- #1
stanley.st
- 31
- 0
I'm reading book and there's proposition with convex function
Function f is convex if and only if for all x,y
[tex](*)\quad f(x)-f(y)\ge\nabla f(y)^T(x-y)[/tex]
It's proven in this way: From definition of convexity
[tex]f(\lambda x+(1-\lambda)x)\le \lambda f(x)+(1-\lambda)f(y)[/tex]
we have
[tex]\frac{f(y+\lambda(x-y))-f(y)}{\lambda}\le f(x)-f(y)[/tex]
Setting [tex]\lambda\to0[/tex] we have (*).
-------------------------------------------------
My problem is in last sentece. I understand formula on left-hand side as directional derivative. But in definition of directional derivative is needed to (x-y) be an unit vector. It is not. So it is not a directional derivative. If I define
[tex]\lambda:=\frac{\mu}{\Vert x-y\Vert}[/tex]
I have directional derivative on left hand side
[tex]\frac{f(y+\mu\frac{(x-y)}{\Vert x-y\Vert})-f(y)}{\mu}\le\frac{f(x)-f(y)}{\Vert x-y\Vert}[/tex]
But in this way I don't obtain result (*) but I obtain this
[tex]\nabla f(y)^T(x-y)\le\frac{f(x)-f(y)}{\Vert x-y\Vert}[/tex]
Function f is convex if and only if for all x,y
[tex](*)\quad f(x)-f(y)\ge\nabla f(y)^T(x-y)[/tex]
It's proven in this way: From definition of convexity
[tex]f(\lambda x+(1-\lambda)x)\le \lambda f(x)+(1-\lambda)f(y)[/tex]
we have
[tex]\frac{f(y+\lambda(x-y))-f(y)}{\lambda}\le f(x)-f(y)[/tex]
Setting [tex]\lambda\to0[/tex] we have (*).
-------------------------------------------------
My problem is in last sentece. I understand formula on left-hand side as directional derivative. But in definition of directional derivative is needed to (x-y) be an unit vector. It is not. So it is not a directional derivative. If I define
[tex]\lambda:=\frac{\mu}{\Vert x-y\Vert}[/tex]
I have directional derivative on left hand side
[tex]\frac{f(y+\mu\frac{(x-y)}{\Vert x-y\Vert})-f(y)}{\mu}\le\frac{f(x)-f(y)}{\Vert x-y\Vert}[/tex]
But in this way I don't obtain result (*) but I obtain this
[tex]\nabla f(y)^T(x-y)\le\frac{f(x)-f(y)}{\Vert x-y\Vert}[/tex]