Phones and sensors that send heavy computation to a nearby edge server have to make a set of joint decisions every few milliseconds. They choose how fast to run their own processor, how much transmit power to spend, and how much of the shared bandwidth to claim. The standard way to make these decisions casts them as a per-slot optimization problem and solves it with a sequential sweep: update user 1's variables, then user 2's using user 1's new values, and so on until the sweep converges. This approach comes from Lyapunov queue control, a method that turns a long-run power-versus-delay goal into one small problem per time slot. It works on paper because two assumptions hide its cost. The channel is taken to be constant within a slot and independent across slots, so the problem is worth solving exactly. And nobody counts the time, signaling rounds, or battery spent running the solver itself.
We propose to drop both assumptions together. Our claim is that the per-slot problem should never be solved to convergence. Instead, every device takes one simultaneous update step per slot, in parallel with all other devices, starting from the previous slot's answer. The rate each device plans with should be a calibrated lower bound on what the fading channel will actually deliver, not the Shannon rate of an idealized channel. If this is right, a mobile controller can keep the known power-delay guarantee, up to a penalty we can predict from how fast the channel moves. It would also cost one broadcast round per slot instead of a number of rounds that grows with the user count.
Core Claim
The setting is the multi-user mobile edge computing model of Power-Delay Tradeoff in Multi-User Mobile-Edge Computing Systems and Stochastic Joint Radio and Computational Resource Management for Multi-User Mobile-Edge Computing Systems. Each slot, these papers minimize a weighted sum of power and queue backlog over device CPU frequencies, transmit powers and bandwidth shares. They solve the coupled power-bandwidth part by Gauss-Seidel alternating minimization and prove that average power is within $O(1/V)$ of optimal while queues grow as $O(V)$, where $V$ is a user-chosen weight on power. Their simulations use frequency-flat block fading, i.i.d. across slots.
Our replacement is a proximal Jacobian ADMM step. ADMM, the alternating direction method of multipliers, splits a problem into pieces coupled by a shared constraint. The "Jacobian" variant updates all pieces at once from the previous iterate, and a proximal term damps oscillation. Applied here, each device updates its own frequency and power in closed form given a bandwidth price that the server broadcasts, and the server updates the price. We run exactly one such iteration per slot.
The theoretical claim we intend to prove is an inexact version of the Lyapunov bound. Let $\epsilon_t$ be how far the one-step solution's per-slot objective sits above the true per-slot minimum, let $\rho<1$ be the contraction factor of one proximal Jacobian step on the per-slot problem, and let $\delta_t$ be how far the per-slot minimizer moves between consecutive slots because the channel and queues changed. Standard results for drift-plus-penalty with approximate minimization then suggest
$$\bar{P} \le P^{\mathrm{opt}} + \frac{B + \bar{\epsilon}}{V}, \qquad \bar{\epsilon} \lesssim \frac{L\,\bar{\delta}}{1-\rho},$$
where $\bar{P}$ is long-run average power, $P^{\mathrm{opt}}$ the optimal average power, $B$ the usual drift constant, $L$ a Lipschitz constant of the per-slot objective, and bars denote time averages. The second inequality is the usual steady-state error of a contracting iteration that tracks a target moving by $\bar\delta$ per step. The prediction we can test is concrete: the extra power over the exact controller should grow roughly linearly with channel drift, and therefore with Doppler, and should shrink as the proximal step is tuned to raise contraction.
The channel half of the claim is about which rate goes into the queue update. When a device plans with a Shannon rate that the fading channel then fails to deliver, the block errors out, the bits stay in the queue, and the queue-stability argument is quietly broken. We substitute the rate of the modulation-and-coding choice made by online conformal inference, a calibration procedure that adjusts a threshold so the long-run error frequency matches a target without assuming a fading distribution. This follows Distribution-Free Adaptive Modulation and Coding via Online Conformal Inference under Non-Stationary Fading. The planned rate then carries a marginal guarantee: it fails at most a target fraction $\alpha$ of the time. The queue update can therefore be written with expected delivered bits $(1-\alpha)R$ rather than an optimistic $R$.
