EB-P2WNet: Env-Aided CSI Prediction

• The paper introduces EB-P2WNet, a neural architecture that exploits a digital twin-based channel subspace basis fused with sparse pilot CSI for complete channel reconstruction.
• It employs a hybrid CNN, Transformer, and LSTM framework to integrate static environmental priors with real-time measurements, significantly reducing pilot overhead.
• Quantitative results demonstrate up to 50% pilot savings and improved NMSE performance under low SNR, interference, and localization errors, highlighting the scheme's robustness.

Searching arXiv for the specified papers to ground the article in the current literature. arXiv search query: (Cai et al., 7 Aug 2025) EB-Based Partial-to-Whole CSI Prediction Network (EB-P2WNet) is a neural architecture for digital-twin-channel-aided channel state information (CSI) prediction that fuses an environment-specific channel subspace basis with real-time partial CSI to reconstruct the entire spatial-frequency domain channel at the present time and to predict CSI for the next channel coherent time. It is introduced within an environment-specific channel subspace basis (EB)-aided partial-to-whole channel state information prediction method (EB-P2WCP) for low-overhead and robust channel prediction in complex sixth-generation mobile communication scenarios, where the static electromagnetic environment is encoded through a digital twin and incorporated as prior information for CSI inference (Cai et al., 7 Aug 2025).

1. Conceptual scope and problem formulation

EB-P2WNet addresses the overhead of CSI acquisition in massive MIMO-OFDM settings by combining two information sources: a sparse pilot-based estimate of the instantaneous channel and an environment-based prior extracted from a digital twin map. In the formulation reported for the method, the true full OFDM MIMO channel is denoted by $\mathbf H\in\mathbb C^{M_t\times N_{sc}}$, while only $N_p\ll N_{sc}$ subcarriers per antenna are sounded in order to save pilots. The partial CSI estimate is represented as

$\mathbf H^0=\mathbf B\odot \mathbf H,$

where $\mathbf B\in\{0,1\}^{M_t\times N_{sc}}$ is a binary mask with each row containing exactly $N_p$ ones, and the vectorized form is

$$\mathbf h^0_{\mathrm{flat}}=\mathrm{vec}(\mathbf H^0)\in\mathbb C^{N_H},$$

with $N_H=M_t\times N_{sc}$ (Cai et al., 7 Aug 2025).

The corresponding prediction tasks are defined as a present-CSI mapping

$$M:\;(\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_t$$

and a future-CSI mapping

$$G:\;(\mathbf H^0_{t-T+1},\dots,\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_{t+1}.$$

Here $\mathbf K_t$ denotes the environment-specific basis associated with the user’s location in the digital twin. The method therefore belongs to a class of partial-to-whole CSI predictors in which subspace priors and sparse pilot observations are jointly exploited rather than treating CSI recovery as an exclusively data-driven interpolation problem (Cai et al., 7 Aug 2025).

A central design premise is that the static properties of the electromagnetic environment can be represented compactly and reused across prediction instances. This differs from approaches that rely only on instantaneous second-order statistics or only on partial CSI. A plausible implication is that EB-P2WNet is intended to shift a portion of CSI acquisition burden from online pilot transmission to offline environmental modeling.

2. Environment-based subspace basis extraction

The environment-specific basis, abbreviated EB, is extracted from a high-precision 3D reconstruction of the static environment in the digital twin, including buildings and roads. The coverage area is partitioned into uniform $N_p\ll N_{sc}$0 grids. In each grid, four corner locations $N_p\ll N_{sc}$1 are selected, and at each vertex $N_p\ll N_{sc}$2 Doppler-shifted channel snapshots $N_p\ll N_{sc}$3 are simulated. From these samples, the covariance matrix is formed as

$N_p\ll N_{sc}$4

Singular-value decomposition is then applied:

$N_p\ll N_{sc}$5

where the singular values satisfy $N_p\ll N_{sc}$6 (Cai et al., 7 Aug 2025).

