Programmable Unitary RF Networks for Narrowband Hybrid Beamforming

arXiv:2603.17616 · eess.SP · Submitted 2026-03-18 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Programmable Unitary RF Networks for Narrowband Hybrid Beamforming".

Dev: Here's the gist of the paper. Hybrid beamforming via programmable unitary RF networks proposes an architecture based on a programmable unitary RF network that achieves power-preserving RF transport from the RF-chains to the antenna ports under ideal matched hardware models.

Rosa: The Core Problem Addressed Conventional hybrid beamforming architectures are often compared with one another and with the fully-digital architecture under the same radiated antenna power.

Dev: This issue is especially pronounced for fully-connected splitter–phase-shifter–combiner networks, whose physical power transfer remains contractive even under ideal passive-component assumptions.

Rosa: The paper addresses the physically relevant budget, which is the power injected by the RF-chain outputs into the passive analog RF network, which is then usually transferred to the antenna ports in a contractive (lossy) manner.

Title and authors: Rosa: So, we're looking at this paper called "Programmable Unitary RF Networks for Narrowband Hybrid Beamforming," and it’s about how we can make the analog part of our beamformer actually move power from where it starts to where the antenna ends without losing anything.

Dev: That's right, Rosa. The core idea is that most current hybrid setups suffer because the power transfer through those passive components is contractive, even when you look at ideal parts. This paper tackles that by proposing a hybrid beamforming architecture based on a programmable unitary RF network <ref:2603.17616#pg1>.

Taro: So it’s not just about getting the same total power out, but about ensuring the injected power actually makes it to the antenna ports without loss under ideal conditions <ref:2603.17616#pg2>.

Rosa: Exactly. They build this analog network as an interlaced mixer-phase architecture, mixing fixed layers with programmable diagonal phase-shifting layers, and they claim that under those ideal passive component assumptions, all the power injected by the RF chain reaches the antenna ports without loss <ref:2603.17616#pg1>.

Dev: That’s a big claim because usually, when you compare hybrid systems to fully-digital ones, you normalize everything to equal radiated antenna power, but that ignores how much of the injected power actually gets transferred <ref:2603.17616#pg3>.

Taro: So the paper focuses on the RF transport efficiency first, suggesting that maximum transfer happens when the underlying N by N analog RF network is unitary <ref:2603.17616#pg3>.

Rosa: And they argue that an analog beamformer that can conduct this semi-unitary transfer of power is only half the story; it also has to be fully expressive, meaning you can program it to realize any r rank semi-unitary matrix <ref:2603.17616#pg3>.

Dev: Which brings us to their main proposal: they derive a closed form digital beamformer and a low complexity programming method for that analog beamformer <ref:2603.17616#pg1>.

Taro: What does that mean for the hardware side? It suggests you can design an AI system where the analog part is programmed via phase shifters to achieve the desired subspace <ref:2603.17616#pg3>.

Rosa: And they found a way to program it by maximizing this objective function, Φ star equals arg max Φ∥F HtarFRF(Φ)∥two F <ref:2603.17616#pg3>. That means the programming is optimized for this power preservation goal.

Dev: It’s interesting because they also gave a guideline for when this subspace containment becomes plausible, which is a stream aware necessary guideline in the depth analysis <ref:2603.17616#pg3>.

Taro: So it gives us some concrete numbers on how many layers we might need, suggesting that the required number of programmable layers M must be greater than or equal to two times S times N minus r, which is M >= 2S(N - r) <ref:2603.17616#pg3>.

Rosa: That points toward a practical hardware design constraint because it tells you how deep you need to go for exact recovery <ref:2603.17616#pg3>.

Title and authors: Dev: And this brings up the issue of quantization, which they address by saying that phase quantization actually preserves RF efficiency because every quantized diagonal layer Dk(ϕ(q)k) remains unitary <ref:2603.17616#pg3>.

