RECURSIVE LEAST MEAN SQUARE FILTER
The Recursive Least Mean Squares (RLMS) filter is an adaptive filtering algorithm used for signal
processing, system identification, and control. It is an extension of the Least Mean Squares (LMS) filter,
leveraging a recursive approach to update filter coefficients for real-time adaptability to varying signal
environments. The least mean square (LMS) algorithm has long been a workhorse of adaptive filtering and
has, since its conception in 1959, found a number of successful applications. Based on the minimization
of an instantaneous approximation to the mean square error, the LMS converges to Wiener-Hopf solution
wop (i.e. the optimal solution), while its ‘instantaneous’ nature also makes it an algorithm of choice for
non-stationary signal processing. Another popular way to reach the Wiener solution, wop, is by using the
method of least squares, through the algorithm called the recursive least squares (RLS). Owing to no
approximations in the derivation, the RLS is capable of converging in much fewer iterations than the LMS,
and achieve better tracking performance. It is nevertheless not without drawbacks, such as poorer ability
to track nonstationary processes or processes with outliers. Moreover, it is less numerically stable and
more computationally expensive, both due to the use of matrix inversion lemma. To date, there has been
no clear consensus on whether to use the LMS or the RLS, and the answer depends on the applications at
hand, while much research effort has been dedicated to improving both algorithms. For LMS, most of its
variants have focused on varying the step size, in an attempt to accelerate the convergence and reduce
steady-state misalignment, while for RLS, the main focus has been on effective implementation and
numerical robustness.
Here’s a detailed explanation:
Key Concepts of RLMS Filter
1. Recursive Framework:
a. Unlike the LMS, which updates its weights using only the most recent error, RLMS
incorporates past information recursively to enhance the convergence rate and stability.
b. RLMS minimizes the mean square error (MSE) between a desired signal and the filter
output by adapting its weights over time.
2. Weight Update Rule: The weight update rule in the RLMS filter typically builds upon the LMS
update formula but incorporates a recursive component:
3. Recursive Error Minimization: The RLMS builds on LMS by maintaining and adjusting past
covariance matrix information recursively to improve tracking in non-stationary environments.
Algorithm 1: The m-NLMS Algorithm
Advantages of RLMS
• Faster Convergence: It often converges faster than the LMS algorithm due to recursive
adjustment.
• Improved Tracking: Adapts more effectively to time-varying signals.
• Reduced Computational Complexity: While recursive, its design avoids the computational
overhead of full matrix inversions seen in algorithms like Recursive Least Squares (RLS).
Applications of RLMS
• Echo Cancellation: In telecommunication systems to cancel echoes.
• Noise Reduction: For adaptive noise suppression in audio and speech processing.
• System Identification: Identifying and modeling unknown systems dynamically.
• Adaptive Control: Controlling systems that are changing in real-time.
Comparison to Related Algorithms
Feature LMS RLMS RLS
Convergence
Moderate Faster Fastest
Speed
Stability High High Medium
Complexity Low Moderate High
Robustness to
High High Moderate
Noise
RLMS balances the trade-off between computational complexity (like LMS) and rapid adaptation (like
RLS), making it a practical choice for many adaptive filtering applications.
