⚡ electrical engineering

Circulant ADMM-Net for Fast High-resolution DoA Estimation

This paper introduces CADMM-Net and CHADMM-Net, two deep neural networks that leverage structured deep unfolding of the ADMM algorithm with circulant and Hermitian-circulant matrices to achieve fast, high-resolution direction of arrival estimation with significantly reduced computational complexity and memory footprint while maintaining competitive performance.

Original authors: Youval Klioui

Published 2026-07-23

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Youval Klioui

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 you are standing in a crowded room, trying to figure out exactly where everyone is talking. You have a special microphone array that can hear the sound waves, but you only get to listen for a split second—a single snapshot. In the world of physics and engineering, this is called "Direction of Arrival" (DoA) estimation. It's the superpower that lets self-driving cars "hear" where other cars, pedestrians, or obstacles are coming from, even when they are too far away to see clearly. The problem is, doing this mathematically is like trying to solve a massive, tangled knot of equations in your head while running a marathon. Traditional methods are either too slow to be useful in a moving car, or they get confused when there's only a tiny bit of data to work with. Scientists have been trying to build a "smart shortcut" using artificial intelligence to solve this knot instantly, but even those shortcuts have been too heavy and slow for the tiny computers inside a car.

This paper introduces two new, super-lightweight neural networks called CADMM-Net and CHADMM-Net. Think of these networks as a team of detectives who have learned a magical trick: instead of checking every single clue in a giant, messy file cabinet (which takes forever), they realize the clues are arranged in a perfect, repeating circle. By recognizing this circular pattern, they can use a mathematical "magic wand" (called a Fast Fourier Transform) to solve the mystery in a flash. The authors found that by forcing their AI to only look for these circular patterns, they could shrink the memory the AI needs by a huge amount and make it run thousands of times faster, all without losing its ability to pinpoint exactly where the sound is coming from. It's like swapping a heavy, slow-moving tank for a nimble, high-speed drone that can still hit the target with perfect accuracy.

The Problem: The Heavy Math of Listening

To understand why this is a big deal, imagine you are trying to find the location of a few friends in a dark room using only a handful of microphones. The math to figure this out is called "LASSO." It's a method that tries to find the simplest explanation for the noise it hears. The problem is, the standard way to solve LASSO is like trying to climb a mountain by taking one tiny, careful step at a time. You might need to take a hundred steps just to get to the top. In a car driving at 60 miles per hour, you don't have time for a hundred steps; you need the answer now.

Scientists tried to speed this up using "Deep Unfolding." Imagine taking that slow, step-by-step climbing process and turning it into a pre-planned slide. You train a neural network to mimic the steps of the climb, but instead of stopping after one step, it slides down the whole mountain in just a few jumps. This is much faster. However, the existing "slides" (like ADMM-Net) were still too heavy. They required storing a massive grid of numbers (a matrix) for every single step, which is like carrying a library of encyclopedias in your backpack just to find a single book. For a car computer with limited space and power, this is a dealbreaker.

The Solution: The Circular Shortcut

The authors of this paper asked a simple question: "Do we really need to carry the whole library?" They realized that in many common setups, the math behind the problem has a special property: it repeats in a circle. This is called a "circulant" structure.

Think of a standard dictionary of clues as a giant, messy spreadsheet where every cell is different. To solve the problem, the computer has to multiply this whole spreadsheet by a vector of numbers. This is slow and memory-hungry. But, if that spreadsheet is actually a "circulant" matrix, it means the rows are just shifted versions of each other, like a pattern on a rolling drum.

The authors built two new networks:

CADMM-Net: This network assumes the pattern is a perfect circle. Instead of storing a massive grid of numbers, it only needs to remember a single list of numbers (a vector) that defines the circle.
CHADMM-Net: This is an even more specialized version that assumes the circle has a mirror symmetry (Hermitian-circulant). It cuts the memory requirement in half again.

By using this "circular" assumption, the networks can use a mathematical tool called the Fast Fourier Transform (FFT). If the standard method is like walking through a forest one tree at a time, the FFT is like teleporting through the forest. It turns a slow, heavy calculation into a lightning-fast one.

