks in terms of spectral efficiency (SE), while also incurring lower computational complexity than the optimal MMSE estimator. II System Model We consider a narrowband point-to-point MIMO link with N_{\textrm{t}} transmit antennas and N_{\textrm{r}} receive antennas. During the channel estimation phase, the pilot overhead is reduced by activating only a subset of the N_{\textrm{t}} antennas instead of sounding the entire array. Specifically, only n_{\textrm{t}}<N_{\textrm{t}} transmit antennas are activated, which relaxes the requirement on the pilot length from T\geq N_{\textrm{t}} to T\geq n_{\textrm{t}} [ 7 ] . The pilots transmitted by these n_{\textrm{t}} active antennas are received across all the N_{\textrm{r}} receive antennas [ 9 ] . Let \Omega_{\textrm{t}}=\{a_{1},\ldots,a_{n_{\textrm{t}}}\}\subset\{1,\ldots,N_{\textrm{t}}\} denote an ordered index set corresponding to the n_{\textrm{t}} antennas, with no repeated elements. We adopt a structured equispaced activation pattern with stride \Delta\in\mathbb{N} , with a_{i}=1+(i-1)\Delta for i=1,\ldots,n_{\textrm{t}} , where n_{\textrm{t}}=\left\lfloor(N_{\textrm{t}}-1)/\Delta\right\rfloor+1 is the maximum integer satisfying the array-size constraint 1+(n_{\textrm{t}}-1)\Delta\leq N_{\textrm{t}} . Let \mathbf{S}\in\mathbb{C}^{n_{\textrm{t}}\times T} denote the pilot matrix transmitted by the n_{\textrm{t}} transmit antennas, whose rows are orthonormal and satisfy \mathbf{S}\mathbf{S}^{\mathsf{H}}=\mathbf{I}_{n_{\textrm{t}}} . Based on \Omega_{\textrm{t}} , we define the antenna-selection matrix \mathbf{F}=\big[\mathbf{e}_{a_{1}}\ \cdots\ \mathbf{e}_{a_{n_{\textrm{t}}}}\big]=\mathbf{I}_{N_{\textrm{t}}}(:,\Omega_{\textrm{t}})\in\{0,1\}^{N_{\textrm{t}}\times n_{\textrm{t}}} , with \mathbf{F}^{\mathsf{H}}\mathbf{F}=\mathbf{I}_{n_{\textrm{t}}} and where \mathbf{e}_{a_{i}} denotes the a_{i} -th canonical basis vector. The resulting transmitted pilot signal is \mathbf{X}=\mathbf{F}\mathbf{S}\in\mathbb{C}^{N_{\textrm{t}}\times T} . The received signal is given by \mathbf{Y}=\mathbf{H}\mathbf{X}+\mathbf{N}=\mathbf{H}\mathbf{F}\mathbf{S}+\mathbf{N} , where \mathbf{H}\in\mathbb{C}^{N_{\textrm{r}}\times N_{\textrm{t}}} denotes the channel matrix and \mathbf{N} is the noise matrix with i.i.d. \mathcal{CN}(0,\sigma_{\textrm{n}}^{2}) entries. De-spreading the pilots by right-multiplication with \mathbf{S}^{\mathsf{H}} yields the observation matrix \tilde{\mathbf{Y}}=\mathbf{Y}\mathbf{S}^{\mathsf{H}}=\mathbf{H}\mathbf{F}+\mathbf{N}\mathbf{S}^{\mathsf{H}}\in\mathbb{C}^{N_{\textrm{r}}\times n_{\textrm{t}}} . Thus, \tilde{\mathbf{Y}} contains noisy observations of the n_{\textrm{t}} sounded columns of \mathbf{H} indexed by \Omega_{\textrm{t}} [ 9 ] , while the remaining N_{\textrm{t}}-n_{\textrm{t}} channel columns remain unobserved and are inferred via the proposed SC-GPR method at the receiver, as detailed in the following section. Equivalently, these matrix observations correspond to noisy samples of individual channel entries indexed on the antenna grid, which we use as training set for GPR. III The Proposed GPR With the SC Kernel This section introduces the proposed SC-GPR estimator, establishes its statistical properties, and presents a scalable inference routine for reduced-pilot CSI acquisition. Our approach adopts the proposed SC kernel for GPR, which: i) encodes the channel’s second-order structure; ii) removes hyperparameter training to reduce the computation cost; iii) yields a closed-form estimator with calibrated uncertainty; and iv) exploits model structure for scalability. The proposed SC kernel prior captures anisotropy and non-separable transmit-receive coupling implied by standard covariance models, which are not represented by conventional distance-based kernels. As a result, the proposed SC-GPR enables more accurate channel reconstruction under sparse sounding. Training and test sets: Define the antenna index grid \mathcal{G}=\{(r,t):r\in\{1,\ldots,N_{\textrm{r}}\},\ t\in\{1,\ldots,N_{\textrm{t}}\}\}\subset\mathbb{N}^{2} , which serves as the input domain of the GPR. Let the observed (training) index set be \mathcal{X}=\{(r,a_{i}):r=1,\ldots,N_{\textrm{r}},\ i=1,\ldots,n_{\textrm{t}}\}\subset\mathcal{G} , with