Rotatable Antenna-Enabled Mobile Edge Computing

Abstract: In the evolving landscape of mobile edge computing (MEC), enhancing communication reliability and computation efficiency to support increasingly stringent low-latency services remains a fundamental challenge. Rotatable antenna (RA) is a promising technology that introduces new spatial degrees of freedom (DoFs) to tackle this challenge. In this letter, we investigate an RA-enabled MEC system where antenna boresight directions can be independently adjusted to proactively improve wireless channel conditions for latency-critical users. We aim to minimize the maximum computation latency by jointly optimizing the MEC server computing resource allocation, receive beamforming, and the deflection angles of all RAs. To address the resulting non-convex problem, we develop an efficient alternating optimization (AO) framework. Specifically, the optimal edge computing resource allocation is derived based on the Karush-Kuhn-Tucker (KKT) conditions. Given the computing resources, the receive beamforming is optimized using semidefinite relaxation (SDR) combined with a bisection search. Furthermore, the RA deflection angles are optimized via fractional programming (FP) and successive convex approximation (SCA). Simulation results verify that the proposed RA-enabled MEC scheme significantly reduces the maximum computation latency compared with conventional benchmark methods.

• The paper introduces an alternating optimization framework that jointly designs task offloading, edge CPU allocation, receive beamforming, and rotatable-antenna orientation to minimize maximum user latency.
• Rotatable antennas reduce latency most effectively at intermediate offloading power and moderate user density, while gains diminish when noise, computing capacity, or multi-user contention becomes dominant.
• The method converges to a stationary solution using KKT-based allocation, semidefinite relaxation, and fractional programming with successive convex approximation, but does not guarantee global optimality.

This letter investigates the integration of rotatable antenna (RA) technology into mobile edge computing (MEC), formulating and solving a joint optimization of computation offloading, edge resource allocation, receive beamforming, and antenna orientation to minimize the maximum computation latency across multiple users. The work is positioned as a first exploration of RA within MEC architectures, addressing the bottleneck that wireless offloading links impose on latency-critical services (2603.16275).

System model

The authors consider an uplink MEC system in which $K$ single-antenna devices offload computation tasks to a base station (BS) equipped with an edge server and a uniform planar array (UPA) of $N = N_y N_z$ directional RAs. Each RA can independently rotate its 3D boresight, parameterized by a zenith deflection angle $\theta_n^e \in [0, \theta_{\max}]$ and azimuth angle $\theta_n^a$, with the pointing vector constrained to unit norm. The channel model combines distance-dependent large-scale fading with Rician small-scale fading, in which the line-of-sight component is weighted by a directional gain pattern $G(\epsilon, \phi) = G_0 \cos^{2p}(\epsilon)$ with $G_0 = 2(2p+1)$ enforcing power conservation. Rotating the boresight thus directly modulates the effective channel gain toward each device.

Each device splits a task of $L_k$ bits between local computation (at CPU frequency $f_k^l$) and edge offloading. The edge latency comprises transmission at rate $R_k = B\log_2(1+\gamma_k)$, where $\gamma_k$ is the SINR under linear receive beamforming, plus edge execution at allocated frequency $N = N_y N_z$0 under a total budget $N = N_y N_z$1. Since local computing and offloading proceed in parallel, the per-device latency is the maximum of the two branches, and the objective is the min–max latency $N = N_y N_z$2. The resulting problem (P1) is non-convex due to coupled integer offloading variables, fractional SINR terms, and the nonlinear dependence of the channel on antenna orientations.

Alternating optimization framework

The proposed solution decomposes (P1) into computation-side and communication-side subproblems solved iteratively via alternating optimization (AO).

Offloading and computing allocation. For fixed beamforming and pointing matrices, the offloaded data size $N = N_y N_z$3 admits a closed-form continuous solution obtained by balancing the local and edge latency branches, with integer feasibility recovered by rounding to the nearest of the floor or ceiling values. Substituting this solution reduces the resource allocation to a convex problem whose KKT conditions yield a closed-form allocation: under the min–max objective, only the bottleneck device's multiplier is active, and $N = N_y N_z$4 is expressed in closed form with the dual variable found by bisection. A notable structural property is that at optimality the local and edge latencies are equal, $N = N_y N_z$5, which allows the communication subproblem to be written purely in terms of edge latency.