What They Found

The researchers tested these new networks against the old heavyweights (like ADMM-Net, LISTA, and TLISTA) and the traditional slow methods (ISTA and ADMM). They simulated a scenario with 30 microphones and up to 8 different sound sources, testing everything from very quiet (0 dB) to very loud (35 dB) environments.

Here is what the simulations showed:

Speed and Size: The new networks are incredibly efficient. While the old ADMM-Net needed to store about 65,000 numbers per layer (for a dictionary size of 256), CADMM-Net only needed about 2,500, and CHADMM-Net needed even fewer. In terms of speed, the new networks performed their calculations in roughly $2N \log(N)$ or $3N \log(N)$ operations, compared to the $N^2$ operations required by the older methods. For a dictionary size of 256, this means the new networks are roughly 16 times faster per step.
Accuracy: Despite being so much smaller and faster, they didn't lose their "ears." In the tests, CADMM-Net and CHADMM-Net performed just as well as the heavy, slow networks at detecting where the sounds were coming from. They measured this using "detection rate" (how often they found the sound) and "RMSE" (how close their guess was to the real angle).
The Trade-off: The authors noted a tiny trade-off. CHADMM-Net, which saves the most memory, is slightly more computationally complex to run than CADMM-Net because of the extra symmetry rules it follows. However, the difference in performance was so small that the memory savings were worth it.

The Verdict

The paper doesn't claim to have solved every problem in the universe, but it suggests a very strong path forward. By proving that you can force a neural network to respect the circular nature of the math, you can build a "DoA estimator" that is small enough to fit in a car's computer and fast enough to react in real-time, without sacrificing the ability to hear a whisper in a storm.

The authors explicitly ruled out the idea that you need massive, unstructured matrices to get high-resolution results. They showed that the "heavy" approach is unnecessary. They also demonstrated that while traditional iterative methods (like running the math 30 times manually) are slow, and older deep learning methods are heavy, this new "circulant" approach hits the sweet spot.

In the end, the paper suggests that for automotive settings—where you have limited snapshots and limited computing power—these new networks are a game-changer. They offer a way to see (or rather, hear) the world with high precision, using a fraction of the resources previously thought necessary. It's a reminder that sometimes, the fastest way to solve a complex problem isn't to work harder, but to realize the problem is actually a circle all along.

Technical Summary: Circulant ADMM-Net for Fast High-resolution DoA Estimation

Problem Statement
Fast and high-resolution Direction of Arrival (DoA) estimation is critical for advanced driver-assistance systems (ADAS). However, automotive settings present unique challenges: the number of available snapshots is often severely limited (frequently a single snapshot), and the array geometry is typically sparse (e.g., Minimum Redundancy Arrays) rather than Uniform Linear Arrays (ULAs).

Traditional subspace-based methods like MUSIC and ESPRIT are inadequate in single-snapshot scenarios unless Forward-Backward Spatial Smoothing (FBSS) is applied, which reduces array aperture and restricts geometry to subarrays with identical structures. Gridless methods (e.g., Atomic Norm Minimization) and Maximum Likelihood estimators suffer from prohibitive computational costs ( $O(M^3)$ or extensive grid searches). While the Least Absolute Shrinkage and Selection Operator (LASSO) framework offers a grid-based sparse recovery alternative that handles single snapshots and arbitrary geometries, standard solvers like Iterative Shrinkage Thresholding Algorithm (ISTA) and Alternating Direction Method of Multipliers (ADMM) require many iterations (often >100) to converge and rely on manually tuned hyperparameters dependent on unknown noise variance and signal sparsity.

Deep unfolding (or unrolling) has been proposed to address the iteration count and hyperparameter tuning issues by training neural networks that mimic iterative algorithms. However, existing deep-unfolded architectures, such as LISTA and ADMM-Net, suffer from high computational complexity ( $O(N^2)$ or $O(N^3)$ per layer) and large memory footprints ( $O(N^2)$ parameters), making them impractical for real-time automotive inference on limited hardware. Structured variants like TLISTA and THADMM-Net reduce parameter counts but often retain high computational complexity or rely on restrictive assumptions (e.g., Gaussian dictionaries) that do not hold for frequency estimation.

Methodology
The paper introduces CADMM-Net and CHADMM-Net, two deep neural networks based on a structured deep unfolding of the ADMM algorithm. The core methodology relies on exploiting the structural properties of the Gram matrix ( $A^HA$ ) within the LASSO formulation.

