Short communicationQuaternion kernel recursive least-squares algorithm
Introduction
The quaternion algebra describes a wide variety of geometrical objects (such as points, lines, and planes), simplifies many physical problems [1], [2], and have been widely used in various fields, such as quaternion neural network [3], three-dimension motion tracking [4], four-dimension biomedical signal processing [5], frequency domain adaptive filtering [6] and nonlinear adaptive system using kernel adaptive filters (KAFs) [7]. In the quaternion domain, the mutual information among all four components can be performed with quaternion multiplication. The 4-D signals in the real domain cannot incorporate the mutual information [7]. The mutual information may help reduce the complexity. It is shown in [3] that the quaternion neural network reduces the complexity by about 30% compared with the real one under the same performance.
Meanwhile, the KAF has developed rapidly in recent years and can cope with nonlinear filtering problems using the kernel trick [8] of the reproducing kernel Hilbert space (RKHS) [9]. The corresponding algorithms were early studied in real- and complex-valued field, including the real kernel least-mean-square (KLMS) [10], real kernel recursive least-square (KRLS) [11], [12], [13], [14], [15] and real kernel recursive maximum correntropy [16], [17] and complex Gaussian KLMS algorithm [18]. Note that the real kernel filters need plentiful similar filters and ignore the connection between each component when dealing with 3-D or 4-D signals. The recent development of sensor technology makes 3-D or 4-D sensor signals accessible, and it requires proper algorithms, especially kernel algorithms to match the properties of hyper-dimension signals. When dealing with hyper-dimension signals, a quaternion kernel filter can combine the mutual information of four nonlinear constituents. Then the KAF was studied in quaternion-valued field using KLMS algorithm, such as quaternion stochastic information gradient algorithm [7] and quaternion minimum error entropy with fiducial point algorithm [19].
The derivation of quaternion algorithms, whether including a kernel or not, are based on the quaternion gradient. Different from the complex-valued and the real-valued domain, the quaternion domain is non-commutative. The quaternion gradient has undergone several versions. From 2009 to 2012, quaternion least-mean-squares (QLMS) [22], [23], HR-QLMS [24], and iQLMS [25], [26] were proposed. Later, the three types of gradient were proved to misuse the non-commutative property [27], [28], [29], [30]. From 2014 to 2019, the quaternion gradient was proposed using three approaches: the generalized Hamilton-real (GHR) calculus [27], [28], the quaternion product [29] and the quaternion involutions [30]. The above mentioned quaternion KAF algorithms in [7], [19] are based on the GHR calculus. The early quaternion RLS (QRLS) algorithms [20], [21] are based on the HR-QLMS [24], which prevents the generalization from RLS to KRLS using the kernel trick.
To the best of our knowledge, the quaternion KRLS has not been tackled by current methods. Generally speaking, the KRLS algorithm converges faster than the KLMS algorithm. This motivates us to study a KRLS algorithm in quaternion domain to coping with hyper-dimension nonlinear input signals. In this paper, we generalize the GHR calculus method [28] to derive a quaternion KRLS algorithm. Firstly, we investigate the QRLS algorithm; Secondly, we generalize the QRLS algorithm to the quaternion KRLS algorithm. The proposed KRLS algorithm can be applied in the nonlinear adaptive systems, such as biomedical signal processing and channel nonlinearities.
Section snippets
Quaternion algebra
A quaternion variable consists of a real component and three imaginary components, written as a linear sumwhere The conjugate of a quaternion isThe relationship among the orthogonal unit vectors areThe above three formulas define the quaterion algebra. The quaternion product is non-commutative, for example,, instead Given we haveq1q2 ≠ q2q1.