Taro: So even if you use lower resolution digital hardware for the phase shifters, the semi-unitary nature of the analog processor stays intact and the power normalization is maintained <ref:2603.17616#pg3>.

Rosa: That’s significant because it means AI systems can actually use less precise hardware without sacrificing power efficiency, which is a major practical gain <ref:2603.17616#pg3>.

Dev: It really highlights the contrast with passive fully connected baselines, where their RF transfer is contractive <ref:2603.17616#pg3>. So the paper shows that this architecture can provide significant gains because its RF transport remains semi-unitary under equal injected power <ref:2603.17616#pg2>.

Taro: So for someone just listening to the show, what does this mean is that we can get performance close to a fully digital precoder while beating baseline hybrid systems when we only inject the same amount of power <ref:2603.17616#pg3>.

Rosa: It means the proposed architecture closely matches the fully-digital precoder, and exact hybrid recovery happens whenever you program that analog subspace to contain your target subspace <ref:2603.17616#pg3>.

Dev: So we are looking at a way to get high spectral efficiency by using unitary analog processing instead of the contractive transfer found in passive fully connected networks <ref:2603.17616#pg3>.

Taro: The implication for real-world failure modes is that if the system is designed around this semi-unitary transport, it has a better safety case when the world misbehaves <ref:2603.17616#pg3>.

Rosa: So to wrap up this paper, the key thing to remember about "Programmable Unitary RF Networks for Narrowband Hybrid Beamforming" is that it achieves power-preserving RF transport from the RF chains to the antenna ports under ideal matched hardware models <ref:2603.17616#pg1>.

Dev: It shows that by structuring the analog network correctly, we can get a hybrid precoder that is nearly as good as a fully-digital one and still perform better than the standard hybrid architectures when you are limited to equal injected power <ref:2603.17616#pg3>.

Taro: The future work I see is applying these depth guidelines to more complex, non-ideal channel environments where the world doesn't behave perfectly <ref:2603.17616#pg3>.

Rosa: Yeah, so the paper suggests that for exact recovery, you need enough layers M greater than two times S times N minus r <ref:2603.17616#pg3>. That's a concrete design hint for hardware engineers.

Dev: It really shows that the physical budget is more important than just the radiated power normalization when designing hybrid systems <ref:2603.17616#pg1>.

Taro: So, it’s about moving past the contractive limitations of passive networks by using a programmable unitary structure to ensure power stays where it needs to go <ref:2603.17616#pg3>.

Rosa: That's what we have with this paper on Programmable Unitary RF Networks for Narrowband Hybrid Beamforming, and that’s where we are for now <ref:2603.17616#pg1>.

The paper's summary: Rosa: So, to recap, this paper is about taking those hybrid beamformers and making sure that every bit of power from the RF chain actually gets to the antenna without any loss under ideal conditions <ref:2603.17616#pg1>.

Dev: Exactly. They show that conventional hybrid setups lose power because they aren't truly unitary in their passive components, which is a big problem for real systems <ref:2603.17616#pg3>.

Taro: The paper proposes this new architecture uses programmable unitary RF networks, meaning the analog part can be steered to maintain that power transfer efficiency <ref:2603.17616#pg2>.

Rosa: They build this analog network by interleaving fixed layers with programmable diagonal phase-shifting layers <ref:2603.17616#pg1>. It’s like you have some basic mixing parts, but you can program the phase shifts to create any desired subspace <ref:2603.17616#pg3>.

Dev: And they claim that under ideal hardware models, this setup preserves injected power exactly at the antenna ports <ref:2603.17616#pg2>. That's the core finding for a controls engineer, it means lower latency because you don't lose signal strength in the analog domain <ref:2603.17616#pg2>.

Taro: But there’s a catch. They have to program that analog beamformer correctly, and they give us this objective function Φ star to maximize for that programming <ref:2603.17616#pg3>. That tells the AI system exactly what settings the phase shifters need to achieve peak power <ref:2603.17616#pg3>.

