Uplink-Downlink CSI Fusion Precoding Network

• The uplink-downlink CSI fusion precoding network is a method that fuses UL-estimated channel geometry with minimal DL feedback to accurately reconstruct downlink CSI in FDD massive MIMO systems.
• It leverages a variety of techniques—from analytical MMSE estimators and Bayesian fusion to end-to-end neural architectures—to synthesize robust precoders even under limited feedback scenarios.
• Key implementations focus on optimizing bit-allocation, reducing reconstruction errors, and enhancing spectral efficiency by balancing uplink-derived features with the residual DL uncertainties.

An uplink-downlink CSI fusion precoding network is a precoding architecture that derives downlink transmit decisions from a fusion of complementary channel information obtained on the uplink and the downlink, rather than from either source alone. The design problem is especially acute in FDD massive MIMO, where strict UL/DL reciprocity is unavailable and explicit downlink CSI feedback becomes prohibitive as antenna count and bandwidth grow. Representative solutions therefore combine uplink-derived geometry, delay structure, or latent channel features with limited downlink phase feedback, predicted CSI, codebook precoders, or quantized CSI-RS observations to reconstruct downlink CSI or directly synthesize robust precoders (Kim et al., 2024, Deutschmann et al., 22 Mar 2025, Wu et al., 9 Sep 2025).

1. Research context and problem formulation

The central motivation is the CSI acquisition bottleneck in large-array systems. In conventional FDD operation, the base station must probe the downlink, the user must estimate the channel, compress it, and feed back a representation suitable for precoding. This pipeline is expensive in feedback bandwidth, UE power, and latency. Several recent lines of work replace or relax full downlink feedback by exploiting structure shared between UL and DL channels: partial reciprocity of geometric parameters, common delay profiles, scene geometry, or learned latent environmental variables (Kim et al., 2024, Lin et al., 2024, Han et al., 13 Jan 2025, Euchner et al., 2022).

Within this literature, at least three design patterns recur. One pattern reconstructs the downlink channel analytically from UL-estimated path parameters plus a very small amount of downlink-side side information, such as quantized per-path phases (Kim et al., 2024). A second pattern performs probabilistic fusion between measured CSI and geometry-predicted CSI, typically in an MMSE or Bayesian form, to improve robustness under low SNR, mobility, or blockage (Deutschmann et al., 22 Mar 2025). A third pattern learns the fusion rule itself, either as a plug-in module that refines codebook precoders using UL CSI (Lin et al., 2024), or as an end-to-end neural architecture that jointly designs CSI-RS, feedback, and BS-side precoding (Wu et al., 9 Sep 2025).

A broader adjacent tradition uses partial-CSI uplink-downlink duality rather than explicit CSI fusion. In a user-centric cell-free setting, uplink combining vectors are reused as downlink precoders, and downlink powers are computed from nominal uplink SINRs under partial CSI (Göttsch et al., 2022). This is not a neural fusion network, but it establishes the same overarching principle: the uplink can supply structural information sufficient to simplify downlink precoder design.

2. Channel representations and reciprocity assumptions

Most fusion formulations depend on a structured channel model. In path-based FDD massive MIMO, the uplink and downlink channels for user $k$ are written as sums of $L_k$ dominant paths,

$$\mathbf{h}_k^{\sf UL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf UL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf UL}, \lambda^{\sf UL}\right), \qquad \mathbf{h}_k^{\sf DL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf DL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf DL}, \lambda^{\sf DL}\right),$$

with array response $\mathbf a(\cdot)$, path angles, and complex path gains (Kim et al., 2024). The essential partial-reciprocity assumptions are that AoA/AoD are frequency-invariant,

$$\theta_{k,\ell}^{\sf UL} = \theta_{k,\ell}^{\sf DL} \triangleq \theta_{k,\ell},$$

and that path-gain magnitudes are quasi-reciprocal,

$$|g_{k,\ell}^{\sf UL}| \approx |g_{k,\ell}^{\sf DL}| \triangleq \beta_{k,\ell}.$$

Under this model, the BS can infer $\theta_{k,\ell}$ and $\beta_{k,\ell}$ from UL pilots, while the non-reciprocal downlink phase remains the missing variable to be fed back or statistically modeled (Kim et al., 2024).