Receive beamforming. For fixed orientations, the SINR constraints are handled by lifting $N = N_y N_z$6 and applying semidefinite relaxation (SDR), dropping the rank-one constraint. The resulting quasi-convex problem is solved by bisection over the SINR targets, with each feasibility check being a convex semidefinite program; rank-one solutions are recovered via Gaussian randomization when necessary. The authors acknowledge that SDR may return high-rank solutions, so global optimality of the beamforming step is not guaranteed.

RA pointing optimization. For fixed beamformers, the fractional SINR structure is addressed with the quadratic transform from fractional programming, introducing auxiliary variables $N = N_y N_z$7 with closed-form updates. The residual non-convexity in the pointing matrix $N = N_y N_z$8 is handled via successive convex approximation (SCA): the desired signal term is linearized by first-order Taylor expansion, and the interference terms $N = N_y N_z$9 are upper-bounded by a quadratic surrogate. The appendices derive the gradient and Hessian of $\theta_n^e \in [0, \theta_{\max}]$0 in closed form and construct the curvature constant $\theta_n^e \in [0, \theta_{\max}]$1 as the absolute value of the scalar Hessian coefficient, exploiting that $\theta_n^e \in [0, \theta_{\max}]$2 is rank-one. This yields a valid local Lipschitz constant and guarantees the majorization condition $\theta_n^e \in [0, \theta_{\max}]$3.

The overall AO iteration is monotonically non-increasing in the objective and lower-bounded, so convergence is guaranteed, though only to a Karush–Kuhn–Tucker-type stationary point rather than a global optimum. The per-iteration complexity is dominated by the SDP step, scaling as $\theta_n^e \in [0, \theta_{\max}]$4.

Simulation results

Simulations use a 2.4 GHz carrier, $\theta_n^e \in [0, \theta_{\max}]$5 RAs with half-wavelength spacing, directivity parameter $\theta_n^e \in [0, \theta_{\max}]$6, $\theta_n^e \in [0, \theta_{\max}]$7, $\theta_n^e \in [0, \theta_{\max}]$8 devices at radius 40 m, and $\theta_n^e \in [0, \theta_{\max}]$9 cycle/s unless varied. The RA-enabled scheme is compared against fixed directional antennas ($\theta_n^a$0), isotropic antennas ($\theta_n^a$1), and random orientations. Three trends emerge:

• Versus offloading power $\theta_n^a$2: the RA scheme consistently achieves lower latency than all benchmarks, but the gap shrinks at both extremes. At low $\theta_n^a$3, noise dominates and directional gain provides little benefit; at high $\theta_n^a$4, latency saturates at the computing-capacity limit, so communication improvements no longer help. This identifies the regime in which RA rotation is most valuable as intermediate SNR, where offloading links are the bottleneck.
• Versus $\theta_n^a$5: latency gains from added computing capacity diminish once $\theta_n^a$6 is sufficiently large, at which point the offloading link becomes dominant. The authors draw the practical implication that equipping edge nodes with extremely large computing capacity is unnecessary for latency minimization when the wireless link is the limiting factor.
• Versus number of devices $\theta_n^a$7: the advantage of the RA scheme over fixed and random-orientation baselines narrows as $\theta_n^a$8 grows, because per-device edge resources shrink and simultaneous boresight optimization for all users becomes increasingly difficult. This is an honest limitation of the multi-user scaling behavior of the approach.

Limitations and open questions

The paper concedes several assumptions and restrictions. The RA model preserves only rotational flexibility—a simplified member of the movable antenna (MA)/six-dimensional movable antenna (6DMA) family—so position adaptation, which could offer larger gains, is excluded. The feedback delay of computation results is neglected, the channel is quasi-static flat fading, and devices are assumed capable of simultaneous local computing and offloading. The SDR step and the AO framework guarantee only stationary solutions, and no optimality gap is quantified. Open questions include how RA orientation optimization scales to dense multi-user regimes where its advantage diminishes, whether joint position-and-rotation (6DMA) control yields further latency reduction, and how the framework extends to dynamic channels where deflection angles must be tracked over time.

Conclusion

This letter establishes a joint communication-and-computation design for RA-enabled MEC, combining closed-form KKT-based computing allocation, SDR-based beamforming, and FP/SCA-based antenna orientation optimization within a convergent AO framework. The results demonstrate that adaptively rotating antenna boresights reduces maximum computation latency relative to fixed, isotropic, and random baselines, while also delineating the operating regimes—moderate offloading power, moderate user density—where the benefit is most pronounced.