Rosa: And they also give a guideline for how deep that network needs to be, which is stream-aware <ref:2603.17616#pg3>. It suggests the number of layers M has to be about two times S times N minus r <ref:2603.17616#pg3>.

Dev: That's a practical constraint, Rosa. For someone working on real-time loops, that tells you how much hardware you need before the performance starts dropping off significantly <ref:2603.17616#pg3>.

Taro: And they also address quantization by saying phase quantization doesn't hurt the power normalization because every quantized layer stays unitary <ref:2603.17616#pg3>. So you can use lower resolution hardware without losing that fundamental power preservation property <ref:2603.17616#pg3>.

Rosa: It means we can build these systems in a way that is more robust to the imperfections of real-world components <ref:2603.17616#pg3>.

Dev: So, it’s about moving away from those passive networks where the RF transport is physically contractive and towards something semi-unitary instead <ref:2603.17616#pg2>.

Taro: The big implication for autonomy is that if you program the analog subspace to contain the target subspace, you get exact recovery <ref:2603.17616#pg3>. That’s a solid safety guarantee for autonomous systems <ref:2603.17616#pg3>.

Rosa: Right. This paper shows how we can get performance close to a fully digital precoder while beating baseline hybrid systems when you only inject the same amount of power <ref:2603.17616#pg3>.

Dev: It really puts the budget issue—the physical injection power versus actual antenna output—front and center in hybrid design <ref:2603.17616#pg1>.

Taro: The next thing we need to look at is how this performs when the world misbehaves, because that's what matters for real deployment <ref:2603.17616#pg3>.

The paper's improvements: Tom: So, we’re talking about how this paper suggests fixing the problems they found in previous hybrid beamforming architectures <ref:2603.17616#pg3>.

Rosa: They show that even with a fixed number of layers, you can get a much better result if you program those diagonal phase-shifting layers correctly <ref:2603.17616#pg2>. It’s about steering the analog beamformer to make the power transport semi-unitary <ref:2603.17616#pg3>.

Dev: That means you can get that high performance without needing a massive, complex digital precoder for every single layer <ref:2603.17616#pg3>. It’s about efficiency in the hardware loop <ref:2603.17616#pg2>.

Taro: And they gave us this stream-aware depth guideline, which is that you need enough layers M to be roughly proportional to two times S times N minus r <ref:2603.17616#pg3>. That’s a concrete way to design the hardware for autonomy systems <ref:2603.17616#pg3>.

Rosa: And that depth guideline is important because it tells you the minimum size you need to start seeing those gains in power preservation <ref:2603.17616#pg3>.

Dev: For my job, that means I can set a hard latency budget based on that guideline, and we know what the safety case is for performance under ideal conditions <ref:2603.17616#pg2>.

Taro: What’s interesting about the quantization part they mentioned is that it doesn't destroy the semi-unitary nature of the processor <ref:2603.17616#pg3>. So you can use lower resolution phase shifters without losing that power efficiency we just talked about <ref:2603.17616#pg3>.

Rosa: That’s a huge win for field trials, because it means the hardware doesn't have to be perfect to deliver near-optimal results in the real world <ref:2603.17616#pg2>.

Dev: Exactly. The paper suggests that we can push the limits of what’s achievable with analog processing by making that programming more efficient and robust against hardware imperfections <ref:2603.17616#pg3>.

Taro: So, the implication is that we can design AI systems with less complex analog circuitry while still maintaining near-exact performance against fully digital baselines <ref:2603.17616#pg3>.

Rosa: Right. This architecture lets us trade off complexity for power efficiency and better spectral efficiency under equal injected power <ref:2603.17616#pg2>.

Dev: It’s a step towards making hybrid beamforming more practical for real-time loops because the performance is tied to the physical budget rather than just radiated power normalization <ref:2603.17616#pg1>.

Taro: So, moving forward, we need to test this not just in ideal lab conditions but see how it handles channel variations and interference when deployed outside a controlled environment <ref:2603.17616#pg2>.

