⚡ electrical engineering

Joint Activity Detection and Channel Estimation for Massive Random Access Using SBL and SCA

This paper proposes a covariance learning-based sparse Bayesian learning (CL-SCA) method that utilizes successive convex approximation to efficiently solve the joint activity detection and channel estimation problem in massive random access scenarios, demonstrating superior performance over existing approaches.

Original authors: Esa Ollila, Majdoddin Esfandiari, Daniel P. Palomar

Published 2026-04-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Esa Ollila, Majdoddin Esfandiari, Daniel P. Palomar

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

Imagine a massive stadium filled with 300 people (devices), but only a tiny crowd of 20 to 40 people (active devices) actually want to speak at any given moment. Everyone is trying to shout their message to a single microphone tower (the Base Station) at the same time, but they can't talk over each other.

In the world of "Massive Machine-Type Communication" (like smart sensors or IoT devices), this is the daily struggle. The tower needs to solve two problems simultaneously:

Who is shouting? (Activity Detection)
What does their voice sound like? (Channel Estimation)

The problem is that the "shouts" are messy, overlapping, and the tower has to figure out who is speaking without anyone raising their hand first (this is called "grant-free" access).

The Old Way: The Slow Detective

Previously, researchers tried to solve this using methods like M-SBL (which uses an Expectation-Maximization algorithm). Think of this as a very thorough but incredibly slow detective. The detective looks at the noise, guesses who might be talking, checks their guess, and then guesses again. They repeat this process hundreds of times until they are sure.

Pros: Very accurate.
Cons: It takes forever. By the time the detective finishes, the conversation is over.

Another method, CL-MP, is like a greedy detective who picks the loudest voice immediately and stops after a few guesses.

Pros: Super fast.
Cons: If there are many people talking, this detective gets confused and misses the quieter ones. Also, once they pick a person, they don't refine their understanding of that person's voice very well.

The New Solution: The "Smart Crowd Manager" (CL-SCA)

The authors of this paper propose a new method called CL-SCA (Covariance Learning with Successive Convex Approximation).

Here is how it works, using a simple analogy:

1. The "Covariance" (Listening to the Rhythm)
Instead of trying to isolate every single voice immediately, the tower listens to the overall rhythm and pattern of the noise. It calculates a "map" of how the signals overlap. This map tells the tower, "Hey, there is definitely activity coming from these specific directions, even if we can't hear the words yet."

2. The "Successive Convex Approximation" (The Smooth Path)
The math behind this is tricky because the problem is like trying to find the lowest point in a landscape full of hills and valleys (non-convex).

The Old Way: Tries to jump from hill to hill, often getting stuck or taking a very long path.
The New Way (SCA): Imagine you are blindfolded on a mountain. Instead of guessing the whole path, you take a small step, look at the ground right in front of you, and approximate the terrain as a smooth, flat slope. You walk down that slope, then stop, look again, and approximate the new spot as a smooth slope.
By doing this in parallel (all 300 potential speakers are checked at the same time, not one by one), the algorithm zooms down the mountain much faster than the old detective.

3. The Result

Speed: It is roughly twice as fast as the previous best optimization methods and much faster than the slow detective.
Accuracy: It catches the "shouters" (active devices) just as well as, or better than, the slow methods.
Voice Clarity: Once it knows who is talking, it estimates their "voice" (channel) better than the greedy method.

Why This Matters

In a world where millions of devices (smart meters, self-driving cars, factory sensors) need to connect instantly without waiting for permission, speed and accuracy are everything.

The Old Detective was too slow for real-time traffic.
The Greedy Detective was fast but missed people and got the voices wrong.
The New CL-SCA Manager is the Goldilocks solution: It's fast enough for real-time use, smart enough to handle a crowded stadium, and accurate enough to hear every voice clearly.

In a nutshell: The paper introduces a new mathematical "shortcut" that lets the base station figure out who is talking and what they sound like, much faster and more reliably than before, by smoothing out the complex math and solving it all at once.

1. Problem Statement

The paper addresses the Joint Device Activity Detection and Channel Estimation (JADCE) problem in the context of Massive Machine-Type Communications (mMTC).

Scenario: A single-cell uplink scenario where a Base Station (BS) with $M$ antennas communicates with $N$ single-antenna devices.
Challenge: The system operates under grant-free random access, meaning a massive number of devices ( $N \gg L$ , where $L$ is the pilot sequence length) attempt to connect sporadically. Only a small subset ( $K \ll N$ ) is active at any given coherence interval.
Goal: Given the received signal matrix $Y$ $Y$ , the known pilot matrix $A$ $A$ , and noise power, the system must simultaneously:
1. Identify which devices are active (Activity Detection).
2. Estimate the channel vectors for the active devices (Channel Estimation).
Limitations of Existing Methods: Traditional Compressed Sensing (CS) approaches require the number of active devices to be smaller than the pilot length. Covariance Learning (CL) methods relax this constraint but often rely on the Expectation-Maximization (EM) algorithm (e.g., M-SBL), which suffers from slow convergence and high computational complexity for large $N$ .

