Subspace leaky LMS

The least mean squared (LMS) adaptive filtering algorithm may experience uncontrolled parameter drift when its input signal is not persistently exciting, leading to serious consequences when implemented with finite word-length. Though so-called "tap-leakage" modifications of LMS have been proposed to mitigate this drift, they inevitably introduce parameter bias which degrades mean-squared error performance. In this letter, we propose a novel algorithm which leaks only in the unexcited modes, thus introducing insignificant bias, while still retaining the low computational complexity of LMS.