PRIMO: GPU-Accelerated Joint Beamforming and RIS Placement Optimization for Outdoor-to-Indoor Coverage at FR1(C) and FR3
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "PRIMO: GPU-Accelerated Joint Beamforming and RIS Placement Optimization for Outdoor-to-Indoor Coverage at FR1(C) and FR3".
Dev: Here's the gist of the paper. Upper mid-band spectrum is promising for 5G and 6G cellular networks, yet concrete and low-emissivity glass impose over 20 dB of building penetration loss above 2 GHz, degrading outdoor-to-indoor (O2I) coverage The paper introduces PRIMO, a training-free joint optimizer of base-station precoding, RIS phase shifts, and RIS placement validated by calibrated ray tracing in a MetroTech Commons digital twin at 6.75 and 16.95 GHz Problem Formulation The core problem addressed is maximizing the minimum user rate over base-station precoding, RIS phase shifts, and RIS position under constraints on total power and unit-modulus constraints on phases Problem (P1) is non-convex due to the unit-modulus constraints forming a manifold S N, the coupling among W, θ, and d through channel equations, and the non-smooth inner mink operator.
Rosa: The standard alternating optimization (AO) approach solves this by cycling through three blocks per iteration: the regularized zero-forcing (RZF) precoder W for fixed phases and position, element-wise (EW) phase updates for fixed W and d, and selection of the best d among Nd candidates.
Dev: For O2I, position d is a genuine third block because varying it changes the building material that the RIS-to-UE path traverses, causing channel quality jumps that phase or precoding optimization cannot address.
Rosa: Proposed Optimization Method PRIMO is proposed as a GPUparallel batch optimizer designed to eliminate the block-coordinate structure entirely, addressing limitations of AO-EW.
Title and authors: Rosa: To kick things off, we're looking at the title and authors of this paper, PRIMO: GPU-Accelerated Joint Beamforming and RIS Placement Optimization for Outdoor-to-Indoor Coverage at FR1(C) and FR3. Rosa, what do you make of that title?
Dev: It’s pretty descriptive. It tells us it’s a joint optimization problem involving beamforming and RIS placement, specifically for outdoor-to-indoor coverage in the upper mid-band frequencies.
Taro: I see the focus on FR1(C) and FR3 right there, which suggests they are looking at different frequency bands to see how it performs across the spectrum.
Rosa: Right. And who’s behind this work? It’s Mingjun Ying, Guanyue Qian, Xinquan Wang, Xingchen Liu, Ishaan S. Gupte, Dipankar Shakya and Theodore S. Rappaport from NYU WIRELESS Brooklyn.
Dev: They have a strong background in the area because they’re at NYU Wireless. That suggests the underlying theory is pretty solid for this kind of complex system design challenge.
Rosa: So, to summarize what we've heard about the title and authors, it’s a paper proposing a new method called PRIMO that uses GPU acceleration to jointly optimize beamforming and RIS placement for O2I coverage in specific frequency bands.
Dev: The implication is that they are trying to solve the complex interplay between signal transmission from the base station, the reflection off the RIS, and where you should physically place that surface.
The paper's summary: Rosa: Now let’s get into what PRIMO actually does according to their summary. It’s a training-free joint optimizer of base station precoding, RIS phase shifts, and RIS placement, validated by calibrated ray tracing in a MetroTech Commons digital twin at six point seven five and sixteen point nine five gigahertz.
Dev: The key takeaway from the summary is that they are solving the problem of maximizing the minimum user rate subject to constraints on total power and unit-modulus constraints on phases.
Taro: They specifically identify why this is hard, saying it’s non-convex because those phase constraints create a kind of manifold, and there's a coupling between all these variables through the channel equations.
Rosa: And they mention that standard alternating optimization methods cycle through precoding, then phase updates, and then choosing the best RIS position among candidates. But they point out that position is a real third block because changing it changes the building material the signal has to go through, causing channel quality jumps that just optimizing phases or precoding can't fix.
Dev: So PRIMO’s approach is different because it treats RIS placement as a genuine third block that needs to be optimized along with the others.
Rosa: Right. And they say their validation used calibrated ray tracing in a digital twin environment, which helps them get concrete results before field trials.
The paper's improvements: Dev: Let’s talk about what PRIMO actually improves over the existing methods, based on what the paper suggests. It’s designed as a GPU-parallel batch optimizer to completely eliminate the block-coordinate structure you see in alternating optimization approaches.
Taro: I like that because it changes how they handle updates. PRIMO refreshes the base station precoder every single gradient step, keeping it tightly coupled to the current phases, which is different from alternating optimization where you only recompute the precoder once per outer iteration.