Conclusion: Rosa: So we’ve gone over how this paper proves that by using programmable unitary RF networks, you can achieve power-preserving RF transport from the RF chains to the antenna ports under ideal matched hardware models <ref:2603.17616#pg1>.

Dev: Right. Basically, they show that you can get almost exactly what a fully-digital precoder gives you, but with better efficiency because the analog network isn't just passive lossy stuff <ref:2603.17616#pg3>.

Taro: The main thing that stands out is the stream-aware depth guideline for when subspace containment becomes plausible <ref:2603.17616#pg3>. That gives us a way to design hardware without needing to guess how deep we need to go <ref:2603.17616#pg3>.

Rosa: And the quantization part, as we discussed, is handled because even when you quantize the phase shifters, the underlying processor still stays unitary and maintains that power normalization <ref:2603.17616#pg3>.

Dev: That’s a big deal for latency. It means our real-time loops can be more aggressive because we know the hardware won't introduce unexpected power drops <ref:2603.17616#pg2>.

Taro: I just think what this means for autonomy is that if you program that analog subspace to contain the target subspace, you get exact recovery when things go wrong <ref:2603.17616#pg3>. That’s a strong safety case for us <ref:2603.17616#pg3>.

Rosa: So we’re looking at a way to get that near-perfect performance while keeping the hardware simpler and more power efficient than current baseline hybrid architectures <ref:2603.17616#pg2>.

Dev: It really puts the physical budget front and center, showing that the way you structure your analog network matters more than just how much total power you inject <ref:2603.17616#pg1>.

Taro: I’m still thinking about how this performs when the world misbehaves, because that's what matters for real deployment <ref:2603.17616#pg3>.

Rosa: So to wrap up, "Programmable Unitary RF Networks for Narrowband Hybrid Beamforming" shows us a way to design hybrid beamformers that preserve power exactly under ideal conditions <ref:2603.17616#pg2>.

Dev: It’s a solid piece of research showing how to leverage unitary analog processing for better spectral efficiency than the contractive transfer found in passive fully connected networks <ref:2603.17616#pg3>.

Taro: The next thing we need to look at is applying those depth guidelines to more complex, non-ideal channel environments where the world doesn't behave perfectly <ref:2603.17616#pg3>.

Rosa: That’s what we’ll be looking into next, pushing this concept outside of the perfect matched hardware model and into the messy reality of the field <ref:2603.17616#pg2>.

Innopolis University

eess.SP

Submitted: 2026-03-18

Updated: 2026-10-29

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 90/100

The gist: Hybrid beamforming via programmable unitary RF networks proposes an architecture based on a programmable unitary RF network that achieves power-preserving RF transport from the RF-chains to the

Key concepts

Programmable Unitary RF Network: This is the core analog component of the beamformer. It uses fixed mixing layers interleaved with programmable diagonal phase-shifting layers. These programmability allows the system to define a specific analog subspace that precisely matches the desired digital precoder, enabling exact hybrid recovery when possible.
Power-Preserving RF Transport: This refers to maintaining the injected power from the RF chains exactly at the antenna ports. The proposed architecture achieves this because its analog beamformer is designed to be semi-unitary. This contrasts with conventional passive hybrid networks where power transfer is inherently lossy or contractive.
Semi-Unitary Analog Beamformer: The derived analog beamformer possesses a semi-unitary property under equal injected power. This crucial mathematical characteristic ensures that the total injected RF chain power is preserved when reaching the antenna ports, which is key to its superior efficiency compared to lossy passive architectures.
Stream-Aware Depth Guideline: This guideline provides a practical rule for designing the hardware depth (number of programmable layers M). It suggests that for exact recovery, M should be at least 2S(N-r), where S, N, and r are system parameters. This helps AI designers select a minimal yet sufficient number of programmable layers to achieve near-exact performance.

