Fast tensorial JADE.

Area of Science:

• Signal Processing
• Machine Learning
• Data Analysis

Background:

• Classical Joint Approximate Diagonalization of Eigenmatrices (JADE) is a foundational algorithm for independent component analysis.
• Recent generalizations, TJADE and k-JADE, extend JADE but face limitations in computational speed and applicability to large datasets.
• Tensorial independent component analysis (TICA) is crucial for analyzing multi-dimensional data.

Purpose of the Study:

• To develop a novel TICA method that enhances computational efficiency while maintaining statistical rigor.
• To address the computational bottlenecks of existing TICA algorithms like TJADE and k-JADE.
• To provide a feasible TICA solution for large-scale datasets, such as video data.

Main Methods:

• The proposed method is based on TJADE and k-JADE, adapting their principles for improved performance.
• Mathematical proofs are provided to establish the statistical properties, including consistency and limiting distribution.
• The method's performance is validated through simulations and timing comparisons against existing algorithms.

Main Results:

• The novel TICA method demonstrates significant improvements in computational speed compared to TJADE and k-JADE.
• Statistical properties, including consistency and limiting distribution, are achieved under mild assumptions.
• The method proves effective for large-scale video data analysis, where prior methods were infeasible.
• Approximate efficiency is achieved for finite samples.

Conclusions:

• The developed TICA method offers a computationally efficient and statistically sound alternative for analyzing complex, multi-dimensional data.
• This approach overcomes the limitations of previous TICA algorithms, enabling applications to previously intractable large-scale datasets.
• An experimental procedure for tuning parameter selection is proposed, enhancing practical usability.