A closely related no-feedback formulation assumes that path count, angle, and delay are frequency invariant,

$$L_k^{\mathrm{dl}} = L_k^{\mathrm{ul}} \triangleq L_k,\qquad \theta_{k,\ell}^{\mathrm{dl}} = \theta_{k,\ell}^{\mathrm{ul}},\qquad \tau_{k,\ell}^{\mathrm{dl}} = \tau_{k,\ell}^{\mathrm{ul}},$$

but explicitly models path-gain mismatch through

$$\alpha_{k,\ell}^{\mathrm{dl}} = \eta_{k,\ell}\alpha_{k,\ell}^{\mathrm{ul}} + \sqrt{1-\eta_{k,\ell}^2}\,g, \qquad g\sim\mathcal{CN}(0,\sigma_{\text{path},k}^2),$$

so perfect CSIT recovery is excluded unless $L_k$0 (Han et al., 13 Jan 2025).

Other fusion frameworks elevate the representation from path parameters to scene geometry. A geometry-based model decomposes the channel into LoS and specular NLoS components, parameterized by a hidden state

$L_k$1

where $L_k$2 contains user position and velocity and $L_k$3 encodes a master virtual anchor representation of the environment map (Deutschmann et al., 22 Mar 2025). In this view, CSI prediction becomes forward propagation through an inferred radio scene rather than interpolation in channel space.

Neural UL-to-DL mapping work adopts a more abstract latent-state viewpoint. The uplink CSI $L_k$4 and downlink channel $L_k$5 are treated as deterministic functions of an underlying environmental state $L_k$6, so the desired mapping is $L_k$7 if $L_k$8 is bijective on the relevant domain (Euchner et al., 2022). In FR3 joint acquisition, the channel is further recast in angle-delay or spatial-frequency tensors so that incomplete UL SRS and DL CSI-RS observations can be aligned and fused by attention or state-space models (He et al., 2 Dec 2025).

3. Fusion mechanisms and representative architectures

The common design objective is to let the uplink provide the structure and the downlink provide the missing, non-reciprocal, or higher-resolution component. The resulting architectures differ mainly in what is fused and where the fusion occurs.

Framework	Fused information	Precoding output/use
Path-phase MMSE fusion	UL AoAs and path gains + quantized DL per-path phase	Reconstructed DL CSI for robust GPIP precoding
Geometry-based channel fusion	Measured CSI + geometry-predicted CSI	Fused CSI for robust beamforming
SRPNet	SB-level codebook precoders + UL delay profile/PDP	RB-level refined precoders
CSC-SA-Net fusion precoding	Quantized CSI-RS feedback branch + SRS-only branch	Final fused precoder
HASCAN/JUDCEN	Feedback-reconstructed DL feature + UL feature	Full CSI for EZF and prediction

In the analytical FDD massive-MIMO formulation, the UE feeds back only the phase of each dominant downlink path. The BS already knows the geometric steering basis from uplink pilots, so the feedback scales with the number of dominant paths rather than with the number of antennas (Kim et al., 2024). This sharply departs from conventional codebook-index or whole-channel feedback.

In geometry-based channel fusion, the uplink is used not merely to estimate CSI but to infer the user position and environment map through a Bayes filter and particle filter. The inferred geometry then yields predicted downlink CSI, which is fused with measured CSI in a Bayesian/MMSE manner (Deutschmann et al., 22 Mar 2025). This suggests a shift from reciprocity-based beamforming toward model-based scene inference.

SRPNet introduces a gNodeB-side plug-in fusion module for cellular FDD systems. Its inputs are SB-level Type II or eType II codebook precoders and UL CSI-derived delay-profile information. The architecture contains an initial precoder upsampling module, a BPF design module using UL CSI delay profile/PDP, and a precoder refinement module with convolutional refinement in beam-domain and angle/frequency-domain (Lin et al., 2024).

The most literal “uplink-downlink CSI fusion precoding network” appears in an end-to-end neural architecture in which two BS-side candidate precoders are formed: a feedback-only precoding network driven by quantized downlink observations, and an SRS-only precoding network driven by uplink SRS. These are then combined by a fusion precoding network. The same framework jointly learns downlink CSI-RS design, UE-side compression/quantization, feedback decoding, and precoding under a spectral-efficiency-oriented loss (Wu et al., 9 Sep 2025).

In FR3 TDD, JUDCEN plays an analogous fusion role at the CSI acquisition level. It shape-matches a feedback-reconstructed DL feature with a UL feature, concatenates them, and processes the result through spatial-frequency attention blocks. The fused CSI is then passed to a Mamba-based channel prediction network for future-slot CSI prediction (He et al., 2 Dec 2025).

