Optimization with Lagrangian Multipliers

Leo Liu
Messages
353
Reaction score
156
Homework Statement
.
Relevant Equations
##\nabla f=\lambda\nabla g##
Problem:
1615503170806.png

Solution:
1615503188617.png

My question:
My reasoning was that if x is max at the point then the gradient vector of g at the point has only x component; that is ##g_y=0,\, g_z=0##. This way I got:
$$\begin{cases}
4y^3+x+z=0\\
\\
4z^3+x+y=0\\
\\
\underbrace{x^4+y^4+z^4+xy+yz+zx=6}_\text{constraint equation}
\end{cases}$$
which produces the same solutions as the list of equations in the official answer does.

What puzzles me is why the answer defines that ##f(x,y,z)=x##. I understand the gradient vector should be parallel to the vector ##<1,0,0>##, and therefore the equation f is x. But this step is reverse engineered. Can someone please explain where ##f(x,y,z)=x## comes from?

Many thanks.
 
Physics news on Phys.org
Leo Liu said:
Can someone please explain where ##f(x,y,z)=x## comes from?
That's just the function you're extremising, ##f(\boldsymbol{x}) = x## (i.e. the function taking a point ##\in \mathbb{R}^3## to its ##x## co-ordinate, which is what you're trying to maximise) subject to the constraint ##g(\boldsymbol{x}) = 6##. It's essentially by construction.

To give some more context, the constraint defines a 2-dimensional surface with normal vector ##\nabla g(\boldsymbol{x})##; displacements ##\mathrm{d}\boldsymbol{x}## within this surface must satisfy ##\mathrm{d}\boldsymbol{x} \cdot \nabla g (\boldsymbol{x} )= 0##, i.e. they must be parallel to the surface. Furthermore, if ##\boldsymbol{x}_0## is the point corresponding to the extremum of ##f(\boldsymbol{x})## under the given constraint, you must have ##\mathrm{d}\boldsymbol{x} \cdot \nabla f (\boldsymbol{x}_0) = 0##, since the change in the value of the function will be zero to first order under a displacement ##\mathrm{d}\boldsymbol{x}##. The only way these two things are possible is if $$\nabla f(\boldsymbol{x}_0) = \lambda \nabla g(\boldsymbol{x}_0)$$It's because if they were not parallel, you would have$$\nabla f(\boldsymbol{x}_0) = \lambda \nabla g(\boldsymbol{x}_0) + \mathbf{a}(\boldsymbol{x}_0) \implies \mathrm{d}\boldsymbol{x} \cdot \nabla f(\boldsymbol{x}_0) = \mathrm{d}\boldsymbol{x} \cdot \mathbf{a}(\boldsymbol{x}_0)$$for some ##\mathbf{a}(\boldsymbol{x}_0)## which is perpendicular to ##\nabla g(\boldsymbol{x}_0)##. But since ##\mathrm{d}\boldsymbol{x}## can be a displacement of any direction in the constraint surface, we can consider letting ##\mathrm{d}\boldsymbol{x} = \mathbf{a}(\boldsymbol{x}_0) \mathrm{d}s##, in which case ##\mathrm{d}\boldsymbol{x} \cdot \mathbf{a}(\boldsymbol{x}_0) = |\mathbf{a}(\boldsymbol{x}_0)|^2 \mathrm{d}s > 0##, which means that ##\mathrm{d}\boldsymbol{x} \cdot \nabla f(\boldsymbol{x}_0)## would no longer be zero.
 
Last edited by a moderator:
There are two things I don't understand about this problem. First, when finding the nth root of a number, there should in theory be n solutions. However, the formula produces n+1 roots. Here is how. The first root is simply ##\left(r\right)^{\left(\frac{1}{n}\right)}##. Then you multiply this first root by n additional expressions given by the formula, as you go through k=0,1,...n-1. So you end up with n+1 roots, which cannot be correct. Let me illustrate what I mean. For this...
Back
Top