The basis dimension $N_p\ll N_{sc}$7 is chosen as the smallest value such that the cumulative energy ratio satisfies

$N_p\ll N_{sc}$8

In the reported simulations, $N_p\ll N_{sc}$9 captures over 95% of the energy, and the environment-specific basis for a grid is

$\mathbf H^0=\mathbf B\odot \mathbf H,$0

One such $\mathbf H^0=\mathbf B\odot \mathbf H,$1 is stored per grid in the digital twin; at runtime, the user’s GPS locates the corresponding grid and the matching basis is fetched (Cai et al., 7 Aug 2025).

This construction makes EB a location-indexed subspace prior rather than a generic statistical prior. Because the basis is extracted from the static environment, it is designed to characterize environmental structure that persists across users and time instants within the same grid. A common misunderstanding is to treat EB as a direct estimate of the instantaneous channel; in the reported method it instead serves as prior information that is fused with partial CSI during reconstruction.

3. Network architecture and fusion pathway

EB-P2WNet is composed of three main blocks: an EB Feature Extraction Subnetwork, an Initial Reconstruction Subnetwork, and a Dual-Input CSI Reconstruction Subnetwork; for future CSI prediction, an LSTM is appended on top of the Initial Reconstruction block (Cai et al., 7 Aug 2025).

The EB Feature Extraction Subnetwork takes $\mathbf H^0=\mathbf B\odot \mathbf H,$2, reshapes it to $\mathbf H^0=\mathbf B\odot \mathbf H,$3, and processes it through several 2D convolutions, batch normalization, LeakyReLU, and max-pooling operations to produce smaller feature maps. It then applies 1–2 Transformer encoder layers to capture long-range dependencies among the $\mathbf H^0=\mathbf B\odot \mathbf H,$4 channels and upsamples through transposed-convolution layers with batch normalization and LeakyReLU back to $\mathbf H^0=\mathbf B\odot \mathbf H,$5. The output is the learned feature map $\mathbf H^0=\mathbf B\odot \mathbf H,$6 (Cai et al., 7 Aug 2025).

The Initial Reconstruction Subnetwork receives the sparse CSI $\mathbf H^0=\mathbf B\odot \mathbf H,$7 and unrolls two iterations of a learned proximal-gradient method:

$\mathbf H^0=\mathbf B\odot \mathbf H,$8

where each $\mathbf H^0=\mathbf B\odot \mathbf H,$9 is a two-layer CNN with LeakyReLU, and $\mathbf B\in\{0,1\}^{M_t\times N_{sc}}$0 is a one-layer CNN producing a scalar step-size map. The initialization is $\mathbf B\in\{0,1\}^{M_t\times N_{sc}}$1, and after two steps the network obtains $\mathbf B\in\{0,1\}^{M_t\times N_{sc}}$2 (Cai et al., 7 Aug 2025).

The Dual-Input CSI Reconstruction Subnetwork takes the concatenation

$\mathbf B\in\{0,1\}^{M_t\times N_{sc}}$3

and processes it with residual CNN blocks of the form conv $\mathbf B\in\{0,1\}^{M_t\times N_{sc}}$4 BN $\mathbf B\in\{0,1\}^{M_t\times N_{sc}}$5 LeakyReLU $\mathbf B\in\{0,1\}^{M_t\times N_{sc}}$6 conv $\mathbf B\in\{0,1\}^{M_t\times N_{sc}}$7 BN with a skip connection. The output is $\mathbf B\in\{0,1\}^{M_t\times N_{sc}}$8. For future CSI prediction, the network uses the past $\mathbf B\in\{0,1\}^{M_t\times N_{sc}}$9 initial reconstructions $N_p$0 as input to an LSTM, which predicts $N_p$1; this is fused with the fixed $N_p$2 in the same dual-input reconstruction block to produce $N_p$3 (Cai et al., 7 Aug 2025).