Terminology

Quantization — 22× in this paper · explained in 15 episodes
Benchmark — 15× in this paper · explained in 2 episodes
Hybrid Architecture — 4× in this paper · explained in 2 episodes
Hybrid System — 4× in this paper · explained in 2 episodes
Universality — 4× in this paper · explained in 2 episodes
Adam — 3× in this paper · explained in 3 episodes
Circuit Depth — 3× in this paper · explained in 2 episodes
Core — 3× in this paper · explained in 2 episodes
Hybrid Architectures — 3× in this paper · explained in 3 episodes
Objective Function — 3× in this paper · explained in 3 episodes
Expressivity — 2× in this paper · explained in 2 episodes
Sufficiency — 2× in this paper · explained in 2 episodes
Calibration — 1× in this paper · explained in 9 episodes
Failure Modes — 1× in this paper · explained in 4 episodes
Gauge Fixing — 1× in this paper · explained in 2 episodes
Papers — 1× in this paper · explained in 3 episodes
Regularization — 1× in this paper · explained in 3 episodes
Steering — 1× in this paper · explained in 2 episodes

Summary

Hybrid beamforming via programmable unitary RF networks proposes an architecture based on a programmable unitary RF network that achieves power-preserving RF transport from the RF-chains to the antenna ports under ideal matched hardware models.

The Core Problem Addressed

Conventional hybrid beamforming architectures are often compared with one another and with the fully-digital architecture under the same radiated antenna power. This issue is especially pronounced for fully-connected splitter–phase-shifter–combiner networks, whose physical power transfer remains contractive even under ideal passive-component assumptions. The paper addresses the physically relevant budget, which is the power injected by the RF-chain outputs into the passive analog RF network, which is then usually transferred to the antenna ports in a contractive (lossy) manner.

Proposed Architecture and Power Preservation

The proposed architecture realizes an analog RF network as an interlaced mixer–phase architecture consisting of fixed (non-programmable) mixing layers interleaved with programmable diagonal phase-shifting layers. Under ideal passive-component assumptions, all injected RF-chain power reaches the antenna ports without loss. The analog beamformer is derived to conduct power-preserving RF transport from the RF-chains to the antenna ports and is programmed via its phase-shifter settings to realize the desired analog subspace. This results in a hybrid precoder that closely matches the fully-digital precoder, and exact hybrid recovery follows whenever the programmed analog subspace contains the target subspace.

Derivation of Digital Beamformer and Programming Method

The authors derive a closed-form digital beamformer and a low-complexity programming method for the analog beamformer, which result in a hybrid precoder that closely matches the fully-digital precoder. The optimal analog beamformer programming is found by maximizing the objective function Φ⋆ = arg max Φ∥F HtarFRF(Φ)∥2 F. A stream-aware necessary guideline for when such containment becomes plausible is provided in the depth analysis.

Performance Evaluation and Comparison

The proposed architecture is evaluated against a fully-digital benchmark, two physically modeled fully-connected hybrid beamforming baselines, and an ideal-lossless Butler/DFT beamselection baseline under equal total injected RF-chain power. Numerical results show that the continuous-phase and 6-bit realizations of the proposed architecture are nearly indistinguishable from the fully-digital benchmark and achieve significant gains over the baseline hybrid beamforming architectures. The gain arises primarily from the fact that its RF transport remains semi-unitary under equal injected power, whereas the passive fully-connected baselines are physically contractive.

Impact of Circuit Depth and Quantization

The performance benefits are guided by a stream-aware depth guideline suggesting that for exact recovery, the required number of programmable layers M must satisfy M ≳ 2S(N − r). Phase quantization preserves RF efficiency because every quantized diagonal layer Dk(ϕ(q)k) remains unitary, and thus the injected-power normalization is maintained. The numerical results show that the proposed architecture benefits primarily from increasing the circuit depth up to about M ≈ 2S, after which the returns diminish markedly.