Rosa: And for those phase shifts, instead of dealing with that non-smooth mink operator that makes direct gradient computation tough in traditional methods, PRIMO uses a smooth log-sum-exp surrogate function.
Dev: That smoothing lets them use Wirtinger calculus to differentiate this surrogate function with respect to the unit-modulus variables. That means they can update all the N phases at once simultaneously.
Taro: They’re using Adam with cosine-annealed step size over T steps in two progressive-pruning stages, which is meant to cut computation time down by about five times compared to a full batch optimization run.
Conclusion: Rosa: So, wrapping up the conclusion of PRIMO: they show that this method improves the minimum user rate by up to a factor of one point nine zero over fully converged alternating-optimization baselines at FR3 with sixteen indoor users <ref:2610.03887#pg1>.
Dev: They also report that it converges about one hundred times faster on just a single GPU, which is a significant speedup for real-time applications.
Taro: And they say the capacity lift is load-dependent, growing from seventy-eight percent over FR1(C) up to three hundred eighty-five percent over FR3 when you have up to twenty indoor users.
Rosa: They also found that with this method, you can support ninety-two percent more indoor users at FR3 above the one bit per second per hertz service floor compared to the alternating optimization approach.
Dev: And they suggest that the optimal RIS placement is actually a first-order design variable, meaning you decide where it goes once for the whole served area instead of every single scheduling instant. They also identify adjacent-facade placement with outdoor line of sight to the base station as minimizing cascaded path loss.
Taro: That’s important because that geometry choice directly addresses material-dependent building penetration loss, or BPL, which is something phase optimization alone just can't fix by moving the RIS along a facade track.
Rosa: So in short, PRIMO outperforms classical methods under multi-user interference and offers much faster tracking of channel variations by allowing it to optimize placement and phases together. That’s what we have for this paper today on PRIMO: GPU-Accelerated Joint Beamforming and RIS Placement Optimization for Outdoor-to-Indoor Coverage at FR1(C) and FR3.
NYU WIRELESS
eess.SP
Submitted: 2026-10-02
Updated: 2026-11-09
Comments: 5 pages, 3 figures, 3 tables. Presented at the 2026 Asilomar Conference on Signals, Systems, and Computers, Pacific Grove, CA, USA, Oct. 25-28, 2026
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 83/100
The gist: Upper mid-band spectrum is promising for 5G and 6G cellular networks, yet concrete and low-emissivity glass impose over 20 dB of building penetration loss above 2 GHz…
Key concepts
- Alternating Optimization (AO)
- A standard approach where optimization is done in separate steps: first optimizing the base station signal, then the RIS phases, and finally choosing the best RIS location. This method is slow because it cycles through these blocks sequentially, which doesn't capture complex interactions between precoding and placement simultaneously.
- PRIMO
- A new GPU-accelerated joint optimizer that eliminates the block-coordinate structure of AO. PRIMO updates the base station precoder and RIS phases simultaneously in every step, leading to faster convergence and better performance by treating all variables together.
- Unit-Modulus Constraints
- Mathematical rules requiring that the phase factors used in beamforming (like RIS phase shifts) must have a magnitude of 1. This constraint makes the optimization problem non-convex and difficult to solve, as it forces the solution onto a complex geometric shape called a manifold.
- Outdoor-to-Indoor (O2I) Coverage
- The challenge of ensuring reliable mobile network service inside buildings when signals are transmitted from outside. This is complicated by building materials like concrete and glass, which cause significant signal loss and unpredictable channel changes that need careful management.
