⚡ electrical engineering

Optimized Power Control for Multi-User Integrated Sensing and Edge AI

This paper proposes an optimized power control framework for multi-user Integrated Sensing and Edge AI systems that offloads features via analog AirComp, deriving closed-form transceiver designs for TDM and FDM settings based on two performance proxies—computation-optimal and decision-optimal—to balance aggregation distortion and inter-class separability.

Original authors: Biao Dong, Bin Cao

Published 2026-05-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Biao Dong, Bin Cao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: A Team of Detectives

Imagine a group of detectives (the devices) trying to solve a mystery about a specific target (like identifying a person or an object). Each detective has a unique vantage point and takes a photo. However, they can't send the whole high-resolution photo to the headquarters (the Access Point) because the mail service is too slow and expensive.

Instead, they each write a short, summarized note (a feature) about what they see. They need to send these notes to the headquarters so a central AI can combine them and make a final decision.

The problem is: How do they send these notes so the headquarters gets the clearest possible picture of the truth, even if the mail service is noisy or the detectives have different amounts of energy?

This paper proposes a smart way to manage the "mailing" process using a technique called AirComp (Over-the-Air Computation).

The Magic Trick: AirComp

Usually, if five people try to talk to a radio station at the same time, it's just a mess of noise. But AirComp is like a special kind of radio where the signals naturally blend together.

Think of it like a choir. If everyone sings their part at the exact same time and volume, the microphone picks up the sum of all the voices. In this system, the devices don't send their notes one by one; they shout them all at once. The radio waves mix in the air, and the headquarters receives a single "mixture" that represents the combined data.

The Two Strategies: "Perfect Math" vs. "Smart Guessing"

The researchers realized that simply making the math perfect isn't always the best way to make a good decision. They tested two different strategies for how the detectives should shout their notes:

1. The "Computation-Optimal" Strategy (The Perfectionist)

The Goal: Make the received mixture look exactly like the sum of the original notes.
The Analogy: Imagine a sound engineer trying to record a choir. They want the recording to be a 100% perfect copy of the live sound. They adjust the volume of every singer so that no one is too quiet or too loud, minimizing any static or distortion.
The Result: This minimizes the "noise" in the data. It's great if you just want the raw data to be clean.

2. The "Decision-Optimal" Strategy (The Detective)

The Goal: Make sure the final decision (e.g., "Is this a cat or a dog?") is as accurate as possible, even if the data isn't mathematically perfect.
The Analogy: Imagine a detective who knows that for this specific case, the shape of the ear is the most important clue, but the color of the fur doesn't matter much. They tell the singers: "Singers with the ear notes, shout louder! Singers with the fur notes, whisper." They intentionally distort the "unimportant" parts to make the "important" parts stand out clearly against the background noise.
The Result: This maximizes the separability between different categories. It cares less about the raw numbers and more about whether the AI can tell the difference between Class A and Class B.

The Findings: When to Use Which?

The paper tested these strategies in two different "mailing" scenarios:

Scenario A: The Slow, Steady Road (Time-Division)

The Setup: The detectives send their notes one after another in time slots (like taking turns speaking).
The Result: In this scenario, both strategies ended up giving the exact same instructions. It turns out that when you take turns, the best way to shout is the same whether you want perfect math or a smart guess. The "Perfectionist" and the "Detective" agree on the plan.

Scenario B: The Busy Highway (Frequency-Division)

The Setup: The detectives send their notes all at once but on different radio frequencies (like different lanes on a highway). Some lanes are bumpy (bad signal), and some are smooth (good signal).
The Result: Here, the two strategies diverged.
- The "Perfectionist" (Computation-Optimal) tried to fix the bumpy lanes by boosting power evenly based on the noise.
- The "Detective" (Decision-Optimal) realized that some lanes carried the most important clues. They dumped all the extra power into those specific lanes to make the critical clues pop, even if it meant the less important lanes got a bit noisier.
- The Winner: In this busy highway scenario, the "Detective" strategy won. By focusing power on the most discriminative information, the AI made better decisions.

The Conclusion

The paper proves that for systems where devices need to sense the world and send data to an AI for a quick decision:

You can mathematically prove how much "noise" hurts the final decision.
If you are sending data over different frequencies (which is common in modern 6G networks), you shouldn't just try to make the data "clean." You should try to make the decision easier.
By intelligently allocating power to the most important features (like the "ear shape" in our analogy), you can get a much smarter AI result than just trying to fix the signal noise.

