Mathematical Foundations of Computer Vision
Michael Breuß Released:18.11.2011
Assigned to:Tutorial at 24.11.2011
Assignment 4 – Matrix Reloaded
Exercise No. 1 – Enter the Matrix
We consider a functiony=ϕ(x)withy∈IRm,x∈IRn, and whereϕis some transformation.
TheJacobian matrixofϕis defined as
∂y
∂x =
∂y1
∂x1 · · · ∂x∂y1 .. n
. . .. ...
∂ym
∂x1 · · · ∂y∂xm
n
∈IRm×n (1)
(a) Lety=Axwithy∈IRm,x∈IRn, and whereA= (aij),A∈IRm×n, does not depend onx.
Prove or disprove: ∂y∂x =A (2pts)
(b) Letf =x>Axbe given wheref ∈IR,x∈IRn,A= (aij),A∈IRn×n.
Compute ∂f∂x for(i)Anot symmetric, and for(ii)Asymmetric. (4pts) (c) Letf(z) =y>(z)x(z)wherez∈IRn,x(z)∈IRn,y(z)∈IRn.
Compute ∂f∂z. (2pts)
(d) Letϕ(x) =kx−vk2, wherex, v∈IRn.
Compute ∂ϕ∂x. (2pts)
Exercise No. 2 – Differentiate the Matrix
Now, letA= (aij)be am×nmatrix withaij =aij(t),t∈IR. Then
d
dtA(t) = A(t) =˙
da11
dt · · · dadt1n ... . .. ...
dam1
dt · · · dadtmn
∈IRm×n (2)
(a) LetB, C ∈IRn×nwithB = (bij),bij =bij(t)andC= (cij),cij=cij(t).
LetBC=I. Compute the equation resulting out of d
dt[BC] = d dt[I]
(4pts)
(b) LetA= (aij)∈IRm×nbe invertible, withaij =aij(t).
Compute dtd A−1
. (4pts)
1
Exercise No. 3 – Matrix Trinity
We consider again thesimilarityproperty.
We remember, that for given A ∈ IRn×n there is a similar matrix Λ if we have an orthogonal matrix Q∈IRn×nwithA=QΛQ>.
(a) Prove that the following implication holds:
IfA can be made similar to a diagonal matrixΛ,thenAis symmetric. (4pts) We have made use already of such matricesQcomposed of the eigenvectors ofA. We give this some more basement:
(b) Prove that the following assertion holds:
For a symmetric matrixA, the eigenvalues are real. (4pts)
We supplement this by:
(c) Prove that the following assertion holds:
For a symmetric matrixA, the eigenvectors to different eigenvalues are orthogonal. (4pts)
2