HW1 Solution - IE 8534 - University of Minnesota Solution - IE 8534 Hamidreza Badri February 16,...
Transcript of HW1 Solution - IE 8534 - University of Minnesota Solution - IE 8534 Hamidreza Badri February 16,...
HW1 Solution - IE 8534
Hamidreza Badri
February 16, 2014
Problem 1.1
Image of convex set under perspective mapping remains convex:
P (x, t) = x/t, x ∈ Rn, t > 0
Let S be a convex set, therefore we have to show that if (x1, t1), (x2, t2) ∈ S, then λP (x1, t1) +(1− λ)P (x2, t2) ∈ P (S) for all λ ∈ [0, 1].
λP (x1, t1) + (1− λ)P (x2, t2)
= λx1t1
+ (1− λ)x2t2
=λt2x1 + (1− λ)t1x2
t1t2
=
λt2λt2+(1−λ)t1x1 + (1−λ)t1
λt2+(1−λ)t1x2t1t2
λt2+(1−λ)t1
By defining λ′
= λt2λt2+(1−λ)t1 :
λP (x1, t1) + (1− λ)P (x2, t2) =λ
′x1 + (1− λ′
)x2λ′t1 + (1− λ′)t2
Since S is a convex set, therefore λ′x1 + (1− λ′
)x2 ∈ S and 0 ≤ λ′= λt2
λt2+(1−λ)t1 ≤ 1.
Inverse image of convex set under perspective mapping remains convex:
The inverse function can be defined as follow
P−1(S) = {(x, t)|x/t ∈ S, t > 0}
Therefore for (x1, t1), (x2, t2) ∈ P−1(S), we must show λ(x1, t1) + (1− λ)(x2, t2) ∈ P−1(S). Wehave
1
λx1 + (1− λ)x2λt1 + (1− λ)t2
=λt1
x1t1
+ (1− λ)t2x2t2
λt1 + (1− λ)t2
=λt1
λt1 + (1− λ)t2
x1t1
+(1− λ)t2
λt1 + (1− λ)t2
x2t2
By defining λ′
= λt1λt1+(1−λ)t2 ∈ [0, 1], it shows that λ
′(x1, t1) + (1− λ′
)(x2, t2) ∈ P−1(S).
Problem 1.2
Image of convex sets under a linear-fractional mapping remains convex:
Linear-fractional mapping function is defined by F (x) = Ax+bcT x+d
, where cTx+ d > 0. Let S be aconvex set, therefore we have to show for every x1, x2 ∈ S, λF (x1) + (1− λ)F (x2) ∈ F (S).
λF (x1) + (1− λ)F (x2)
λF (x1) + (1− λ)F (x2) = λAx1 + b
cTx1 + d+ (1− λ)
Ax2 + b
cTx2 + d
If we pick
λ =λ
′(cTx1 + d)
λ′(cTx1 + d) + (1− λ′)(cTx2 + d), λ, λ
′ ∈ [0, 1]
We can get
λF (x1) + (1− λ)F (x2) =
λ′(cTx1 + d)
λ′(cTx1 + d) + (1− λ′)(cTx2 + d)
Ax1 + b
cTx1 + d+ (1− λ
′(cTx1 + d)
λ′(cTx1 + d) + (1− λ′)(cTx2 + d))Ax2 + b
cTx2 + d
=A(λ
′x1 + (1− λ′
)x2) + b
cT (λ′x1 + (1− λ′)x2) + d
Since S is a convex set and λ′ ∈ [0, 1], λ
′x1 +(1−λ′
)x2 ∈ S. Therefore λF (x1)+(1−λ)F (x2) ∈F (S).
Inverse image of convex sets under a Linear-fractional mapping remains convex:
The inverse image of thew convex set S under a Linear-fractional mapping function is F−1(S) ={x| Ax+b
cT x+d∈ S, cTx+d > S}. We need to show that λx1+(1−λ)x2 ∈ F−1(S) for all x1, x2 ∈ F−1(S),
0 ≤ λ ≤ 1.
2
A(λx1 + (1− λ)x2) + b
cT (λx1 + (1− λ)x2) + d
=λ(Ax1 + b) + (1− λ)(Ax2 + b)
λ(cTx1 + d) + (1− λ)(cTx2 + d)
=λ(cTx1 + d) (Ax1+b)
(cT x1+d)+ (1− λ)(cTx2 + d) (Ax2+b)
(cT x2+d)
λ(cTx1 + d) + (1− λ)(cTx2 + d)
=λ(cTx1 + d) (Ax1+b)
(cT x1+d)+ (1− λ)(cTx2 + d) (Ax2+b)
(cT x2+d)
λ(cTx1 + d) + (1− λ)(cTx2 + d)
=λ(cTx1 + d)
λ(cTx1 + d) + (1− λ)(cTx2 + d)
(Ax1 + b)
(cTx1 + d)+
(1− λ)(cTx2 + d)
λ(cTx1 + d) + (1− λ)(cTx2 + d)
(Ax2 + b)
(cTx2 + d)
Bt defining λ′
= λ(cT x1+d)λ(cT x1+d)+(1−λ)(cT x2+d) , 0 ≤ λ′ ≤ 1, we have
A(λx1 + (1− λ)x2) + b
cT (λx1 + (1− λ)x2) + d= λ
′F (x1) + (1− λ′
)F (x2)
Since S is convex, λx1 + (1− λ)x2 ∈ S and it completes the proof.
Problem 2.1
epi(f) = {(x, t)|f(x) ≤ t, x ∈ dom(f)}
Suppose f be a convex function, then for (x1, t1), (x2, t2) ∈ epi(f) we have
f(λx1 + (1− λ)x2) ≤ λf(x1) + (1− λ)f(x2)
and since f(x1) ≤ t1 and f(x2) ≤ t2, we can get
λf(x1) + (1− λ)f(x2) ≤ λt1 + (1− λ)t2
Therefore {λx1 + (1 − λ)x2, λt1 + (1 − λ)t2} ∈ epi(f) and it shows that epi(f) is a convex setfor convex functions. Conversely assume epi(f) is a convex set, then if (x1, t1), (x2, t2) ∈ epi(f) wehave (λx1 + (1− λ)x2, λt1 + (1− λ)t2) ∈ epi(f), which means:
f(λx1 + (1− λ)x2) ≤ λt1 + (1− λ)t2
3
We can always pick t1 = f(x1) and t2 = f(x2) which imply the definition of convex functions.
Problem 2.2
Assume x1, x2 ∈ dom(f), then
g(λx1 + (a− λ)x2) = inf(f(λx1 + (1− λ)x2), y)
≤ f(λx1 + (1− λ)x2, λy1 + (1− λ)y2)
≤ λf(x1, y1) + (1− λ)f(x2, y2)
≤ λg(x1) + (1− λ)g(x2) + ε
By taking the limit of the right hand side of the last inequity when ε → 0, proof completes.Also the last inequality comes from the fact that for ε ≥ 0, there are y1 and y2 such that
f(x1, y1) ≤ g(x1) + ε
f(x2, y2) ≤ g(x2) + ε
4