Theoretical Motivation: The authors prove that under a uniform frequency grid and specific array geometry assumptions (specifically $\gamma = 1/2$ ), the matrix $(A^HA + \rho I)^{-1}$ , which is central to the ADMM update step, becomes a circulant matrix. If the matrix is also Hermitian (which holds for $A^HA$ ), the inverse is a Hermitian-circulant matrix.
Network Architecture:
- CADMM-Net: Imposes a circulant constraint on the learnable weight matrix $W$ in the ADMM-Net architecture. A circulant matrix is fully defined by a single vector of length $N$ . The matrix-vector multiplication is replaced by Fast Fourier Transform (FFT) and Inverse FFT (IFFT) operations.
- CHADMM-Net: Imposes both Hermitian and circulant constraints. A Hermitian-circulant matrix is defined by approximately $\lfloor N/2 \rfloor + 1$ parameters.
Computational Efficiency: By utilizing the convolution theorem, the matrix operations in each layer are performed via FFT/IFFT. This reduces the computational complexity per layer from $O(N^2)$ (in ADMM-Net) to $O(N \log N)$ . Specifically, CADMM-Net requires $2N \log N$ operations, and CHADMM-Net requires $3N \log N$ operations per layer.
Training: The networks are trained using a "support-aware" loss function. To address the non-smoothness of the standard Normalized Mean Squared Error (NMSE) regarding support shifts, the output and ground truth are convolved with a Laplacian kernel before computing the loss. The networks are initialized with weights derived from the corresponding iterative ADMM steps.

Key Contributions

Novel Architectures: Introduction of CADMM-Net and CHADMM-Net, which apply circulant and Hermitian-circulant constraints to the deep unfolding of ADMM.
Complexity Reduction: Demonstration that these constraints reduce the memory footprint from $O(N^2)$ to $O(N)$ (or $\approx N/2$ for CHADMM-Net) and the computational complexity per layer from $O(N^2)$ to $O(N \log N)$ .
Performance Benchmarking: Comprehensive evaluation against state-of-the-art baselines, including classical iterative methods (ISTA, ADMM), standard deep-unfolded networks (LISTA, ADMM-Net), and structured variants (TLISTA, THADMM-Net, THLISTA).
Experimental Validation: Testing conducted on both Uniform Linear Array (ULA) and Sparse Linear Array (SLA) scenarios with single-snapshot data.

Results
The experiments utilized $N=256$ dictionary size and $M=30$ array elements.

Detection Rate: In SLA settings, CHADMM-Net and THADMM-Net maintained the highest detection rates across the Signal-to-Noise Ratio (SNR) range. In ULA settings, CADMM-Net outperformed others at low SNR (<10 dB), while CHADMM-Net and THADMM-Net dominated at higher SNRs. All deep-unfolded networks significantly outperformed their classical iterative counterparts (ISTA/ADMM) with the same number of steps.
Accuracy (RMSE & NMSE): Structured networks (CADMM-Net and CHADMM-Net) exhibited performance competitive with the unstructured ADMM-Net. In SLA settings, the NMSE performance was nearly identical to ADMM-Net across all SNRs. In ULA settings, CHADMM-Net showed a slight advantage over ADMM-Net at SNRs above 15 dB.
Efficiency: The proposed networks achieved a drastic reduction in parameters (e.g., ~3,930 parameters for CHADMM-Net vs. ~1.9 million for ADMM-Net) and Complex Floating Point Operations (CFLOPS) per inference, making them suitable for resource-constrained environments.

Significance and Claims
The paper claims that by leveraging the circulant nature of the Gram matrix in specific DoA estimation contexts, it is possible to construct deep-unfolded networks that are computationally efficient and memory-light without sacrificing the high-resolution performance of the original ADMM algorithm. The authors assert that CADMM-Net and CHADMM-Net offer a viable solution for real-time, high-resolution DoA estimation in automotive applications where single snapshots and sparse arrays are the norm, providing a competitive alternative to existing deep unfolding methods while drastically reducing the hardware requirements for inference.

The Problem: The Heavy Math of Listening

The Solution: The Circular Shortcut

What They Found

The Verdict

Technical Summary: Circulant ADMM-Net for Fast High-resolution DoA Estimation

More like this