Research Program
The work has three parts.
The first is algorithmic. We rewrite Mao et al.'s per-slot problem so that bandwidth is the only coupling constraint, and derive the closed-form per-device proximal update and the server's price update. We then prove the inexact drift-plus-penalty bound above, including the effect of conformal rate miscoverage on queue stability. Two prior papers supply the method. A Proximal Jacobian ADMM Approach for Fast Massive MIMO Signal Detection in Low-Latency Communications applies proximal Jacobian ADMM to MIMO detection, updating all users' symbols simultaneously without matrix inversions. Joint Energy-Bandwidth Allocation for Multi-User Channels with Cooperating Hybrid Energy Nodes uses it to split an energy-bandwidth allocation into closed-form blocks.
Both solve a static problem to convergence. The first models no device computation or queues. The second relies on non-causal knowledge of channels over a horizon, which is exactly what fast fading denies. We instead run the method as a one-step tracker inside a stochastic queue controller. The tensor decomposition literature constrains this design. Generalized Higher-Order Tensor Decomposition via Parallel ADMM shows large speedups from Jacobi-style updates, but that line of work also records that switching from sequential to simultaneous updates can slow or destabilize convergence unless proximal damping is added. We therefore treat the proximal weight as the main design knob and measure $\rho$ empirically rather than assuming it.
The second part is channel realism. Mao et al. assume block fading. Delay Minimization in Pinching-Antenna-Enabled NOMA-MEC Networks assumes exact instantaneous channel knowledge and negligible waveguide loss, and alternates sequentially among its variables. We evaluate under 3GPP tapped-delay-line multipath profiles with speed transitions, as Batra et al. do. Batra et al. motivate this choice: they show static link adaptation degrading under 3 to 500 km/h transitions. They also limit it, since their learning rate is tied to the Jakes Doppler model and their guarantee is marginal rather than per-SNR. We inherit both caveats and test them.
The third part is device cost. Offloading papers model device energy as $\kappa f^3$ and ignore the solver. We will implement the per-device update and the sequential alternatives on a phone-class ARM processor and measure wall time and energy per slot. We will also count signaling rounds between devices and server, since in a distributed deployment each round costs radio energy as well as latency.
Baselines
The naive reference is local-only execution with maximum CPU frequency. Under i.i.d. block fading, Power-Delay Tradeoff in Multi-User Mobile-Edge Computing Systems reports roughly 1000 ms average delay for this policy, against 1.05 ms for their Lyapunov offloading controller. We rerun it in our simulator to anchor the realistic-channel numbers.
The strongest existing method is Mao et al.'s Lyapunov controller with Gauss-Seidel alternating minimization solved to convergence every slot, planning with the Shannon rate under perfect instantaneous channel knowledge. It is the strongest competitor because it is the exact solution of the per-slot problem our method only approximates, and it carries the proven $[O(1/V), O(V)]$ tradeoff. We run it on identical traces, task arrivals and $V$ values, with the same convergence tolerance used to define our "exact" reference. Its reported 1.05 ms delay comes from block fading. We found no reported result on Lacuna for this controller, or any offloading controller, under 3GPP TDL non-stationary fading, so we will establish that number ourselves.
Beating it means two things. First, we must match its average power at equal average delay, within the predicted $\bar\epsilon/V$ gap, while using a fixed one broadcast round per slot rather than its user-dependent sweep count. Second, under TDL speed transitions, we must show lower delay and fewer retransmission-driven queue spikes, because its Shannon-rate planning is optimistic.
Several ablations isolate each ingredient. One-sweep warm-started Gauss-Seidel separates "simultaneous" from "one step with warm start." Proximal Jacobian ADMM run to convergence per slot separates "one step" from "Jacobian." Dual decomposition with a subgradient price update is the other simultaneous scheme. The full method with Shannon rates instead of conformal rates isolates the channel ingredient. For link adaptation we compare conformal rates against 3GPP outer-loop link adaptation (OLLA) and a fixed SNR backoff margin. On TDL channels, Batra et al. report 10.3% (±0.012) empirical block error at a 10% target and an 82.4% singleton rate for their conformal method, versus 75.3% for OLLA. We expect to reproduce these before plugging the method into the controller.
