Lemma 2.2.3 (The General Case) 在满足normal function前提下,有总结如下:
(a) When H is full rank, the unique solution is given by x = (H* H)-1 H*y.
(b) When H is not full rank, the normal equations always have more than one solution,where any two solutions x1and x2 differ by a vector in the nul/space of H, i.e.,
H(x1- x2) = 0.
(c) The projection of y onto C(H) is unique and is defined by , where
x is any solution to the normal equations; when H has full rank, we can write
Let be a linear subspace of and let be an arbitrary element of The projection of onto denoted by or often just is a unique element of such that:
Lemma 2.3.1 (Orthogonality and Approximation)
Let be a subspace of a linear vector space and let be any element of Then the projection, \hat{y} c , has the property that
for any
证明: We can write
But since and and, by definition, is orthogonal to Therefore
2.3.2 几何见解(Geometric Insights)
引理2.3.1证明了本节开头使用的几何参数的合理性:
The subspace is now the space spanned by the columns of the matrix
H. The least-squares solution, is characterized by the fact that the residual vector is orthogonal to or equivalently, is given by the unique projection of onto This is depicted in Fig. 2.1
is Hermitian and positive-definite, 所以数值计算上喜欢用柯列斯基分解来解。
Solve the lower triangular system of equations for
Solve the upper triangular system of equations for .
直接用回代法即可求解,因为左边的系数矩阵都是三角阵。
好处:不用求逆矩阵,并且比LU分解快两倍。
坏处:但是,这个程序在实现时可能会遇到数值困难这是因为对于病态数据矩阵(ill-conditioned matrices) h,当矩阵乘积 h * h 形成时,数值精度就丧失了( h * h 不可逆)。
ill-conditioned matrices are those that have a very large ratio of the largest to the smallest singular values For further discussion, see any textbook on numerical linear algebra, those cited at the end of
Lemma 2.6 .1 (Recursive Updating: The RLS Algorithm) The solution of problem can be computed as
where and satisfies the Riccati recursion
and is the regularized least-squares solution of The effort required for one step of the recursion is flops.so that for N steps we need only operations vs 直接解nomal function时间复杂度为
Lemma 2.7.1 (Recursive Downdating) Suppose. is the regularized least-squares solution to the overdetermined system of linear equations obtained by deleting the first equation in the system Then
TLS解最终证实了对正则化最小二乘解的解释.More specifically, assume is full rank with smallest singular value . Assume also is full rank with smallest singular value When a unique solution to the TLS problem exists and it can be expressed in the form(满足奇异值分解的某种条件,TLS就有如下唯一解)
与(2.4.7)相比,我们可以将TLS解视为正则化成本函数的解,但具有负定矩阵的解不存在。被替换
正则化最小二乘的解:(2.4.7)
替换后,正则化最小二乘的目标函数由:
可写为:
2.8.2数据不确定性有条件的标准Criteria with Bounds on Data Uncertainties
Appendix for Chapter 2
Part 1
解的分析有四种情况
(1)有不止一个解,当且仅当 y 在column space of H并且null space of H不为零。== H不列满秩且有解
此时任意的两个解使成立:
即
(2)方程(2.A.1)有唯一解,当且仅 y 在column space of H并且H列满秩
方程?xH=y?则是同理由row space of H来入手.
(3)H为方阵且满秩,则方程(2.A.1)有解且有唯一解
(4)方程无解,y不在column space of H
Part 2
Lemma Let be an matrix with comnlex entries. Then we have