HYPERDOA: Robust and Efficient DoA Estimation using Hyperdimensional Computing

Basic Information

• Paper ID: 2510.10718
• Title: HYPERDOA: Robust and Efficient DoA Estimation using Hyperdimensional Computing
• Authors: Rajat Bhattacharjya, Woohyeok Park, Arnab Sarkar, Hyunwoo Oh, Mohsen Imani, Nikil Dutt
• Institutions: University of California, Irvine, USA; Kookmin University, Seoul, South Korea; Indian Institute of Technology, Kharagpur, India
• Classification: eess.SP cs.AI cs.AR cs.SC
• Publication Date: October 12, 2025 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2510.10718

Abstract

Direction of Arrival (DoA) estimation techniques face critical trade-offs: classical methods lack accuracy under low signal-to-noise ratio (SNR) conditions, while modern deep learning methods are too energy-intensive and opaque for resource-constrained safety-critical systems. This paper proposes HYPERDOA, a novel estimator based on hyperdimensional computing (HDC). The framework introduces two distinct feature extraction strategies for its HDC pipeline—mean spatial lag autocorrelation and spatial smoothing—and recasts DoA estimation as a pattern recognition problem. The method leverages HDC's inherent noise robustness and transparent algebraic operations to circumvent expensive matrix decomposition in classical methods and the "black-box" nature of deep learning approaches, respectively. Evaluation demonstrates that HYPERDOA achieves approximately 35.39% higher accuracy than state-of-the-art methods in low SNR coherent source scenarios and is approximately 93% more energy-efficient than competing neural baselines on an embedded NVIDIA Jetson Xavier NX platform.

Research Background and Motivation

Problem Definition

Direction of Arrival (DoA) estimation is a critical technique in array signal processing used to determine the angle at which incident signals strike an antenna array, thereby localizing signal sources. This technique has widespread applications in autonomous vehicle localization, biomedical sensor signal analysis, seismic monitoring systems, and related fields.

Problem Importance

DoA estimation is crucial for safety-critical and resource-constrained edge device applications, particularly in scenarios requiring real-time processing, low power consumption, and high reliability. Both traditional and deep learning approaches exhibit significant limitations, necessitating novel solutions.

Limitations of Existing Methods

1. Classical Methods (MUSIC, Root-MUSIC, ESPRIT):
   • Accuracy significantly degrades under low SNR, coherent sources, or model mismatch conditions
   • Rely on subspace decomposition requiring eigenvalue decomposition (EVD) or singular value decomposition (SVD)
   • High computational complexity, difficult to deploy efficiently on low-power embedded systems
2. Deep Learning Methods (DeepMUSIC, SubspaceNet):
   • "Black-box" nature makes analysis, verification, and trust difficult
   • Ignore system-level complexity, incurring high costs in FLOPs, training overhead, and device power consumption
   • Unsuitable for edge computing deployment

Research Motivation

To address these issues, the authors propose using hyperdimensional computing (HDC), a brain-inspired computing paradigm. HDC provides inherent fault tolerance through distributed representations of high-dimensional vectors (hypervectors), employing simple parallel algebraic operations that combine robustness with computational efficiency.

Core Contributions

1. Complete HDC Pipeline Design: Recasts DoA estimation as a pattern recognition problem using associative memory for angle detection through similarity search, completely bypassing expensive matrix decomposition at inference time
2. Two Novel Feature Extraction Strategies:
   • Mean spatial lag autocorrelation
   • Spatial smoothing Delivering 35.39% higher accuracy than state-of-the-art methods in challenging low SNR and coherent source scenarios
3. System-Level Energy Evaluation: Empirical evaluation on embedded NVIDIA Jetson Xavier NX platform confirms HYPERDOA is 92.93% more energy-efficient than modern neural baselines
4. Multi-Label Learning Adaptation: Improves the OnlineHD learning rule for DoA estimation's multi-source nature, addressing the problem where a single sample corresponds to multiple ground-truth labels

Methodology Details

Task Definition

Consider a uniform linear array (ULA) with N antennas at λ/2 spacing receiving T snapshots from M narrowband sources, modeled as a data matrix:

$X = A(\theta)S + V$

Where:

• $A(\theta) \in \mathbb{C}^{N \times M}$: Steering matrix for source DoA $\theta$
• $S \in \mathbb{C}^{M \times T}$: Source signals
• $V \in \mathbb{C}^{N \times T}$: Spatial white noise

Model Architecture

The HYPERDOA architecture comprises four main stages:

1. Feature Extraction Module

Mean Spatial Lag Autocorrelation:

• Compute sample spatial covariance matrix: $\hat{R}_X = \frac{1}{T}XX^H$
• Extract mean autocorrelation for each spatial lag k: $r_k = \frac{1}{N-k}\sum_{i=1}^{N-k}[\hat{R}_X]_{i,i+k}$
• Separate complex-valued vectors into real and imaginary parts and concatenate to form real-valued feature vectors

Spatial Smoothing:

• Partition the full antenna array of size N into L overlapping subarrays
• Compute covariance matrix for each subarray and average: $\hat{R}_{SS} = \frac{1}{L}\sum_{j=1}^L R_j$
• Construct feature vectors through vectorization of the upper triangular portion

2. HDC Encoding Module

Uses a fractional power encoder based on Fourier Holographic Reduced Representation (FHRR):

• Assign unique random basis hypervectors $B_i \in \mathbb{C}^D$ to each feature dimension
• Apply feature values as phase rotations to corresponding basis vectors
• Generate query hypervector: $H_q = \bigotimes_{i=1}^{2N} \rho^{f_i}(B_i)$

3. Associative Memory Module

Training Phase:

• Store prototype hypervectors (centroids) for each discrete candidate angle
• Improved multi-label learning rule with positive updates only for true labels: $C_{\theta_i} \leftarrow C_{\theta_i} + \eta H_q, \forall i \in \{1,...,M\}$

Inference Phase:

• Compute dot product similarity between query hypervector and all trained centroids
• Generate angle pseudospectrum

4. Multi-Source Decoding Module

Uses non-maximum suppression algorithm to identify M most prominent peaks:

1. Identify global maximum in similarity spectrum corresponding to angle
2. Apply suppression window, discarding all similarity scores within minimum angular separation (6°)
3. Repeat process until M sources are identified

Technical Innovations

1. Pattern Recognition Reformulation: Transforms traditional subspace decomposition problems into pattern matching in hyperdimensional space
2. Multi-Label Adaptation: Innovatively modifies HDC learning rules for DoA estimation's inherent multi-source characteristics
3. Transparent Computing: Replaces complex matrix decomposition with simple algebraic operations, improving interpretability and efficiency
4. Feature Engineering: Combines classical signal processing insights with feature extraction strategies suited for HDC

Experimental Setup

Dataset

• Signal Model: Half-wavelength spaced ULA with N elements and M sources
• Data Generation: Source DoAs uniformly sampled from -90°, 90° with minimum separation of 15°
• Scenarios: Both coherent and incoherent source cases
• Noise Model: Complex Gaussian noise
• Data Scale: 45,000 training samples, 2,250 test samples (5%)

Evaluation Metrics

Mean square periodic error (MSPE, dB) used as primary accuracy metric

Comparison Methods

• Classical Methods: MUSIC, Root-MUSIC, ESPRIT
• Deep Learning Methods: SubspaceNet+MUSIC, SubspaceNet+Root-MUSIC, SubspaceNet+ESPRIT

Implementation Details

• HDC dimensionality: D = 10,000
• Angular grid resolution: 0.1°
• Multi-source decoding peak separation: 6°
• Evaluation configuration: N=8, T=100, M=2,3,4

Experimental Results

Main Results

Incoherent Source Scenario:

• SNR 1,5 dB range: HDC(lag) achieves 6.29% higher accuracy than SOTA methods
• SNR -5,-1 dB range: HDC(lag) achieves 18.41% higher accuracy than SOTA methods

Coherent Source Scenario:

• SNR 1,5 dB, M=3: HDC(lag) achieves 45.38% accuracy improvement, HDC(spatial smoothing) achieves 53.13% improvement
• SNR -5,-1 dB: HYPERDOA maintains best performance as source count increases
• Overall SNR -5,5 dB, M=3,4: HYPERDOA achieves 35.39% accuracy improvement over SOTA methods

