Low-Complexity Multi-Target Detection in ELAA ISAC
Diluka Galappaththige, Shayan Zargari, Chintha Tellambura, Geoffrey Ye Li · 2025
Read Low-Complexity Multi-Target Detection in ELAA ISAC by Diluka Galappaththige, Shayan Zargari, Chintha Tellambura, Geoffrey Ye Li (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-17776.txt.
Introduction
Diluka Galappaththige and co-authors speed up multi-target sensing in a 2025 letter. Huge antenna arrays at high frequencies can talk and sense at once. They must form many beams to serve users and spot targets. Classic methods are slow: one past run took two weeks for a mid-size array.
The joint problem is non-convex, with rate goals, sensing targets, and power caps. Their method searches on a curved manifold so the rules hold by design. A stochastic gradient method with an augmented Lagrangian drives the search. The result runs far faster than standard benchmarks on large arrays. The work makes real-time use of giant arrays more plausible.
Catalog record
| CREATOR | Diluka Galappaththige, Shayan Zargari, Chintha Tellambura, Geoffrey Ye Li |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 31 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-17776
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Low-Complexity Multi-Target Detection in ELAA ISAC Diluka Galappaththige, Shayan Zargari, Chintha Tellambura, Geoffrey Ye Li Abstract Multi-target detection and communication with extremely large-scale antenna arrays (ELAAs) operating at high frequencies necessitate generating multiple beams. However, conventional algorithms are slow and computationally intensive. For instance, they can simulate a 200 200 -antenna system over two weeks, and the time complexity grows exponentially with the number of antennas. Thus, this letter explores an ultra-low-complex solution for a multi-user, multi-target integrated sensing and communication (ISAC) system equipped with an ELAA base station (BS). It maximizes the communication sum rate while meeting sensing beampattern gain targets and transmit power constraints. As this problem is non-convex, a Riemannian stochastic gradient descent-based augmented Lagrangian manifold optimization (SGALM) algorithm is developed, which searches on a manifold to ensure constraint compliance. The algorithm achieves ultra-low complexity and superior runtime performance compared to conventional algorithms. For example, it is 56 56 times faster than the standard benchmark for 257 257 BS antennas. I Introduction Multi-target detection in integrated sensing and communication (ISAC) systems is a critical application. However, ISAC base stations (BSs) will use large-scale antenna arrays (ELAA) at high frequencies [ 1 , 2 ] , which exhibit near-field (NF) propagation [ 3 ] . Of course, far-field (FF) ISAC research is also emerging [ 4 , 5 , 6 ] . The high dimensionality of ELAA systems, e.g., a vast number of antenna elements, exponentially increases the computational complexity [ 1 , 2 ] . Traditional successive convex approximation (SCA) and semi-definite relaxation (SDR) methods are inefficient due to overhead and extended running times. For example, in [ 2 ] , a 200-antenna BS ISAC system required two weeks to generate a single simulation curve, with time complexity growing exponentially as the number of antennas increases. This challenge calls for super-efficient algorithms. The problem’s search space has size MKN , where M , K , and N are the numbers of BS antennas, users, and targets, respectively. Consequently, the search space grows rapidly as network dimensions increase, leading to exponential complexity. Thus, a fundamentally different algorithm is required. The key novelty of our work lies not only in the algorithm but in addressing the high computational complexity of beamforming for multi-user, multi-target NF ISAC systems with ELAAs. To our knowledge, multi-target detection in NF ISAC systems remains largely unexplored. The closest related work, our previous paper [ 2 ] , employs SCA and SDR for FF ISAC systems. This study is the first to introduce a low-complexity beamforming for multi-target detection and communication in NF ISAC systems. In contrast, many NF ISAC studies focus on single-target scenarios. Building on [ 7 ] , our algorithm extends beyond standard optimization (including manifold optimization (MO) and fractional programming (FP)) by integrating a penalty-based augmented Lagrangian method (ALM) directly into the MO framework. While [ 7 ] developed a low-complexity iterative augmented Lagrangian MO (IALMO) algorithm for FF ISAC systems, it focuses on a 3\text{\,}\mathrm{GHz} frequency with fewer than 28 BS antennas. At higher frequencies, like 54\text{\,}\mathrm{GHz} , smaller wavelengths enable antenna arrays with up to 500 elements. Although adaptable, IALMO’s performance has not yet been evaluated for ELAA-ISAC systems over the 