⚡ electrical engineering

Generative Site-Specific Beamforming for UPAs via Decoupled Channel Sensing

This paper proposes a cross-fused generative framework that combines decoupled azimuth-elevation channel sensing with a bidirectional cross-attention encoder and conditional normalizing flow to resolve angular ambiguity and generate high-fidelity beam candidates, achieving significant beamforming gain improvements and drastically reduced overhead compared to exhaustive 2D searches in UPA systems.

Original authors: Yao Tang, Zhaolin Wang

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yao Tang, Zhaolin Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Finding the Perfect Signal in a Crowded Room

Imagine you are at a massive, noisy concert (the wireless network). You are the User (your phone), and the Base Station is the giant speaker system on stage. To hear the music clearly, the speaker needs to focus a tight beam of sound directly at you, rather than blasting noise in every direction. This is called beamforming.

However, the room is full of walls, pillars, and other people (buildings and obstacles) that bounce the sound around. The speaker needs to figure out exactly where you are and which path the sound takes to reach you.

The Problem:
Traditionally, the speaker tries to find you by shouting in every single direction one by one (like a lighthouse spinning its light). If the room is huge and the speaker has thousands of tiny speakers (antennas), this "shouting in every direction" takes forever and wastes a lot of energy. This is the overhead problem.

The New Idea:
This paper proposes a smarter way to find you. Instead of shouting in every direction at once, the speaker does two quick, separate checks:

Horizontal Check: "Are you to the left or right?"
Vertical Check: "Are you high up or low down?"

This is much faster. But here's the catch: By checking them separately, the speaker loses the connection between the two. It knows you are "Left" and "High," but it doesn't know if you are "Left-High" or if there are two different people (one Left-Low, one Right-High). It creates a puzzle with missing pieces.

The Solution: The "Generative Detective"

The authors created a new system called GenSSBF (Generative Site-Specific Beamforming) to solve this puzzle. Think of it as a Detective who uses a special toolkit to guess the answer.

1. The "Decoupled" Clues (The Separate Checks)

The system first gathers two separate lists of clues:

Azimuth List: A list of how loud the signal is at different left/right angles.
Elevation List: A list of how loud the signal is at different up/down angles.
The Issue: These lists are "marginal." They tell you the loudness of the left side and the loudness of the top, but they don't tell you which specific corner (Left-Top) is the real target.

2. The "Cross-Fused" Brain (Connecting the Dots)

To fix the missing connection, the system uses a Cross-Attention Encoder.

Analogy: Imagine two detectives, one looking at the Left/Right clues and one looking at the Up/Down clues. They sit at a table and talk to each other. The Left/Right detective says, "I see a strong signal here," and the Up/Down detective replies, "Ah, that matches a strong signal I see there."
By sharing information, they reconstruct the hidden relationship between the horizontal and vertical angles. They figure out the "joint" picture that was lost when they checked separately.

3. The "Generative" Artist (Making Multiple Guesses)

Because the clues are still a bit fuzzy (there might be more than one possible answer), the system doesn't just pick one guess. It uses a Normalizing Flow Generator.

Analogy: Instead of a single detective pointing to one spot, imagine an artist who paints multiple possible portraits of where you might be.
The system generates a small set of "candidate beams" (a few best guesses). It's like saying, "You are likely in one of these three spots."

4. The "Lightweight" Verification (The Final Check)

Finally, the system doesn't waste time checking every possible spot in the room. It only tests the few spots the artist painted.

Analogy: The speaker sends a quick, tiny "ping" to just those three spots. You (the phone) say, "I heard the ping loudest in Spot #2!"
The speaker then locks its beam onto Spot #2.

Why This is a Big Deal

The paper claims this method is a huge improvement over the old ways:

Speed & Efficiency: It reduces the time and energy needed to find the signal by 93.8%. Instead of checking 1,024 directions (like checking every seat in a stadium), it checks a few dozen and then uses its "brain" to guess the rest.
Better Accuracy: Even though it checks fewer spots, it finds the best spot more often than the old method. In their tests, it improved signal strength by 38% to 83% compared to the traditional "check everything" method.
Handling Confusion: It is specifically designed to handle the confusion caused by checking angles separately. The "Cross-Attention" part is the key that unlocks the missing information.

Summary in One Sentence

This paper teaches a wireless system to find your phone much faster by checking horizontal and vertical angles separately, then using a smart AI "detective" to connect those clues and generate a few high-quality guesses, which are quickly verified to lock in the perfect signal.

Technical Summary: Generative Site-Specific Beamforming for UPAs via Decoupled Channel Sensing

Problem Statement

In Uniform Planar Array (UPA) systems, particularly for mmWave communications, accurate beam alignment is critical for achieving high spectral efficiency and link reliability. However, traditional beam alignment methods face a fundamental tradeoff between beamforming fidelity and online sensing overhead:

Channel-dependent beamforming requires accurate Channel State Information (CSI), leading to prohibitive pilot training and feedback overhead in large-scale antenna systems.
Grid-of-beams sweeping (e.g., exhaustive 2D DFT codebook search) avoids explicit CSI acquisition but incurs multiplicative complexity ( $M_\phi \times M_\theta$ ) as the number of probing beams increases with angular resolution.
Site-Specific Beamforming (SSBF) attempts to learn environmental regularities to infer beams from limited measurements. However, existing generative SSBF approaches often rely on full 2D sensing or generic codebook measurements, failing to address the specific challenges of decoupled sensing in UPAs.

