Solving strange system of equations arising from Lagrange multipliers

83 Views Asked by At

Let $Q \in \mathbb{R}^{n \times n}$ be positive definite. I am trying to solve $$\min f(x) = ||x||^2$$ $$\text{s.t } \hspace{0.3 cm} h(x) = x^TQx-1 = 0$$ The constraint set is closed and bounded, so there is indeed a minimum.

We have $\nabla f(x) = 2x$ and $\nabla h(x) = 2Qx$, so the Lagrange multiplier system is

$$2x + \lambda 2Qx=0$$ $$x^TQx = 1$$

The problem is, I cannot seem to get an explicit solution to this system. The only thing I can come up with is an algorithm.

The first equation is equivalent to $$Qx = - \dfrac {1}{\lambda}x.$$

($\lambda$ can't be zero since that would imply $x = 0$, which cannot happen since $0$ is not in the constraint set)

So $x$ has to be an eigenvector of $Q$. Since $Q$ is symmetric, it has $n$ distinct eigenvalues $\mu_i$, so the geometric multiplicity of each eigenvalue is $1$, so all the eigenspaces are of the form $\text{span} \{v_i \}.$

Thus $x = c_i v_i$ for some $i$, so then for each $i$ I would plug in $cv_i$ into $x^TQx = 1$, which yields $$c_i = \pm \sqrt{\dfrac {1}{v_i^TQV_i}}.$$

Finally, I would plug in $c_iv_i$ into the objective function $f$ for every $i$, and see which one gives me the lowest value.

My question is: is my algorithm correct, and is there a better way to solve this problem?

Thank you.

2

There are 2 best solutions below

2
On BEST ANSWER

You have an error here. $Q$ does not necessarily have distinct eigenvalues (take $Q$ to be the identity matrix, for example). But there is nevertheless an orthonormal basis of eigenvectors. If $\mu_1$ is the largest eigenvalue of $Q$ with corresponding eigenvector $v$ chosen so that $h(v)=v^\top Qv = \mu\|v\|^2=1$, then this $v$ will give you the minimum value of $f(x)$.

If we choose $\{v_i\}$ to be an orthonormal basis of eigenvectors and write $x=\sum y_iv_i$, then $Q(x) = \sum \mu_i y_i^2$. Thus, if $\mu_1\ge\mu_j$ for all $j$, we have $$Q(x)=\sum \mu_i y_i^2 \le \mu_1 \|y\|^2 = \mu_1 \|x\|^2.$$ We conclude that if $Q(x)=1$, then $\|x\|^2\ge 1/\mu_1$, with equality holding precisely when $x=\pm \dfrac1{\sqrt{\mu_1}}v_1$.

0
On

This is indeed the best that you can do. Geometrically, the set ${x:x^TQx=1} $ forms an ellipse, with the principle axes being the directions of the eigenvectors. The eigenvector associated with the largest eigenvalue will minimize the norm of $x$. This solution method works for any such optimization problem with positive definite Q, so you know that you can’t do any better because you can essentially encode any principal eigenvalue problem into this problem.