4. Estimation theory, reconstruction, and uncertainty handling

The analytical benchmark for CSI fusion is the MSE-optimal downlink reconstruction in which quantized per-path downlink phases are combined with UL-extracted path angles and magnitudes. The resulting estimator is

$L_k$9

with phase-error compensation factor

$$\mathbf{h}_k^{\sf UL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf UL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf UL}, \lambda^{\sf UL}\right), \qquad \mathbf{h}_k^{\sf DL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf DL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf DL}, \lambda^{\sf DL}\right),$$0

The corresponding per-user MMSE is

$$\mathbf{h}_k^{\sf UL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf UL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf UL}, \lambda^{\sf UL}\right), \qquad \mathbf{h}_k^{\sf DL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf DL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf DL}, \lambda^{\sf DL}\right),$$1

and the normalized per-path MSE is strictly decreasing in $$\mathbf{h}_k^{\sf UL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf UL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf UL}, \lambda^{\sf UL}\right), \qquad \mathbf{h}_k^{\sf DL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf DL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf DL}, \lambda^{\sf DL}\right),$$2 and convex after piecewise-linear interpolation. This leads to an incremental bit-allocation rule that assigns each additional phase bit to the path with the largest marginal MSE reduction, rather than using uniform allocation (Kim et al., 2024).

A companion object is the reconstruction-error covariance,

$$\mathbf{h}_k^{\sf UL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf UL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf UL}, \lambda^{\sf UL}\right), \qquad \mathbf{h}_k^{\sf DL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf DL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf DL}, \lambda^{\sf DL}\right),$$3

which is explicitly fed into robust precoding (Kim et al., 2024). The same uncertainty-aware philosophy appears in no-feedback FDD RSMA, where the downlink CSIT error $$\mathbf{h}_k^{\sf UL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf UL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf UL}, \lambda^{\sf UL}\right), \qquad \mathbf{h}_k^{\sf DL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf DL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf DL}, \lambda^{\sf DL}\right),$$4 is modeled through $$\mathbf{h}_k^{\sf UL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf UL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf UL}, \lambda^{\sf UL}\right), \qquad \mathbf{h}_k^{\sf DL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf DL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf DL}, \lambda^{\sf DL}\right),$$5. Because direct evaluation is difficult after nonlinear 2D-NOMP reconstruction, the covariance is approximated using a CRLB-inspired observed Fisher information construction, and, for $$\mathbf{h}_k^{\sf UL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf UL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf UL}, \lambda^{\sf UL}\right), \qquad \mathbf{h}_k^{\sf DL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf DL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf DL}, \lambda^{\sf DL}\right),$$6, an additional identity term captures gain decorrelation (Han et al., 13 Jan 2025).

Bayesian geometry-based fusion makes the same MMSE principle explicit. With predicted prior mean $$\mathbf{h}_k^{\sf UL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf UL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf UL}, \lambda^{\sf UL}\right), \qquad \mathbf{h}_k^{\sf DL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf DL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf DL}, \lambda^{\sf DL}\right),$$7, prediction covariance $$\mathbf{h}_k^{\sf UL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf UL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf UL}, \lambda^{\sf UL}\right), \qquad \mathbf{h}_k^{\sf DL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf DL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf DL}, \lambda^{\sf DL}\right),$$8, and measurement covariance $$\mathbf{h}_k^{\sf UL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf UL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf UL}, \lambda^{\sf UL}\right), \qquad \mathbf{h}_k^{\sf DL} = \sum_{\ell=1}^{L_k} g_{k,\ell}^{\sf DL}\,\mathbf{a}\!\left(\theta_{k,\ell}^{\sf DL}, \lambda^{\sf DL}\right),$$9, the fused CSI can be written in LMMSE form as

$\mathbf a(\cdot)$0

When measured CSI is reliable it dominates; when it is noisy or stale, the model-based prediction dominates (Deutschmann et al., 22 Mar 2025).

These formulations show that “fusion” is not merely feature concatenation. In the strongest versions, it is an uncertainty-weighted estimator whose posterior mean, covariance, and induced error geometry are all propagated into the downstream precoder.