All methods share the same proximal and step-size tuning budget: a grid over the same number of settings, chosen on held-out traces.
Why not just…?
Why not just run one warm-started Gauss-Seidel sweep per slot? This is the most serious threat. If the server holds all channel and queue information centrally, one sequential sweep may track the optimum as well as one Jacobi step or better, because Gauss-Seidel typically contracts faster per iteration. The case for simultaneous updates rests on the distributed setting, where each device owns its own variables. There, a sequential sweep needs $K$ ordered exchanges per slot for $K$ users, while a Jacobi step needs one broadcast and one reply. The ablation reports both tracking error and signaling rounds. If centralized one-sweep Gauss-Seidel matches our power-delay curve at negligible solver cost, we will say so. The contribution then narrows to the distributed case and the conformal rate correction.
Why not just use dual decomposition, which is already parallel? Dual price updates are simultaneous, and they are the natural textbook answer. The concern is that with one iteration per slot, the primal allocation produced by a subgradient price step can violate the bandwidth budget and oscillate. The proximal term exists to suppress exactly this. We measure bandwidth constraint violation, oscillation, and tracking error for both. If dual decomposition with a tuned step matches, the proximal Jacobian machinery is unnecessary, and the paper's claim reduces to "one-step tracking plus calibrated rates."
Why not just plan with a fixed SNR backoff or OLLA instead of conformal rates? A fixed margin is cheap and may be adequate at a single speed. Batra et al.'s results show OLLA losing efficiency relative to conformal selection across speed transitions. Their guarantee is only marginal, though, and their learning rate assumes Jakes Doppler, so a well-tuned backoff could be competitive on TDL profiles that depart from Jakes. Our experiment reports queue delay and delivered-bit shortfall for all three rate rules across each speed regime separately, which will show whether the guarantee matters for queue stability or only for link-level statistics.
Proposed Experiments
The simulator follows the Mao et al. multi-user MEC setup: stochastic task arrivals, DVFS-controlled local CPUs, and a shared uplink bandwidth budget. Channels are replaced by 3GPP TDL-A, TDL-B and TDL-C traces with user speeds switching among 3, 120 and 500 km/h. We keep i.i.d. Rayleigh block fading as a sanity condition, where we should recover Mao et al.'s qualitative tradeoff.
For each $V$ we report average device power, average queue length and delay, the 99th-percentile delay, block error rate, delivered-bit shortfall, per-slot solver wall time and energy on phone-class hardware, and signaling rounds. The user count ranges from 4 to 64, which tests the claim that sequential cost grows with $K$ while ours does not.
Two checks are decisive. First, plot the power gap between our one-step controller and the exact Gauss-Seidel controller against measured per-slot channel drift. The theory predicts a roughly linear relation whose slope falls as the proximal weight is tuned toward stronger contraction. If the gap does not track drift, or the queues destabilize at 500 km/h, the inexact-Lyapunov argument fails. Second, at matched average delay, the full method must beat the exact controller with Shannon-rate planning on 99th-percentile delay under speed transitions. It must also beat its own Shannon-rate ablation, which shows that the calibrated rate, not the parallel update alone, removes the retransmission-driven queue spikes.
References
- Power-Delay Tradeoff in Multi-User Mobile-Edge Computing Systems
- Stochastic Joint Radio and Computational Resource Management for Multi-User Mobile-Edge Computing Systems
- Distribution-Free Adaptive Modulation and Coding via Online Conformal Inference under Non-Stationary Fading
- A Proximal Jacobian ADMM Approach for Fast Massive MIMO Signal Detection in Low-Latency Communications
- Joint energy-bandwidth allocation for multi-user channels with cooperating hybrid energy nodes
- Generalized Higher-Order Tensor Decomposition via Parallel ADMM
- Delay Minimization in Pinching-Antenna-Enabled NOMA-MEC Networks