Introducing the three perpendicular
Quaternion RLS
Similar with the above quaternion LMS algorithm, the quaternion RLS minimizes the cost function in estimating w. The cost function [14] is defined aswhere λ is a real regularization parameter and . The gradient of Eq. (14) is obtained from (8), (12) and (13), yieldingDefining
Quaternion Kernel Method
The theory of the RKHS greatly promotes the development of the theory of adaptive signal processing. With the kernel trick, the KAF solves nonlinear filtering problems. In this section, we focus on providing the quaternion kernel RLS algorithm.
For a sequence of n input–output pairs, , the input ui is transformed into the feature space , which is an infinite-dimensional space. The parametric model of the feature space becomeswhere , and ψ(ui
Quaternion RLS
We test the proposed QRLS algorithm and the real RLS algorithm. The linear model in quaternion model has been given in (9). The variables in (9) is expressed bywhere (i) is omitted for simplicity.
The corresponding expression in the real domain is given byIntroduce
Conclusion
We have generalized the GHR calculus method to calculate the quaternion RLS filter, and have proposed a new quaternion KRLS algorithm, which has the same form as the real-valued KRLS algorithm. The simulations verified the feasibility of the proposed algorithm. The proposed QRLS algorithm had almost the same steady error as that of the CFS under different SNRs. Moreover, the proposed quaternion KRLS converged faster than the algorithms, Quat-KLMS, and WL-Quat-KLMS.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (34)
- et al.
Uncini, Frequency domain quaternion adaptive filters: algorithms and convergence performance
Signal Process.
(2017) - et al.
Kernel recursive maximum correntropy
Signal Process.
(2015) - et al.
The quaternion minimum error entropy algorithm with fiducial point for nonlinear adaptive systems
Signal Process.
(2019) Quaternions and matrices of quaternions
Linear Algebra Appl.
(1997)- T.A. Ell, N. Le Bihan, S.J. Sangwine, Quaternion Fourier Transforms for Signal and Image Processing, Wiley, London, UK,...
Elements of quaternions
(1881)- et al.
Lecture notes in control and information sciences: Vol. 234
Neural Networks in Multidimensional Domains
(1998) - et al.
Multiple-model adaptive estimation for 3-D and 4-D signals: a widely linear quaternion approach
IEEE Trans. Neural Netw. Learn. Syst.
(2019) - et al.
A kernel adaptive algorithm for quaternion-valued inputs
IEEE Trans. Neural Netw. Learn. Syst.
(2015) - et al.
The quaternion stochastic information gradient algorithm for nonlinear adaptive systems
IEEE Trans. Signal Process.
(2019)
Quaternion reproducing Kernel Hilbert spaces: existence and uniqueness conditions
IEEE Trans. Inf. Theory
Theory of reproducing kernels
Trans. Am. Math. Soc.
Transfer learning in adaptive filters: the nearest instance centroid-estimation Kernel least-mean-square algorithm
IEEE Trans. Signal Process.
Kernel recursive least squares with multiple feedback and its convergence analysis
IEEE Trans. Circuits Syst. II: Express Briefs
A class of weighted quantized kernel recursive least squares algorithms
IEEE Trans. Circuits Syst. II: Express Briefs
The kernel recursive least-squares algorithm
IEEE Trans. Signal Process.
Extended Kernel recursive least squares algorithm
IEEE Trans. Signal Process.
Cited by (9)
Quantized kernel recursive minimum error entropy algorithm
2023, Engineering Applications of Artificial IntelligenceWidely nonlinear quaternion-valued second-order Volterra recursive least squares filter
2023, Signal ProcessingCitation Excerpt :As we all know, some systems or signals in practical applications have some inherent nonlinear characteristics, and the aforementioned quaternion linear algorithms cannot handle nonlinear problems well. Therefore, a variety of nonlinear quaternion models such as quaternion neural network, quaternion polynomial filtering and quaternion kernel filtering [11,13,15,16,18] came into being. Polynomial filtering is widely used in real and complex-valued domains because of its simplicity and high nonlinearity [22], but its research in quaternion domain is relatively few.
A kernel recursive minimum error entropy adaptive filter
2022, Signal Processing