The nonlinear activation throughout the architecture is LeakyReLU with negative slope $N_p$4. This layered design separates environmental feature extraction, data-consistent reconstruction from sparse pilots, and final feature fusion. This suggests that the model is explicitly structured to prevent the environmental prior from overwhelming the measurement-driven component.

4. Objectives, training protocol, and simulation setup

Training is organized in two stages. Stage 1, for present CSI only, uses the mean-squared error loss

$N_p$5

Stage 2, for joint present and future CSI fine-tuning, uses

$N_p$6

with $N_p$7 and $N_p$8 (Cai et al., 7 Aug 2025).

The data are generated by ray-tracing in Wireless InSite over a $N_p$9 urban macro scenario containing 22 buildings, with 6th-order reflections and 1st-order diffractions, at center frequency $$\mathbf h^0_{\mathrm{flat}}=\mathrm{vec}(\mathbf H^0)\in\mathbb C^{N_H},$$0, bandwidth $$\mathbf h^0_{\mathrm{flat}}=\mathrm{vec}(\mathbf H^0)\in\mathbb C^{N_H},$$1, $$\mathbf h^0_{\mathrm{flat}}=\mathrm{vec}(\mathbf H^0)\in\mathbb C^{N_H},$$2, and $$\mathbf h^0_{\mathrm{flat}}=\mathrm{vec}(\mathbf H^0)\in\mathbb C^{N_H},$$3. The coverage is divided into 1,000 grids of $$\mathbf h^0_{\mathrm{flat}}=\mathrm{vec}(\mathbf H^0)\in\mathbb C^{N_H},$$4 side, with 25 random user points per grid and four vertices times ten Doppler instants per grid used for basis extraction. Noise is added at SNR values $$\mathbf h^0_{\mathrm{flat}}=\mathrm{vec}(\mathbf H^0)\in\mathbb C^{N_H},$$5. For the future prediction task, the model uses $$\mathbf h^0_{\mathrm{flat}}=\mathrm{vec}(\mathbf H^0)\in\mathbb C^{N_H},$$6 past instants with $$\mathbf h^0_{\mathrm{flat}}=\mathrm{vec}(\mathbf H^0)\in\mathbb C^{N_H},$$7 mobility and a $$\mathbf h^0_{\mathrm{flat}}=\mathrm{vec}(\mathbf H^0)\in\mathbb C^{N_H},$$8 interval (Cai et al., 7 Aug 2025).

The reported hyperparameters are batch size 32 and Adam optimization. Learning rates are $$\mathbf h^0_{\mathrm{flat}}=\mathrm{vec}(\mathbf H^0)\in\mathbb C^{N_H},$$9 for the EB feature extraction block and $N_H=M_t\times N_{sc}$0 for the initial and dual-input reconstruction blocks over 200 epochs. For the future LSTM stage, stage 1 uses learning rate $N_H=M_t\times N_{sc}$1 for 100 epochs, followed by stage 2 fine-tuning with learning rate $N_H=M_t\times N_{sc}$2 for 50 epochs. These settings define the experimental regime under which the quantitative results are reported (Cai et al., 7 Aug 2025).

5. Quantitative performance and robustness characteristics

Performance is evaluated under SNR values $N_H=M_t\times N_{sc}$3 and pilot ratios $N_H=M_t\times N_{sc}$4. The baselines are LMMSE, which is assumed to know the full second-order statistics; IEB-PR, an ideal noise-free EB projection upper bound; EB-PR, a practical noisy EB projection using

$N_H=M_t\times N_{sc}$5

and P2WCP, which uses the same network as EB-P2WNet but without any $N_H=M_t\times N_{sc}$6-branch. The reported metrics are NMSE,

$N_H=M_t\times N_{sc}$7

cosine similarity,

$N_H=M_t\times N_{sc}$8

and average achievable rate under beamforming (Cai et al., 7 Aug 2025).