Conclusion on Power Efficiency

The paper concludes that because the induced analog beamformer is semi-unitary, the architecture preserves the injected RF-chain power exactly at the antenna ports under the ideal matched hardware model. This power-preserving property survives phase quantization, sharply distinguishing it from passive fully-connected hybrid beamforming architectures whose RF transfer is contractive. The numerical results for the spherical-wave downlink show that the proposed hybrid beamforming architecture can closely reproduce the performance of the fully-digital architecture while providing significant performance gains over the baseline architectures under equal injected power.

References

[1] R. W. Heath, Jr., N. Gonzalez-Prelcic, S. Rangan, W. Roh, and A. M. Sayeed, “An overview of signal processing techniques for millimeter wave MIMO systems,” IEEE Journal of Selected Topics in Signal Processing, vol. 10, no. 3, pp. 436–453, 2016

[2] A. Garcia-Rodriguez, V. Venkateswaran, P. Rulikowski, and C. Masouros, “Hybrid analog–digital precoding revisited under realistic RF modeling,” IEEE Wireless Communications Letters, vol. 5, no. 5, pp. 528–531, Oct. 2016

[3] F. Sohrabi and W. Yu, “Hybrid digital and analog beamforming design for large-scale antenna arrays,” IEEE Journal of Selected Topics in Signal Processing, vol. 10, no. 3, pp. 501–513, Apr. 2016

[4] X. Yu, J. Zhang, and K. B. Letaief, “Doubling phase shifters for efficient hybrid precoder design in millimeter-wave communication systems,” Journal of Communications and Information Networks, vol. 4, no. 2, pp. 51–67, Jun. 2019

[5] M. Reck, A. Zeilinger, H. J. Bernstein, and P. Bertani, “Experimental realization of any discrete unitary operator,” Physical Review Letters, vol. 73, no. 1, pp. 58–61, Jul. 1994

[6] W. R. Clements, P. C. Humphreys, B. J. Metcalf, W. S. Kolthammer, and I. A. Walmsley, “Optimal design for universal multiport interferometers,” Optica, vol 3, no 12, pp 1460–1465

[7] R. Keshavarz, K. Zelaya, N. Shariati, and M.-A. Miri, “Programmable circuits for analog matrix computations,” Nature Communications, vol 16, p 8514

[8] J. Alvarez-Vizoso and D. Barral, “Universality and optimal architectures for layered programmable unitary decompositions,” arXiv preprint arXiv:2510.19397, Oct 2025

[9] X. Gao, L. Dai, S. Han, C.-L. I, and R. W. Heath, Jr., “Energy-efficient hybrid analog and digital precoding for mmWave MIMO systems with large antenna arrays,” IEEE Journal on Selected Areas in Communications, vol 34, no 4, pp 998–1009

[10] Z. Wu, M. Nerini, and B. Clerckx, “Microwave linear analog computer (MiLAC)-aided multiuser MISO: Fundamental limits and beamforming design,” arXiv preprint arXiv:2601.

Improvements for AI systems

Programmable unitary RF networks allow AI systems to implement analog beamformers that exactly reproduce fully-digital precoders under equal injected power, achieving exact hybrid recovery when the programmed analog subspace contains the target subspace.
Stream-aware depth scaling provides a necessary guideline for when subspace containment becomes plausible, suggesting that the required number of programmable layers M is roughly proportional to 2S(N-r). This allows AI systems to design hardware architectures with minimal complexity while maintaining near-exact performance against fully-digital baselines.
Phase quantization does not alter the semi-unitary nature of the analog processor, meaning quantizing the programmable phase layers reduces the reachable analog subspace, but it does not change the semi-unitarity of the analog processor and therefore does not alter the equal-injected-power normalization. This means AI systems can use lower resolution digital hardware for phase shifters without sacrificing power efficiency.
The system can achieve significant performance gains over baseline architectures like FC1 and FC2 because its RF transport remains semi-unitary under equal injected power, whereas the passive fully-connected baselines are physically contractive. This allows AI systems to leverage unitary analog processing for better spectral efficiency than conventional hybrid designs.