Computer Science/Mathematics 550 Numerical Linear Algebra 2022
Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: daixieit
Computer Science/Mathematics 550
Numerical Linear Algebra
Assignment Three
2022
This entire homework will take the full rank least squares problem LS = arg miny∈R亢 ,b − Xy,2(2)
and compare it to the perturbed solution
LS = arg miny∈R亢 ,b + δb − (X + δX)y,2(2)
You are going to prove a bound on ,LS − LS ,2 and ,LS − LS ,2 where
LS = b − X LS
LS = b + δb − (X + δX)LS
1. Inverse of Augmented Matrix. We know that LS and LS satisfy \ ╱ \ = ╱ 0n(b) \
If X has full column rank, show that
(a) The matrix
M = \
is nonsingular.
(b) The inverse is given by
M-l = ╱ X(P)十 − (十)-l \ ,
where X十 is the Moore–Penrose pseudoinverse and P = Im − XX十 .
2. Perturbation of ,X十 ,2 . Show the following. (a) If X has full column rank then
,X十 ,2 = σn (X)-l
where σn (X) is the smallest singular value of X .
(b) Using the result in the notes that
σn (X) = min ,Xw,2
|w|2 =l
show that if
ηX = ,(δX)X十 ,2 < 1
then
σn (X + δX) ≥ (1 − ηX )σn (X) > 0.
Thus X + δX has full column rank and
,(X + δX)十 ,2 ≤ ,X十 ,2 /(1 − ηX ).
3. Perturbation Bound on the Solution. Assume that ηX = ,(δX)X十 ,2 < 1. Let
fLS = \ , gLS = \
and let
M˜ = \ .
(a) Noting the fLS solves
MfLS = ╱ 0(b) \
and
M˜ gLS = ╱ b δb \
show that
LS − LS = P [(δb) − (δX)LS ] − (X十 )T (δX)T LS
LS − LS = X十 [(δb) − (δX)LS ] + (XT X)-l (δX)T LS where P = Im − XX十 .
(b) Using norm bound and some identities, show that
,LS − LS ,2 ≤ ζ
,LS − LS ,2 ≤ ,X十 ,2 ζ
where
ζ ≤ ,δb,2 + ,(δX ),2 ,LS ,2 + ηX ,LS ,2 .
4. Meaning for Q–R factorization. Let D be a nonsingular diagonal matrix and let
X = XS D
(a) Show that if X has the Q–R factorization
X = Q ╱ 0(R) \
then XS has the Q–R factorization
XS = Q ╱ 0(R)S \ , RS = RD-l
and
XS(十) = DX十 .
(b) Show that if X + δX = (XS + δXS )D then
ηX = ,(δXS )XS(十) ,2 .
and thus
ηX ≤ min{,(δX)D-l ,2 ,(XD-l )十 ,2 : D diagonal and nonsingular }.
(c) In floating point arithmetic, Householder Q–R decomposition yields a computed Qˆ and Rˆ such that
X + δX = Qˆ ╱ 0(Rˆ) \
where
,δXej ,2 = ξ,Xej ,2 , ξ = O(mnεM ).
Let D = diag(dl , . . . , dn ), dj = ,Xej ,2 . Show that
ηX ≤ √nξ,XS(十) ,2 .
5. Perturbation of Subspaces. Let the full column rank matrix X ∈ Rm×n have the Q-R factorization
X = Q m(n) − n ╱ 0(R)\
where
n m − n
Q = ╱Ql Q2 、.
Let X + δX have the Q-R decomposition
X + δX = V m(n) − n ╱ 0(S)\
where
n m − n
V = ╱ Vl V2 、.
Assume that ηX = ,(δX)X十 ,2 < 1. The“angle”Θ between range(X) = range(Ql ) and range(X + δX) = range(Vl ) is given by
| sin Θ| = ,VlT Q2 ,2 = ,Ql(T)V2 ,2 .
Show that
(a)
X十 = ╱ R-0l \ QT
and that ,X十 ,2 = ,R-l ,2 .
(b)
,VlT Q2 ,2 ≤ ,(δX)R-l ,2 = ,(δX)X十 ,2 = ηX ,
,Ql(T)V2 ,2 ≤ ,(δX)S-l ,2 = ,(δX)(X + δX)十 ,2 = ηX+δX ,
and thus
| sin Θ| ≤ min{ηX , ηX+δX}.
2022-02-09