For present CSI at $N_H=M_t\times N_{sc}$9 and pilot ratio $$M:\;(\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_t$$0, EB-P2WNet yields $$M:\;(\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_t$$1 versus $$M:\;(\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_t$$2 for P2WCP, corresponding to an approximately $$M:\;(\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_t$$3 gain. Under the same conditions, the method saves 40–50% of pilots to match NMSE. The abstract states that incorporating EB provides significant benefits under low signal-to-noise ratio and pilot ratio conditions, achieving a reduction of up to 50% in pilot overhead (Cai et al., 7 Aug 2025).

In the multi-user interference experiment, two users’ CSI are mixed as

$$M:\;(\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_t$$4

EB-P2WNet remains within $$M:\;(\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_t$$5 NMSE of interference-free performance for $$M:\;(\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_t$$6 and outperforms all baselines by $$M:\;(\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_t$$7–$$M:\;(\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_t$$8 at $$M:\;(\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_t$$9. In the localization-error study, with random position errors from 1 to $$G:\;(\mathbf H^0_{t-T+1},\dots,\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_{t+1}.$$0, a $$G:\;(\mathbf H^0_{t-T+1},\dots,\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_{t+1}.$$1 error at $$G:\;(\mathbf H^0_{t-T+1},\dots,\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_{t+1}.$$2 and pilot ratio $$G:\;(\mathbf H^0_{t-T+1},\dots,\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_{t+1}.$$3 causes only a $$G:\;(\mathbf H^0_{t-T+1},\dots,\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_{t+1}.$$4 NMSE increase for EB-P2WNet, while P2WCP is unaffected because it does not use EB. EB-P2WNet tolerates approximately $$G:\;(\mathbf H^0_{t-T+1},\dots,\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_{t+1}.$$5 error before the EB priors become neutral (Cai et al., 7 Aug 2025).

The method is also tested under coarser EB granularity by merging $$G:\;(\mathbf H^0_{t-T+1},\dots,\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_{t+1}.$$6 grids into $$G:\;(\mathbf H^0_{t-T+1},\dots,\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_{t+1}.$$7 grids. In that setting, the accuracy drop is less than $$G:\;(\mathbf H^0_{t-T+1},\dots,\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_{t+1}.$$8, and the localization tolerance extends to $$G:\;(\mathbf H^0_{t-T+1},\dots,\mathbf H^0_t,\mathbf K_t)\longmapsto \widehat{\mathbf H}_{t+1}.$$9 for a $\mathbf K_t$0 NMSE penalty. For future CSI prediction under pilot ratio $\mathbf K_t$1 at $\mathbf K_t$2, EB-P2WNet achieves $\mathbf K_t$3 and $\mathbf K_t$4, compared with $\mathbf K_t$5 and $\mathbf K_t$6 for P2WCP, and completes prediction in approximately $\mathbf K_t$7 (Cai et al., 7 Aug 2025).

Setting	EB-P2WNet result	Comparator
Present CSI, $\mathbf K_t$8, pilot $\mathbf K_t$9	NMSE $N_p\ll N_{sc}$00	P2WCP: $N_p\ll N_{sc}$01
Future CSI, $N_p\ll N_{sc}$02 pilot, $N_p\ll N_{sc}$03	NMSE $N_p\ll N_{sc}$04, CS $N_p\ll N_{sc}$05	P2WCP: NMSE $N_p\ll N_{sc}$06, CS $N_p\ll N_{sc}$07
Localization error, $N_p\ll N_{sc}$08	$N_p\ll N_{sc}$09 NMSE increase	P2WCP unaffected

These results are reported as evidence that the environmental subspace prior remains useful under low SNR, colored interference, and location uncertainty, rather than only in idealized inference conditions.

