A new preprint reports that a dynamic way of assigning radar sensors and time-frequency resources performed better than equal sharing when communication capacity was scarce. At high capacity, both approaches reached the same maximum achievable estimation accuracy.
When the links are the bottleneck
The study examines resource allocation in a capacity-constrained, two-hop cloud radar network. In the proposed buffered access protocol, sensors apply local spectral windowing before sending measurements through intermediate Edge Servers to a Fusion Center. The architecture makes communication limits central to the allocation problem: the system must decide which sensors to use and how to divide time and frequency resources.
That decision is framed as a mixed-integer optimization. It minimizes the combined Cramér-Rao lower bound, or CRLB, across targets while jointly choosing sensors and assigning time-frequency resources under fronthaul and backhaul limits. A CRLB is a theoretical lower limit on estimation error, so a lower value represents more precise estimation within the model. The proposed solution uses a Big-M formulation and successive convex approximation, an iterative procedure that tackles the original problem through a series of approximations. In the hybrid version, sensor-selection and fronthaul-allocation variables can take continuous values between zero and one, while backhaul-allocation variables remain binary, or on and off.
A surrogate for the target measurement
The paper does not optimize geometric-parameter CRLB directly. It describes that problem as intractable and instead uses complex-reflectivity estimation, which the paper says is directly related to integrated signal-to-noise ratio. That choice provides a tractable objective, but it also narrows what the result means: the reported accuracy is based on a reflectivity surrogate rather than a direct geometric-parameter CRLB.
What was tested
The evaluation was entirely simulated. The model contained two edge servers and one fusion center, with sensors uniformly distributed inside a circle with a radius of 10 metres and four targets placed at random. Across 50 Monte Carlo trials, target placements and channel realizations were randomized. The outcome was the sum of normalized CRLBs across all targets, while communication capacity was normalized by the total sensor flow rate to form a capacity-to-load comparison. The dynamic method was compared with a static baseline that allocated resources equally across all sensors.
The analysis tracked the sum of normalized CRLBs as the capacity-to-load ratio changed. The supplied results give the direction of the comparison, but no numerical effect size or uncertainty interval. That means the summary shows which approach was favored, not how large the gap was.
The gap appears at low capacity
At low communication capacity, the proposed allocation retained reasonable performance. The equal-resource baseline degraded and eventually broke down when the capacity available per sensor could not cover all targets. The authors describe this as better performance from dynamic prioritization under tight constraints, but the paper does not report a numerical size for the advantage.
At high capacity, the two approaches converged. When the capacity-to-load ratio was greater than 2, both methods reached the maximum achievable estimation accuracy. In these simulations, the reported difference was concentrated in the constrained regime, while the equal-resource approach caught up once capacity was ample.
More sensors do not tell a simple story
The paper also gives opposing names to the effect of adding sensors. For the proposed framework, the reported pattern is called Sensor Gain. For the static baseline, the reported deterioration is called Dilution Loss. These labels describe the direction of the simulated trends; the supplied analysis does not give a numerical sensor-count effect.
A result still bounded by its model
These findings are a comparison inside a particular model, not a demonstration in a deployed radar system. The evaluation uses Monte Carlo simulations and does not establish performance in hardware or field conditions. The setup has two edge servers, one fusion center and four randomly placed targets, while more complex network topologies and moving-target scenarios remain open questions. The objective also uses complex-reflectivity CRLB as a surrogate, and the iterative Big-M and successive-convex-approximation solution does not establish global optimality.
The work was co-funded by the LOEWE initiative within the emergenCITY center and partly by BMFTR through the Open6GHub plus project. The document is an arXiv version 1 preprint dated 28 Aug 2026.
Paper data and sources
Original title: Resource Allocation for Cloud Radar Networks with Communication Constraints
Authors: Christian Eckrich, Abdelhak M. Zoubir, Vahid Jamali
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text