Energy Consumption Results:

• HDC(lag): 135 mJ/inference
• HDC(spatial smoothing): 142 mJ/inference
• SubspaceNet+MUSIC: 4645 mJ/inference (3253.8% higher than HDC on average)
• 92.93% more energy-efficient compared to all neural baselines

Accuracy-Energy Trade-off Analysis

Experimental results demonstrate that HYPERDOA achieves top-tier accuracy across both SNR ranges while significantly reducing energy consumption, establishing its feasibility in resource-constrained, mission-critical environments.

Related Work

Main Directions in DoA Estimation

1. Classical Subspace Methods: Eigendecomposition-based methods such as MUSIC and ESPRIT
2. Deep Learning Methods: Data-driven approaches including DeepMUSIC and SubspaceNet
3. Robustness Enhancement: Preprocessing techniques such as spatial smoothing and forward-backward averaging

Hyperdimensional Computing Applications

HDC, as a brain-inspired computing paradigm, has demonstrated inherent robustness and computational efficiency advantages in biological signal processing, wearable health applications, and related fields.

Advantages of This Work

Compared to existing work, HYPERDOA is the first to apply HDC to DoA estimation, combining the interpretability of classical methods with the adaptability of deep learning approaches, while overcoming the primary deficiencies of both.

Conclusions and Discussion

Main Conclusions

1. HYPERDOA successfully transforms DoA estimation into an HDC pattern recognition problem
2. Achieves significant accuracy improvements in challenging low SNR and coherent source scenarios
3. Substantially reduces energy consumption, suitable for edge device deployment
4. Provides transparent, interpretable computational processes

Limitations

1. Currently validated only on synthetic data, lacking real-world environment testing
2. Does not account for practical factors such as array calibration errors
3. Angular resolution limited by discrete grid precision
4. Performance boundaries for extremely low SNR scenarios not fully explored

Future Directions

1. Test in more challenging real-world scenarios
2. Investigate robustness to array imperfections
3. Explore novel HDC techniques and optimization strategies
4. Extend to other array geometries

In-Depth Evaluation

Strengths

1. Strong Innovation: First application of HDC to DoA estimation, providing novel problem-solving approaches
2. High Practical Value: Simultaneously addresses accuracy and energy efficiency, suitable for real deployment
3. Complete Methodology: Comprehensive solution from theoretical framework to system implementation
4. Sufficient Experimentation: Covers multiple scenarios with comprehensive performance evaluation
5. Good Interpretability: Avoids the black-box problem of deep learning

Weaknesses

1. Limited Validation: Validated only on synthetic data, lacking real-world data verification
2. Insufficient Theoretical Analysis: Lacks theoretical analysis of HDC performance bounds in DoA estimation
3. Parameter Sensitivity: Insufficient analysis of critical parameter impacts (e.g., HDC dimensionality, grid resolution)
4. Scalability Concerns: Scalability to large-scale arrays and high-dimensional scenarios remains unknown

Impact

1. Academic Contribution: Provides new research directions for the DoA estimation field
2. Practical Value: Offers feasible solutions for signal processing on edge devices
3. Cross-Domain Significance: Demonstrates HDC's application potential in traditional signal processing problems

Applicable Scenarios

• DoA estimation on resource-constrained edge devices
• Safety-critical system applications requiring interpretable AI
• Challenging environments with low SNR and coherent sources
• Mobile platforms with real-time processing requirements

References

The paper cites 28 relevant references covering classical DoA estimation methods, deep learning approaches, HDC theoretical foundations, and applications, providing solid theoretical grounding for the research.

Overall Assessment: This is a high-quality interdisciplinary research paper that successfully applies hyperdimensional computing, an emerging computing paradigm, to the classical DoA estimation problem. The paper makes significant contributions in methodological innovation, experimental validation, and practical value, offering promising solutions for addressing signal processing challenges on edge devices. While there remains room for improvement in theoretical analysis and practical validation, the overall research quality is high with important academic value and application prospects.