6. Relation to earlier environment-aided CSI prediction

A directly relevant precursor is the environment-aware CSI prediction framework introduced in “Can Wireless Environmental Information Decrease Pilot Overhead: A CSI Prediction Example” (Shi et al., 2024). That work proposes WEI-CSIP, where scatterers around the mobile station are abstracted from environmental information using multiview images, an environmental feature map is extracted by a CNN, a deep probabilistic subsampling network acquires an optimal fixed pilot pattern, and a CNN-based channel prediction network predicts complete CSI from the environmental feature map and partial CSI (Shi et al., 2024).

In that earlier formulation, the environmental feature map $N_p\ll N_{sc}$10 is extracted from a panoramic image built from $N_p\ll N_{sc}$11 RGB camera views,

$N_p\ll N_{sc}$12

using three convolutional layers and one pooling layer. Pilot pattern optimization is performed with a binary mask $N_p\ll N_{sc}$13 having exactly $N_p\ll N_{sc}$14 ones, learned through a Gumbel-Top-k or Gumbel-Softmax relaxation. The reconstruction stage combines a small proximal iterative block with a fusion CNN so that

$N_p\ll N_{sc}$15

The reported test-set result for EB-P2WNet in that framework at pilot ratio $N_p\ll N_{sc}$16 is $N_p\ll N_{sc}$17 and cosine similarity $N_p\ll N_{sc}$18, compared with $N_p\ll N_{sc}$19 and cosine similarity $N_p\ll N_{sc}$20 for the random-sampling, no-environment baseline; the reported pilot overhead reduction is from $N_p\ll N_{sc}$21 to $N_p\ll N_{sc}$22 (Shi et al., 2024).

The later digital-twin-channel formulation differs in the form of the environmental prior. Instead of multi-view imagery and a learned environmental feature map tied to scatterer abstraction, it employs a grid-indexed environment-specific channel subspace basis extracted from a digital twin map. This suggests a methodological transition from image-conditioned environmental side information toward explicit subspace priors derived from simulated electromagnetic structure. A common misconception is that these are interchangeable representations; the available descriptions indicate that they use different sensing modalities, different prior parameterizations, and different fusion pathways.

7. Significance, interpretation, and limitations of the reported evidence

Within the reported framework, EB-P2WNet is presented as a low-overhead and robust channel prediction scheme for various complex scenarios, combining a small number of environment-extracted subspace bases ($N_p\ll N_{sc}$23) with sparse pilot measurements through a CNN+Transformer+LSTM architecture (Cai et al., 7 Aug 2025). The core empirical claims are threefold: reduction of required pilot ratio by up to 50%, robustness under low SNR and multi-user interference, and prediction of the next coherence-time CSI within approximately $N_p\ll N_{sc}$24 (Cai et al., 7 Aug 2025).

The evidence also delineates the operating assumptions under which these conclusions hold. The environmental prior is tied to a preconstructed digital twin and to grid-based localization; the method stores one basis per grid and fetches it using GPS at runtime. Consequently, robustness to localization errors is not absolute but quantified: approximately $N_p\ll N_{sc}$25 error before EB priors become neutral under the reported settings, with different behavior under coarser EB granularity (Cai et al., 7 Aug 2025). This clarifies that the method does not eliminate dependence on environmental registration; rather, it moderates its sensitivity.

A further interpretive point concerns the role of EB relative to purely learned CSI completion. Because P2WCP is defined as the same network without the $N_p\ll N_{sc}$26-branch, the comparison isolates the contribution of the environment-specific basis within the reported architecture. The measured gap between EB-P2WNet and P2WCP at low pilot ratios and low SNR suggests that the principal gain arises from introducing structured environmental priors into the reconstruction process rather than from architectural scaling alone (Cai et al., 7 Aug 2025). A plausible implication is that the framework is most advantageous when pilot scarcity and channel uncertainty are sufficiently severe that the static environmental subspace becomes a useful regularizer.