10\text{\,}\mathrm{GHz} - 10\text{\,}\mathrm{THz} range. Moreover, [ 7 ] reuses the same communication beams for target detection. In NF, this dual usage is impractical due to beam-focusing effects [ 3 ] , where energy is concentrated in both angle and distance, reducing coverage and limiting multi-purpose applications. To address this, our approach employs separate beams for communication and sensing (C&S) with ELAA (see Remark 1 ). Unlike [ 7 ] , which uses Riemannian conjugate (RC) methods, we adopt Riemannian stochastic gradient descent (SGD) to solve the proposed beamforming problem. Riemannian SGD improves large-scale optimization by approximating gradients using subsets of data at each iteration. This reduces computational overhead compared to full gradient descent, allowing faster updates and enhanced scalability [ 8 ] . This letter presents a Riemannian SGD-based augmented Lagrangian MO (SGALM) algorithm to overcome the aforementioned computational complexity. It is designed to maximize communication sum rates while meeting sensing thresholds and minimum rate requirements. Since the optimization problem is non-convex and NP-hard, we augment the objective function with a penalty for beampattern gain violations and minimize over a complex sphere manifold, following the ALM [ 9 ] . SGALM iteratively adjusts optimization variables and Lagrange parameters to satisfy constraints. Compared to standard techniques, SGALM archives dramatic complexity savings. For example, with 257 and 321 BS antennas, SGALM is 56 and 61 times faster than the benchmark. This efficiency gain scales up with the antenna array size, making SGALM ideal for ELAA-ISAC systems and dynamic networks. While this study focuses on NF propagation, SGALM is versatile and can be applied to various operating scenarios, including FF environments and different antenna configurations. This flexibility makes it suitable for diverse applications, ranging from urban environments with dense multi-path interference to rural areas with line-of-sight conditions. II Preliminaries Fig. 1: A ELAA-ISAC system. II- 1 System and Channel Models Fig. 1 shows the narrowband ELAA-ISAC system consisting of M=2\tilde{M}+1 -uniform linear array (ULA) antenna BS with antenna spacing d=\lambda/2 , K single-antenna communication users, and N sensing targets, where \lambda is the wavelength [ 2 ] . The NF spherical wave channel model is used. The NF channel vector between the BS and the b -th node ( b=k for the k -th user and b=n for the n -th target), \mathbf{f}_{b}\in\mathbb{C}^{M\times 1} , is given as \mathbf{f}_{b}=\beta_{b}\mathbf{a}_{f}(r_{b},\theta_{b}), where \mathbf{f}\in\{\mathbf{h},\mathbf{g}\} , \mathbf{a}_{f}(r_{b},\theta_{b})\in\mathbb{C}^{M\times 1} is the NF array response vector with [\mathbf{a}_{f}(r_{b},\theta_{b})]_{m}=e^{-j\frac{2\pi}{\lambda}(r_{mb}(r_{b},\theta_{b})-r_{b})} as the m -th element, and r_{mb}(r_{b},\theta_{b}) is given in [ 2 , Eqn. (1)] . Here, r_{b} and \theta_{b} are the b -th node distance and angle from the center of the ULA, respectively. Moreover, \beta_{b}=\tilde{\beta}_{b}e^{-j\frac{2\pi}{\lambda}r_{b}} is the complex channel gain, and \tilde{\beta}_{b}=\sqrt{{\lambda}/{4\pi}}r_{b}^{-1} is the free-space path-loss between the 0 -th antenna element and the b -th node. Further details of the channel and system models can be found in [ 2 ] . Channel estimation and data transmission/sensing occur in two separate time slots. The initial slot in each coherent interval is allocated to estimating channel state information (CSI), feasible with emerging methods [ 10 ] . Thus, CSI is available. This assumption is standard and widely used. Nevertheless, the proposed beamforming algorithm is robust to imperfect CSI. The details of how it can adapt to imperfect CSI can be found in [ 7 ] . II- 2 Signal Model The BS transmits signal \mathbf{x}\in\mathbb{C}^{M\times 1} , which is jointly designed for communicating with users while sensing the targets, i.e, \mathbf{x}=\sum_{k\in\mathcal{K}}\mathbf{w}_{k}q_{k}+\sum_{n\in\mathcal{N}}\mathbf{s}_{n} . Here, q_{k}\in\mathbb{C} denotes the intended data symbol for the k -th user with unit power, i.e., \mathbb{E}\{|q_{k}|^{2}\}=1 , \mathbf{w}_{k}\in\mathbb{C}^{M\times 1} is the BS data beamforming vector for the k -th user, and \mathbf{s}_{n}\in\mathbb{C}^{M\times 1} is the sensing signal intended for detecting the n -th target. It is assumed that q_{k} and \mathbf{s}_{n} are independent of each other for k\in\mathcal{K} and n\in\mathcal{N} , and the beamforming at the BS is achieved through designing \mathbf{w}_{k} and \mathbf{s}_{n} [ 11 , 12 ] . The received signal at the k -th user is given by \displaystyle y_{k}=\mathbf{h}_{k}^{\rm{H}}\mathbf{w}_{k}q_{k}+\sum\nolimits_{i\in\mathcal{K}_{k}}\mathbf{h}_{k}^{\rm{H}}\mathbf{w}_{i}q_{i}+\sum\nolimits_{n\in\mathcal{N}}\mathbf{h}_{k}^{\rm{H}}\mathbf{s}_{n}+\varsigma_{k},\quad (1)