When azimuth and elevation domains are probed independently to reduce overhead to linear complexity ( $M_\phi + M_\theta$ ), the resulting Reference Signal Received Power (RSRP) observations suffer from:

Information Loss: The explicit joint azimuth-elevation coupling of the channel is lost; only marginal power distributions are preserved.
Lobe Mismatch: Strongest azimuth and elevation lobes may not correspond to the same physical path, leading to virtual angular directions that do not exist.
Phase Ambiguity: Power-only RSRP feedback discards phase information, making the mapping from separate observations to a beamformer non-unique and highly ambiguous.

Methodology

The paper proposes a Cross-Fused Generative Site-Specific Beamforming (GenSSBF) framework designed to generate high-fidelity beam candidates from low-overhead, decoupled sensing.

1. Decoupled Channel Sensing

The Base Station (BS) independently probes the azimuth and elevation domains using separate codebooks.

Azimuth Sensing: Probes different azimuth sectors with broad elevation coverage.
Elevation Sensing: Probes different elevation sectors with broad azimuth coverage.
Output: Two low-dimensional RSRP vectors, $r_\phi$ and $r_\theta$ , reducing online overhead from multiplicative to linear complexity.

2. Cross-Fused GenSSBF Architecture

To recover the lost joint angular information, the framework employs a generative model with three core components:

Angular-Domain Channel Encoding: The target beamformer is represented in the angular domain (DFT basis). To avoid phase-wrapping issues during training, complex channel coefficients are encoded as amplitude and a cosine-sine phase pair ( $\rho, \cos\phi, \sin\phi$ ).
Bidirectional Cross-Attention Encoder: This module addresses the information loss of decoupled sensing.
- Separate MLPs encode $r_\phi$ and $r_\theta$ into feature tokens.
- A bidirectional cross-attention mechanism allows the azimuth branch to query elevation features (and vice versa) to extract latent dependencies.
- The refined features are fused into a single condition vector $c$ that summarizes the site-specific 2D angular structure.
Conditional Normalizing Flow Generator:
- Instead of a deterministic predictor (which would average out multiple plausible solutions), a conditional normalizing flow maps a simple Gaussian latent variable $z$ to a complex distribution of beam candidates conditioned on $c$ .
- This stochastic generator produces a set of $K$ candidate beamformers, handling the multimodal nature of the mapping from decoupled RSRP to beam directions.

3. Task-Oriented Training and Verification

Training Objective: The model is trained using a normalized beamforming gain loss. Rather than fitting the entire conditional distribution (likelihood-based), the objective maximizes the gain of the best candidate among the $K$ generated beams. This encourages the generator to assign probability mass to high-gain regions.
Online Verification: The BS generates $K$ candidate beams. The User Equipment (UE) performs lightweight pilot measurements on these candidates and feeds back the index of the beam with the highest RSRP for final data transmission.

Key Contributions

Low-Overhead Decoupled Sensing: The paper formalizes a decoupled sensing strategy for UPAs that reduces online overhead from $O(M_\phi M_\theta)$ to $O(M_\phi + M_\theta)$ , while explicitly characterizing the resulting loss of joint angular coupling.
Cross-Fused GenSSBF Framework: A novel architecture combining a bidirectional cross-attention encoder to recover latent azimuth-elevation dependencies and a conditional normalizing flow to generate multiple high-fidelity beam candidates from ambiguous inputs.
Task-Oriented Training Loss: A training objective that optimizes for the inclusion of at least one high-gain beam in the candidate set, rather than minimizing average reconstruction error. The paper provides a theoretical interpretation showing that success probability increases with the generator's probability mass in the high-gain region and the number of candidates.
Performance Validation: Extensive simulations using DeepMIMO scenarios (indoor, outdoor blockage, and urban mmWave) demonstrate the framework's efficacy.

Results

Simulation results based on DeepMIMO scenarios (I2 28, O1B 28, Boston5G 28) show:

Beamforming Gain: The proposed GenSSBF (with $K=8$ $K = 8$ candidates) consistently outperforms deterministic MLP-based prediction, GenSSBF without cross-attention, and conventional exhaustive DFT codebook search.
- Compared to a full 1024-beam 2D DFT search, the proposed method achieved normalized beamforming gain improvements of 83.6% (I2 28), 74.6% (O1B 28), and 38.1% (Boston5G 28).
- It achieves beam patterns significantly closer to the Maximum Ratio Transmission (MRT) upper bound than fixed codebook searches.
Overhead Reduction: The decoupled sensing design reduces the sweeping overhead by 93.8% compared to full 2D DFT sweeping.
Robustness: The framework maintains superior performance over a wide range of Signal-to-Noise Ratios (SNR) compared to benchmarks, though performance degrades at very low SNR (< -5 dB) where sensing noise overwhelms the signal.
Ablation Studies: Removing the cross-attention module significantly degrades performance, confirming its necessity for recovering latent dependencies. Stochastic generation (GenSSBF) outperforms deterministic prediction (MLP), validating the need for multimodal candidate generation in ambiguous sensing scenarios.

Significance

The paper claims that reducing sensing overhead alone is insufficient for reliable UPA beam alignment due to the inherent ambiguity of decoupled observations. The significance of this work lies in demonstrating that generative artificial intelligence, specifically when combined with cross-domain fusion and task-oriented stochastic generation, can effectively infer high-quality, site-specific beamformers from low-dimensional, marginal power measurements. This approach offers a viable path to achieving near-optimal beamforming gains with drastically reduced online sensing costs, addressing a critical bottleneck in the deployment of large-scale UPA systems in 5G/6G networks.