2. Methodology

The authors propose a Covariance Learning-based Sparse Bayesian Learning (CL-SCA) framework. The methodology is divided into three main stages:

A. Signal Model and Covariance Formulation

The received signal is modeled as a Multiple Measurement Vector (MMV) problem. The columns of the received signal matrix $Y$ are modeled as independent and identically distributed (i.i.d.) circular Gaussian vectors with a covariance matrix $\Sigma$ :
$\Sigma = A\Gamma A^H + \sigma^2 I$
Where:

$\Gamma = \text{diag}(\gamma)$ is a diagonal matrix of signal powers.
$\gamma_n = \alpha_n \rho_n \beta_n$ represents the effective power of device $n$ (incorporating activity indicator $\alpha_n$ , transmit power $\rho_n$ , and large-scale fading $\beta_n$ ).
Since only $K$ devices are active, $\gamma$ is a sparse vector.

B. Optimization via Successive Convex Approximation (SCA)

The core contribution is the development of an efficient algorithm to estimate the sparse power vector $\gamma$ by minimizing the scaled negative log-likelihood function (LLF):
$\ell(\gamma) = \text{tr}((A\Gamma A^H + \sigma^2 I)^{-1}S) + \log |A\Gamma A^H + \sigma^2 I|$
Where $S$ is the Sample Covariance Matrix (SCM).

Problem: The objective function is non-convex because it is the sum of a convex term and a concave term (the log-determinant).
Solution (SCA): The authors decompose the objective into a convex part $g(\gamma)$ and a non-convex part $f(\gamma)$ . They linearize the non-convex part $f(\gamma)$ around the current iterate $\gamma^k$ using a first-order Taylor expansion.
Parallel Updates: Unlike coordinate-wise optimization, the SCA framework allows all $N$ variables ( $\gamma_1, \dots, \gamma_N$ ) to be updated simultaneously and in parallel at each iteration.
Closed-Form Solution: The authors derive a closed-form minimizer for the approximated convex sub-problem for each device, ensuring a significant reduction in computational steps per iteration compared to EM.
Convergence: A smoothing operation with a diminishing step-size is applied to guarantee convergence to a stationary point.

C. Joint Detection and Estimation

Once the sparse power vector $\hat{\gamma}$ is estimated:

Activity Detection: Active devices are identified using either a threshold rule ( $\hat{\gamma}_n \geq \gamma_{th}$ ) or a Top-K rule (selecting the $K$ largest entries).
Channel Estimation: Using the estimated support (active devices), the effective channel matrix $\hat{X}$ is computed via Empirical Bayesian estimation (Type-II Maximum Likelihood):
$\hat{X} = \hat{\Gamma} A^H (A\hat{\Gamma}A^H + \sigma^2 I)^{-1} Y$

3. Key Contributions

Novel Algorithm (CL-SCA): Development of a fast, iterative algorithm for JADCE that utilizes the SCA framework to solve the non-convex covariance learning problem.
Parallelization: Unlike the standard M-SBL (EM algorithm) which updates variables sequentially or slowly, the proposed method updates all power estimates in parallel, drastically reducing runtime.
Theoretical Guarantees: The algorithm is proven to converge to a stationary point of the non-convex objective function.
Complexity Analysis: The per-iteration complexity is $O(NL^2)$ , which is comparable to Coordinate-Wise Optimization (CWO) and greedy methods, but with superior convergence speed.

4. Simulation Results

The authors evaluated the proposed method against three competitors: M-SBL (EM), Coordinate-Wise Optimization (CWO), and Covariance Matching Pursuit (CL-MP).

Activity Detection Performance (Probability of Missed Detection - PMD):
- CL-SCA and CWO achieved the best performance across all tested scenarios (varying pilot lengths $L$ , antennas $M$ , and active devices $K$ ).
- CL-SCA significantly outperformed the standard M-SBL (EM) algorithm, especially with short pilot lengths ( $L=20$ ), where the performance gap was noticeable.
- CL-MP (a greedy method) performed well only when $K$ was small and known; its performance degraded as $K$ increased.
Channel Estimation Performance (NMSE):
- CL-SCA provided state-of-the-art Normalized Mean Square Error (NMSE) performance.
- Even in cases where CL-MP had slightly better activity detection, its channel estimation performance was inferior to CL-SCA.
Computational Efficiency (Running Time):
- CL-SCA was approximately twice as fast as CWO.
- CL-SCA was roughly one order of magnitude (10x) faster than the M-SBL (EM) algorithm, primarily because EM requires hundreds of iterations to converge.
- While CL-MP was the fastest computationally (due to its greedy nature and fixed $K$ iterations), it failed to match the accuracy of CL-SCA in channel estimation and robustness in activity detection.

5. Significance

This paper presents a critical advancement for grant-free massive random access systems, which are foundational for 5G/6G IoT networks.

Scalability: The proposed CL-SCA algorithm makes it feasible to handle massive numbers of devices ( $N$ ) with short pilot sequences ( $L$ ) without the prohibitive computational cost of traditional Bayesian methods.
Robustness: It offers a robust solution that does not strictly rely on knowing the exact number of active users ( $K$ ) beforehand, unlike greedy approaches.
Practicality: By balancing high accuracy in both activity detection and channel estimation with low latency (fast convergence), the CL-SCA method is highly suitable for real-time implementation in future mMTC networks.

In conclusion, the authors successfully bridge the gap between the high accuracy of Sparse Bayesian Learning and the computational efficiency required for massive-scale systems, establishing CL-SCA as a superior alternative to existing optimization-based JADCE methods.