5. Precoder synthesis, objectives, and learning paradigms

Once CSI has been fused or reconstructed, precoding can be posed as a robust optimization problem. In the path-phase FDD formulation, the BS maximizes a lower bound on the sum rate that depends on

$\mathbf a(\cdot)$1

and solves the resulting NP-hard problem with Generalized Power Iteration Precoding (GPIP). The GPIP update has the generalized-eigenvector form

$\mathbf a(\cdot)$2

followed by normalization. This jointly addresses user selection, power allocation, and beamforming under channel-reconstruction error (Kim et al., 2024).

Learned precoding works often optimize the communication objective directly rather than an MSE proxy. In UL-to-DL precoding prediction, the target metric is normalized received power,

$\mathbf a(\cdot)$3

and all neural models are trained with loss $\mathbf a(\cdot)$4, aligning learning with beamforming gain rather than coefficient error (Euchner et al., 2022).

SRPNet formulates SB-to-RB precoder reconstruction as learning a nonlinear mapping $\mathbf a(\cdot)$5 that maximizes beamforming gain across RBs,

$\mathbf a(\cdot)$6

with UL CSI used only as auxiliary information for bandpass-filter design and alias suppression (Lin et al., 2024).

In end-to-end learned massive MIMO, the physical-layer subproblem is also trained by a spectral-efficiency objective. The CSI-RS design network, UE-side CSI semantic extractor, BS-side CSI semantic decoder, and fusion precoding network are pretrained with

$\mathbf a(\cdot)$7

where $\mathbf a(\cdot)$8 is the summed spectral efficiency across users, subcarriers, and symbols; only after this stage are the modules jointly adapted with the downstream task loss (Wu et al., 9 Sep 2025). In FR3 joint acquisition, the reconstructed and predicted CSI are evaluated for downlink transmission using eigen zero-forcing, and cosine-similarity training is preferred because channel direction matters more directly for precoding than raw MSE (He et al., 2 Dec 2025).

A non-neural limiting case is the partial-CSI duality formulation for cell-free networks, where downlink precoders are simply set equal to uplink combining vectors, $\mathbf a(\cdot)$9, and power allocation follows from a nominal UL/DL SINR mapping computable from partial CSI and large-scale fading (Göttsch et al., 2022). This illustrates that a “fusion precoding” viewpoint can be instantiated either as explicit estimation plus robust optimization or as a structural reuse of uplink-side signal processing.

6. Empirical behavior, limitations, and recurring misconceptions

A consistent empirical result is that modest downlink-side supplementation can dramatically improve uplink-derived precoding. In path-phase FDD massive MIMO, using only UL-derived geometry degrades as path count increases or UL-DL frequency separation grows, but adding only a few bits per path of downlink phase feedback substantially improves reconstructed CSI and yields sum-spectral efficiency close to WMMSE with perfect CSI when combined with MSE-optimal reconstruction, error covariance, and robust GPIP precoding (Kim et al., 2024).

Geometry-based fusion is particularly effective when direct reciprocity is weak. On measured hallway data, reciprocity-based beamforming under low SNR suffers a loss of at least $$\theta_{k,\ell}^{\sf UL} = \theta_{k,\ell}^{\sf DL} \triangleq \theta_{k,\ell},$$0 dB relative to perfect CSI at $$\theta_{k,\ell}^{\sf UL} = \theta_{k,\ell}^{\sf DL} \triangleq \theta_{k,\ell},$$1 dB, whereas the geometry-based beamformer using predicted CSI outperforms reciprocity-based beamforming by about $$\theta_{k,\ell}^{\sf UL} = \theta_{k,\ell}^{\sf DL} \triangleq \theta_{k,\ell},$$2 dB on average at the chosen low SNR. The same framework also reports horizontal position RMSE $$\theta_{k,\ell}^{\sf UL} = \theta_{k,\ell}^{\sf DL} \triangleq \theta_{k,\ell},$$3 cm and vertical position RMSE $$\theta_{k,\ell}^{\sf UL} = \theta_{k,\ell}^{\sf DL} \triangleq \theta_{k,\ell},$$4 cm with the stochastic likelihood, while an LoS-only model degrades to $$\theta_{k,\ell}^{\sf UL} = \theta_{k,\ell}^{\sf DL} \triangleq \theta_{k,\ell},$$5 cm and $$\theta_{k,\ell}^{\sf UL} = \theta_{k,\ell}^{\sf DL} \triangleq \theta_{k,\ell},$$6 cm, respectively (Deutschmann et al., 22 Mar 2025).