Terminology
Summary
Upper mid-band spectrum is promising for 5G and 6G cellular networks, yet concrete and low-emissivity glass impose over 20 dB of building penetration loss above 2 GHz, degrading outdoor-to-indoor (O2I) coverage The paper introduces PRIMO, a training-free joint optimizer of base-station precoding, RIS phase shifts, and RIS placement validated by calibrated ray tracing in a MetroTech Commons digital twin at 6. p. 0 A reconfigurable intelligent
surface (RIS) on a neighboring building bypasses this blockage
by focusing energy onto indoor users. This paper introduces
PRIMO (Parallel RIS Min-Rate Optimization), a training-free
joint optimizer of base-station precoding, RIS phase shifts, and
RIS placement, validated by calibrated ray tracing in a MetroTech
Commons digital twin (Brooklyn, NY) at 6.Where the paper says it
Problem Formulation
The core problem addressed is maximizing the minimum user rate over base-station precoding, RIS phase shifts, and RIS position under constraints on total power and unit-modulus constraints on phases Problem (P1) is non-convex due to the unit-modulus constraints forming a manifold S N, the coupling among W, θ, and d through channel equations, and the non-smooth inner mink operator. The standard alternating optimization (AO) approach solves this by cycling through three blocks per iteration: the regularized zero-forcing (RZF) precoder W for fixed phases and position, element-wise (EW) phase updates for fixed W and d, and selection of the best d among Nd candidates. p. 1 (3d)
Problem (P1) is non-convex due to (i) the unit-modulus
constraints (3c) forming the manifold S
N , (ii) the coupling
among W, θ, and d through (1), and (iii) the non-smooth inner
mink operator; no closed-form joint solution exists. The standard alternating optimization (AO) approach [12]
solves (P1) by cycling through three blocks per iteration: the
regularized zero-forcing (RZF) precoder W for fixed phases
and position, element-wise (EW) phase updates for fixed W
and d, and selection of the best d among Nd candidates. For
O2I, position d is a genuine third block: varying d changes
the building material (concrete vs. glass) that the RIS-to-UE
path traverses, causing channel quality jumps that phase or
precoding optimization cannot address.Where the paper says it
Where the paper says it
Proposed Optimization Method
PRIMO is proposed as a GPUparallel batch optimizer designed to eliminate the block-coordinate structure entirely, addressing limitations of AO-EW. PRIMO refreshes W at every gradient step, keeping the precoder tightly coupled to the current phases, unlike AO-EW which recomputes W once per outer iteration. p. 2 ,hK]
H, and W =
√
P W˜ /∥W˜ ∥F . Unlike
AO-EW, which recomputes W once per outer iteration, PRIMO
refreshes W at every gradient step, keeping the precoder tightly
coupled to the current phases; this closed-form per-step RZF
with first-order phase updates avoids both an inner convex
solve and implicit differentiation, making GPU batching over
hundreds of restarts tractable. The non-smooth mink operator in (3a) prevents direct gradient
computation. PRIMO therefore replaces the mink operator
with the following smooth log-sum-exp surrogate L(θ), with
smoothing parameter αsm >0,
L(θ) = −
1
αsm
log X
K
k=1
e
−αsmRk(θ)
! PRIMO updates all N phases simultaneously, differentiating (5) with respect to the unit-modulus
variables ϕ=[e
jθ1
, . , ejθN ]
T via Wirtinger calculus [23] as
∇ϕ∗L =
X
K
k=1
pk
ln 2 (1+γk)Ik
sk,k ak,k − γk
X
j̸=k
sk,j ak,j
,
(6)
where pk = e
−αsmRk /
P
j
e
−αsmRj are softmax weights concentrating the gradient on the worst UE, sk,j = h
H
k wj ,
ak,j = f
∗
k ⊙ (Gwj ), Ik =
P
j̸=k
|sk,j |
2 + σ
2
, and γk is the
SINR.Where the paper says it
Where the paper says it
Where the paper says it
Validation and Performance Gains
The framework is validated using NYURay RT in a MetroTech Commons digital twin, where BS and RIS are held at the same physical aperture across both bands. The paper demonstrates that for equal physical aperture, half-wavelength spacing yields about 6. p. 0 This paper introduces
PRIMO (Parallel RIS Min-Rate Optimization), a training-free
joint optimizer of base-station precoding, RIS phase shifts, and
RIS placement, validated by calibrated ray tracing in a MetroTech
Commons digital twin (Brooklyn, NY) at 6. At
equal physical aperture, half-wavelength spacing packs 6. At
equal physical aperture, half-wavelength spacing packs 6. 25 times
as many elements per array at FR3 as at FR1(C), so FR1(C)
leads at low indoor-user counts while FR3 dominates at higher
counts, where the additional spatial degrees of freedom suppress
inter-user interference. 25× more elements at FR3 (BS M = 256→1600,
RIS N = 1024 → 6400). Sweeping the load from 4 to
20 indoor users shows the dual-band advantage is loaddependent, with FR3 delivering a 4. 25 times
as many elements per array at FR3 as at FR1(C), so FR1(C)
leads at low indoor-user counts while FR3 dominates at higher
counts, where the additional spatial degrees of freedom suppress
inter-user interference. Above a 1 bps/Hz service floor, PRIMO
supports 92% more indoor users at FR3 than at FR1(C), and
raises the minimum user rate by up to 90% over a fully converged
