Condition number. Square root filtering. Potter’s algorithm. M. Sami Fadali Professor EBME Department UNR 1 Singular For How close is a matrix to being singular? Singular value decomposition gives the singular values. MATLAB: condition number >> cond(P) value decomposition ∗ Singular 2 ∗ ∗ ∗ ∗ ∗ ∗ Values symmetric positive semidefinite. 3 Properties 4 Error covariance matrix can become illconditioned. Occurs if one state variable has a much higher uncertainty than another. For symmetric positive definite, use the Cholesky decomposition. MATLAB Matrix is computationally singular. >> Ps= chol(P,'lower') Ps = lower triangular matrix such that P=Ps.PsT 5 6 A priori error covariance matrix =square root of Example: Ps better conditioned than P A posteriori error covariance matrix =square root of 7 8 = state-transition matrix = zero mean white Gaussian process noise with covariance matrix. Factorization, 9 QR Factorization: factorize in to the product (here orthogonal nonunique. 10 Find any orthogonal matrix T such that Methods to find are QR factorization methods: Householder transformation, Givens transformation, Gram-Schmidt, modified Gram-Schmidt. matrix ) orthogonal matrix upper triangular matrix We only need R from the factorization. 11 12 Rows of measurement matrix diagonal Gain , , , , , Rewrite as: 13 14 Can show , , Recursion 15 for square root matrix 16 1. Compute the a priori square root matrix and estimate and initialize 2. Calculate for , , , 3. Set 17 18 Approximation Computing 19 20 , 21 Next 22 gain matrix , , Using: Actual 23 value 24 Square root filtering increases the precision of the Kalman filter. It is computationally more costly and harder to program. Was more important in the early days of Kalman filtering with less powerful computers. Is still useful for practical implementation in embedded systems. 25 D. Simon, Optimal State Estimation: Kalman, H∞, and Nonlinear Approaches, WileyInterscience, Hoboken, NJ, 2006. J. W. Demmel, Applied Numerical Linear Algebra, SIAM, Philadelphia, 1997. 26