Plug-in learned fusion is effective but not free. SRPNet improves Type II and eType II RB-level precoder quality most strongly for high delay-spread channels, and both threshold-based and learning-based PDP switches outperform random switching; maximum excess delay is the most effective simple rule, while the learning-based switch gives the best overall performance and lowest complexity. At the same time, SRPNet is reported to be roughly $$\theta_{k,\ell}^{\sf UL} = \theta_{k,\ell}^{\sf DL} \triangleq \theta_{k,\ell},$$7 more complex than linear interpolation, which motivates conditional activation (Lin et al., 2024).

A recurrent misconception is that strong random-split performance guarantees deployable UL-to-DL generalization. On indoor measurements, a plain DNN reaches about $$\theta_{k,\ell}^{\sf UL} = \theta_{k,\ell}^{\sf DL} \triangleq \theta_{k,\ell},$$8 dB mean power under a random 50/50 split, but under a $$\theta_{k,\ell}^{\sf UL} = \theta_{k,\ell}^{\sf DL} \triangleq \theta_{k,\ell},$$9 m checkerboard split the same model shows about $$|g_{k,\ell}^{\sf UL}| \approx |g_{k,\ell}^{\sf DL}| \triangleq \beta_{k,\ell}.$$0 dB on seen regions and about $$|g_{k,\ell}^{\sf UL}| \approx |g_{k,\ell}^{\sf DL}| \triangleq \beta_{k,\ell}.$$1 dB on unseen regions, revealing substantial spatial overfitting (Euchner et al., 2022). Another misconception is that fusion always eliminates feedback. Some frameworks remove direct CSIT feedback altogether and compensate with RSMA plus covariance-aware robust optimization, achieving $$|g_{k,\ell}^{\sf UL}| \approx |g_{k,\ell}^{\sf DL}| \triangleq \beta_{k,\ell}.$$2 higher minimum spectral efficiency than WMMSE, $$|g_{k,\ell}^{\sf UL}| \approx |g_{k,\ell}^{\sf DL}| \triangleq \beta_{k,\ell}.$$3 higher than the same framework without error covariance approximation, and about $$|g_{k,\ell}^{\sf UL}| \approx |g_{k,\ell}^{\sf DL}| \triangleq \beta_{k,\ell}.$$4 ms latency reduction at the $$|g_{k,\ell}^{\sf UL}| \approx |g_{k,\ell}^{\sf DL}| \triangleq \beta_{k,\ell}.$$5th percentile for a $$|g_{k,\ell}^{\sf UL}| \approx |g_{k,\ell}^{\sf DL}| \triangleq \beta_{k,\ell}.$$6-bit payload; others still rely on limited but structured feedback, such as per-path phase, compressed CSI-RS features, or codebook precoders (Han et al., 13 Jan 2025).

In FR3 joint UL/DL acquisition, the benefit of explicit fusion remains pronounced even in TDD when pilot ports are insufficient for full probing. HASCAN improves NMSE over DACEN by $$|g_{k,\ell}^{\sf UL}| \approx |g_{k,\ell}^{\sf DL}| \triangleq \beta_{k,\ell}.$$7 dB and over KDD-SFCEN + TUDCEN by $$|g_{k,\ell}^{\sf UL}| \approx |g_{k,\ell}^{\sf DL}| \triangleq \beta_{k,\ell}.$$8 dB at $$|g_{k,\ell}^{\sf UL}| \approx |g_{k,\ell}^{\sf DL}| \triangleq \beta_{k,\ell}.$$9 dB, while the cosine-similarity-trained variant improves spectral efficiency by $\theta_{k,\ell}$0 over the MSE-trained variant with no extra computational cost; the Mamba-based prediction stage also reduces FLOPs by about $\theta_{k,\ell}$1 relative to Transformer-based baselines (He et al., 2 Dec 2025). In the end-to-end fusion-precoding setting, separated methods degrade sharply when CSI-RS is limited to $\theta_{k,\ell}$2, whereas the jointly trained design learns to compensate for time-varying channel mismatch and resource scarcity, with non-orthogonal transmission outperforming orthogonal transmission when resources are tight (Wu et al., 9 Sep 2025).

Taken together, these results indicate that uplink-downlink CSI fusion precoding is best understood not as a single algorithm but as a design doctrine: extract the component of downlink structure that the uplink reveals reliably, identify the residual uncertainty that the uplink cannot resolve, and fuse a minimally sufficient downlink-side signal—feedback, predicted geometry, CSI-RS observations, or learned semantic features—into a precoder that is explicitly robust to the remaining error.