Gradient-Ascent Eigen Decomposition for Fast Covariance Convergence
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for eigen-decomposition, such as the power method, require a large number of iterations to converge, especially when eigenvalues are clustered, leading to high computational complexity in applications like user-specific eigen-beamforming in MIMO systems.
Innovation Solution
The method employs gradient ascent iterations with an optimal step size that maximizes the Rayleigh quotient, allowing for iterative determination of eigenvalues and eigenvectors with reduced computational complexity, specifically requiring only one quadratic-complexity operation per iteration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If the power method is used for eigen-decomposition, then the per-iteration complexity is low, but the number of iterations required to converge is large, especially when eigenvalues are clustered
Solution Approach 1:
The patent transforms the eigen-decomposition problem into an optimization problem by changing the parameter representation from direct iterative methods to gradient-based optimization of the Rayleigh quotient. This parameter transformation allows the use of gradient ascent with optimally chosen step sizes, achieving faster convergence while maintaining computational efficiency per iteration.
2Productivity
If advanced methods like Lanczos iteration or partial SVD are used, then convergence is improved, but the per-iteration complexity becomes cubic or higher
Solution Approach 1:
The patent replaces complex linear algebra operations (cubic complexity methods like Lanczos or SVD) with a gradient-based optimization approach that relies on simpler matrix-vector multiplications and gradient computations. This substitution reduces per-iteration complexity from cubic to quadratic while achieving comparable or better convergence rates through optimal step size selection.
3Reliability
If user-specific eigen-beamforming is implemented, then beamforming gains are significantly increased, but computational requirements increase due to eigen-decomposition of channel covariance matrices
Solution Approach 1:
The patent introduces dynamic step size adjustment in the gradient ascent algorithm, where the step size is optimized at each iteration to maximize the Rayleigh quotient. This dynamic adaptation allows the algorithm to converge faster with fewer iterations, reducing the overall computational burden while maintaining the high beamforming gains required for user-specific eigen-beamforming in massive MIMO systems.
Data Source
AI summary
This specification relates to systems and methods for determining eigenvalues and eigenvectors of a covariance matrix, for example for use in eigen-beamforming in MIMO systems. According to a first aspect of this specification, there is describes a method comprising: receiving an input covariance matrix; and determining one or more eigenvalues and/or eigenvectors of the matrix iteratively using gradient ascent, wherein each iteration of the gradient ascent has a step size, μ, that maximises a Rayleigh quotient along the gradient.