In short: Don't just fix the signal; fix the signal to help the AI make the right choice.

Technical Summary: Optimized Power Control for Multi-User Integrated Sensing and Edge AI

Problem Statement
This work addresses the challenges within Integrated Sensing and Edge AI (ISEA) systems, a paradigm emerging for 6G networks to support latency-sensitive applications like smart factories and digital twins. In a multi-user ISEA system, distributed devices perform target sensing, extract local features, and offload these features to an Access Point (AP) for collaborative inference. A critical bottleneck in this process is the transmission of high-dimensional feature vectors, which strains communication resources. While Over-the-Air Computation (AirComp) offers a promising solution by leveraging signal superposition for analog aggregation, existing designs often lack a unified framework to characterize the relationship between AirComp-induced errors and the ultimate task performance (inference accuracy). Specifically, conventional "computation-optimal" designs minimize aggregation distortion without considering task characteristics, while "task-aware" approaches often rely on proxies without rigorous theoretical justification or optimal transceiver derivation.

Methodology
The authors propose a theoretical framework linking AirComp errors to inference performance through two distinct performance proxies:

Computation-Optimal Proxy: Based on minimizing the Mean Squared Error (MSE) of the aggregated signal. The paper establishes a theoretical lower bound on classification accuracy, demonstrating that minimizing computation MSE is a sufficient condition for improving inference performance.
Decision-Optimal Proxy: Based on maximizing the inter-class Mahalanobis Distance (MD). Derived from Bayesian hypothesis testing, this proxy aims to maximize the separability between classes in the feature space, directly targeting the reduction of classification error probability.

The study derives optimal transceiver designs (specifically closed-form power allocation) for two transmission schemes:

Time-Division Multiplexing (TDM): Features are transmitted sequentially over time slots. The optimization problems for both proxies are solved independently per slot.
Frequency-Division Multiplexing (FDM): Features are transmitted over subcarriers. Due to frequency-selective fading and joint power constraints, the optimization problems are non-convex. The authors employ the Lagrange duality method, leveraging the time-sharing condition to solve for optimal power allocation via subgradient methods.

Key Contributions

Theoretical Proxies: The paper formally establishes the relationship between AirComp error and inference performance, justifying the use of both computation-optimal (MSE minimization) and decision-optimal (MD maximization) proxies.
Closed-Form Solutions:
- For TDM, the optimal power allocation reveals a threshold-based structure. Interestingly, under quasi-static channels, both the computation-optimal and decision-optimal designs yield identical structural solutions.
- For FDM, the optimal designs exhibit dual-decomposition structures. Here, the two proxies diverge: the decision-optimal design allocates power based on discriminative importance (feature separability), whereas the computation-optimal design follows variance-based regularization.
System Design: The work provides specific algorithms for transceiver design that balance feature importance, channel conditions, and power budgets.

Experimental Results
Numerical results, validated using a Support Vector Machine (SVM) for four-class human posture recognition, demonstrate the following:

Low SNR: All schemes degrade to random guessing.
High SNR: Most schemes approach a saturation upper bound, except for Channel Inversion, which is limited by total power constraints.
Moderate SNR: The computation-optimal scheme outperforms others, benefiting from non-uniform resource allocation.
User Scaling: As the number of users ( $K$ ) increases, class uncertainty decreases, and all schemes converge to a common performance bound.
Subcarrier Scaling: As the number of subcarriers ( $N$ ) increases, the performance gap between the decision-optimal and other schemes narrows due to reduced frequency-domain heterogeneity. However, a persistent gap remains because the decision-optimal design specifically targets discriminative subcarriers.

Significance and Claims
The paper claims to provide a foundational framework for optimizing ISEA systems by moving beyond generic signal reconstruction to task-oriented optimization. It highlights that while computation-optimal designs are effective in quasi-static environments, decision-optimal designs offer additional inference gains in frequency-selective channels by aligning resource allocation with the discriminative importance of features. The work validates that a unified approach characterizing the trade-off between AirComp errors and task performance is essential for overcoming communication bottlenecks in next-generation edge AI systems.