alternating-optimization baseline while running about 100 times
faster. 70× minimum user
rate over FR1(C) at 12 indoor users (6. 38 bps/Hz),
where the larger element count suppresses inter-user interference (IUI).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion and Impact
The results show that PRIMO improves the minimum user rate by up to a factor of 1.90 over a fully converged alternating-optimization baseline at FR3 with 16 indoor users, while converging about 100 times faster on a single GPU The capacity lift is load-dependent, growing from 78% over FR1(C) to 385% over FR3 as the number of indoor users increases up to K=20. p. 0 PRIMO improves the minimum user rate by up to a factor
of 1. 19 bps/Hz, mean of 5 trials) over a
fully converged alternating-optimization baseline at FR3
with 16 indoor users, while converging about 100 times
faster on a single GPU (57 vs. 25 times
as many elements per array at FR3 as at FR1(C), so FR1(C)
leads at low indoor-user counts while FR3 dominates at higher
counts, where the additional spatial degrees of freedom suppress
inter-user interference. Above a 1 bps/Hz service floor, PRIMO
supports 92% more indoor users at FR3 than at FR1(C), and
raises the minimum user rate by up to 90% over a fully converged
alternating-optimization baseline while running about 100 times
faster. • We sweep the RIS position along a 40-meter facade track
and identify deployment geometry as a first-order design
variable, decided once per site over the served indoor
area rather than per scheduling instant. Adjacent-facade
placement with outdoor LoS to the BS minimizes cascaded
path loss, and RIS phase optimization alone cannot recover
from a poor position. 85× (K = 20), a multi-user capacity-scaling
benefit rather than a constant single-link gain. (2) PRIMO and true AO-EW are the only methods that
remain competitive under multi-user IUI (Table III); PRIMO
additionally supports ∼ 2 ms warm-start tracking and runs
∼100× faster than AO-EW on a single GPU at FR3.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Parallel RIS Min-Rate Optimization (PRIMO) allows for joint optimization of base-station precoding, RIS phase shifts, and RIS placement without relying on alternating optimization or training data. This enables a system to achieve a minimum user rate improvement by up to a factor of 1.90 over fully converged alternating-optimization baselines at FR3 with 16 indoor users while running about 100 times faster on a single GPU.
-
The proposed algorithm supports tracking channel variation for slowly-varying O2I channels by initializing from the previous slot's theta with B=1 and T=10 (warm-start mode). This reduces latency to approximately 2 milliseconds on a single graphics processing unit, which is well within the 10–80 ms beam management periodicity of 5G New Radio.
-
The system can determine optimal RIS placement as a first-order design variable alongside phase optimization. Moving the RIS along the facade changes which building material (concrete vs. glass) the RIS-to-UE path traverses, producing channel quality jumps that no phase optimization alone can recover from.
-
The framework can select a deployment geometry by identifying adjacent-facade placement with outdoor LoS to the BS as yielding the lowest cascaded path loss at both bands. This minimizes cascaded path loss and improves coverage guarantees by addressing material-dependent building penetration loss (BPL).
Abstract
Upper mid-band spectrum is promising for 5G and 6G cellular networks, yet concrete and low-emissivity glass impose over 20 dB of building penetration loss above 2 GHz, degrading outdoor-to-indoor (O2I) coverage. A reconfigurable intelligent surface (RIS) on a neighboring building bypasses this blockage by focusing energy onto indoor users. This paper introduces PRIMO (Parallel RIS Min-Rate Optimization), a training-free joint optimizer of base-station precoding, RIS phase shifts, and RIS placement, validated by calibrated ray tracing in a MetroTech Commons digital twin (Brooklyn, NY) at 6.75 and 16.95 GHz. At equal physical aperture, half-wavelength spacing packs 6.25 times as many elements per array at FR3 as at FR1(C), so FR1(C) leads at low indoor-user counts while FR3 dominates at higher counts, where the additional spatial degrees of freedom suppress inter-user interference. Above a 1 bps/Hz service floor, PRIMO supports 92% more indoor users at FR3 than at FR1(C), and raises the minimum user rate by up to 90% over a fully converged alternating-optimization baseline while running about 100 times faster.
Sources
- Site-Specific RIS Deployment in Cellular Networks via Calibrated Ray Tracing
- The Complex Gradient Operator and the CR-Calculus
Related papers
- Open-H-RF: A Large-Scale Ultrasound Raw Channel Dataset Initiative
- Scalable Attention for 5G NR Channel Estimation
- Low-Altitude Wireless Networks: A Comprehensive Survey
- Sensing-Assisted OFDM with Index Modulation: A Waveform for Optical Wireless Integrated Communication and Localisation
- Iteration-Free Candidate Pruning and an Index-Domain VLSI Architecture for Sort-Free MIMO Box Decoding up to 1024-QAM
- A Foundation Model for Large-Scale Wireless Network Planning, Operation and Optimization