I'm trying to interpret the two results when deriving the dual of the convex optimization problem:

$$\text{minimize }\|x\|_2$$ $$\text{subject to }Ax=b$$

where we assume that $x \in \mathbb{R}$; that the domain, $D$, of the problem is the set of reals.

We begin by writing the problem's Lagrangian:
$$L(x, \nu)=\|x\|_2+\nu^TAx - \nu^Tb$$

and the dual function: $$g(\nu)=\underset{x \in D}{\inf}L(x,\nu)=\underset{x \in D}{\inf}(\|x\|_2+\nu^TAx - \nu^Tb)$$

Solution #1: minimization by zero gradient condition

Standard practice at this point for differentiable $L(x,\nu)$ seems to be to find the optimal $x^*$ that satisfies: $$\nabla_xL(x,\nu)=\frac{x}{\|x\|_2} + A^T\nu=0$$

which gives: $$x^* = \left\{\begin{array}{ll} -A^T\nu & \|A^T\nu\|_2=1 \\ -\infty & otherwise \end{array}\right.$$

Then, substituting $x^*$ into $g(\nu)$ we have one (incomplete) interpretation of the dual function (for $ \|A^T\nu\|_2=1 $): $$\begin{align} g(\nu) &= \|-A^T\nu\|_2 - \|A^T\nu\|^2_2-b^T\nu\\ &= (1-1)-b^T\nu\\ &= \left\{\begin{array}{ll} -b^T\nu & \|A^T\nu\|_2=1\\ -\infty & otherwise \end{array}\right. \end{align}$$

Solution #2: general minimization over x

Alternately, we can choose to analyze the dual function to minimize the lagrangian over the domain of $x$ to give a more complete picture of the dual function. Again we start with the definition of the dual function: $$g(\nu) = \underset{x\in D}{\inf}L(x,\nu)=\underset{x\in D}{\inf}(\|x\|_2+\nu^TAx - b^T\nu) $$

and we make the choice of $x=-t\cdot A^T\nu, t\ge0$ (though $x=t\cdot A^T\nu, t\le0$ should work similarly) and substitute into $g(\nu)$ giving: $$ \begin{align} g(\nu) &= \underset{t\ge0}{\inf}(t\cdot(\|A^T\nu\|_2 - \|A^T\nu\|_2^2) - b^T\nu) \\ &= \left\{\begin{array}{ll} -b^T\nu & \|A^T\nu\|_2 \le 1 \\ -\infty & otherwise \end{array}\right. \end{align} $$


In solution #2 we get the additional finite expression, $-b^T\nu$, for the dual function when $\|A^T\nu\|_2\lt 1$.

My questions are:

  1. what is the interpretation of the different solutions when $\|A^T\nu\|_2\lt1$?
  2. why is it necessary to restrict $t$ to be non-negative in the substitution step for solution #2? How can we justify this constraint, and does this constraint influence the answer to question #1?

I'm looking for some geometric intuition into this problem that can explain the apparent difference here. I appreciate the help!

  • 3
    $\begingroup$ Why make it more difficult on yourself than necessary? Minimizing $\|x\|_2^2$ is equivalent, and it's a heck of a lot easier to work with. $\endgroup$ – Michael Grant Feb 18 '17 at 2:51
  • 1
    $\begingroup$ Thanks @MichaelGrant, I'll keep that in mind. The example was presented to me as the unsquared $L_2$ norm, probably as a demonstration of the zero gradient condition failing for this choice of non-differentiable objective, so I was just looking for some clarification of the understanding. $\endgroup$ – Ryan Neph Feb 18 '17 at 4:19
  • 1
    $\begingroup$ I would take Michael's suggestion here as it simplifies things. If not, for #1, note that the objective is not differentiable, so you have the condition $0 \in \partial_x L(x, \nu)$ (subdifferential). A little work shows that if $\|A^* \nu \| <1$ then $x^* = 0$. We have $\partial_x L(0, \nu) = \overline{B}(0,1) + \{ A^* \nu \}$. $\endgroup$ – copper.hat Feb 20 '17 at 6:52

This is an elaboration of my comment.

Michael's suggestion hits the nail on the head as it avoids the pitfalls encountered in both Solution #1 & Solution #2.

In both cases, the issue is that the objective is not differentiable everywhere, and so a little care is needed in the characterisation of a solution. To get some intuition, plot the function $x \mapsto |x|+bx$ for $|b|<1, |b|=1, |b|>1$ to see how the infimal value varies with $b$.

Regarding Question #1, let $\lambda(x) = L(x,\nu)$, we are looking for a solution of $\inf_x \lambda(x)$. Note that $\lambda$ is not differentiable everywhere, so we need to take care at points where it is not differentiable. Since it is convex and defined everywhere, we can use the subdifferential instead.

Since $\lambda$ is convex, we have a minimiser (that is, some $\hat{x}$ such that $\lambda(\hat{x})=\inf_x \lambda(x)$) iff $0 \in \partial \lambda (x)$ for some $x$, where $\partial \lambda (x)$ is the subdifferential at $x$.

We have $\partial \lambda (x) = \partial \|\cdot\|_2 (x) + \{ A^T \nu \} = \begin{cases} \overline{B}(0,1)+ \{ A^T \nu \}, & x=0 \\ \{ { x \over \|x\|_2 } \} + \{ A^T \nu \}, & \text{otherwise}\end{cases}$. From this we see that there is no minimiser if $\|A^T \nu \|_2 > 1$, the minimiser is $\hat{x}=- A^T \nu$ if $\|A^T \nu \|_2 = 1$, and the minimiser is $\hat{x}=0$ if $\|A^T \nu \|_2 < 1$.

From this we get $\inf_x \lambda(x) = \begin{cases} -\nu^T b, & \|A^T \nu \|_2 \le 1\\ -\infty, & \text{otherwise}\end{cases}$, which matches Solution #2. (Note that the non existence of a minimiser does not imply that the function is unbounded below, but it is easy to explicitly choose an $x$ in this case that shows that the function is unbounded below.)

Regarding Question #2, as you have (mostly) observed, we have $\inf_x \lambda(x) = \inf_t \lambda(t A^T \nu) $. Expanding, $\lambda(t A^T \nu) = |t| \|A^T \nu\|_2 - t \|A^T \nu \|_2^2 - \nu^T b$ (note again the non differentiability at $t=0$ because of the $|t|$ term).

We can either perform a case analysis or compute the subdifferential as above to compute $\inf_t \lambda(t A^T \nu) = \begin{cases} -\nu^T b, & \|A^T \nu \|_2 \le 1\\ -\infty, & \text{otherwise}\end{cases}$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.