I want to solve the two optimization problems:

$$ (1) \text{ } \min f(x) \text{ subject to } g(x) = 0 $$


$$ (2) \text{ } \max f(x) \text{ subject to } g(x) = 0 $$

separately of course (one then the other).

I was thinking of defining the auxiliary equations:

$$ L_1(x,\lambda) = f(x) + \lambda g(x) $$

for the minimization version and for the maximization:

$$ L_2(x,\lambda) = - f(x) + \lambda g(x) $$

since both solutions satisfy the standard $ \nabla f(x) = \lambda \nabla g(x) $ (i.e. that the lines and gradients are parallel) it seems I can't distinguish when I am getting a max or a min. So which one am I suppose to use for each one?

Or perhaps I can't distinguish them by designing a auxiliary equation and instead just optimize $L_1$ with gradient descent to optimize constraint (1) and gradient ascent to optimize $L_1$ to get (2) OR gradient descent on $L_2$? I am really confused because I don't understand why my optimization problem if I try to minimize $L_2$ with GD it wouldn't just drive $-f(x)$ to -infinity and ignore g(x) (since if $-f(x) = -\infty $ then $L_2$ is minimized...but it seems to have the weird property that it ignore my constraint... so did I design my auxiliary function wrong or am I using the wrong algorithm or what am I doing wrong?)

To add context is because I am trying to do the following:

I wanted to design an optimization problem where I maximize a non-negative quantity and another one where minimize it both subjected to a constraint.

$$ \min \| r \| \text{ subject to } f(x+r,w) = l$$

and separately then compute:

$$ \max\| r \| \text{ subject to } f(x+r,w) = l$$

the reason I am doing this is in the context of Neural Networks and adversarial examples. Thus, $f(x+r,w)$ is some complicated (non-convex) Neural Network function. In this framework I was wondering what exactly guaranteed as they claim in the paper that they get:

the closest image to x classified as l by f.

For example how do they know its not the farthest imagine classified as $l$? Perhaps not exactly fartherst because the problem might be non-convex but at least find the locally furthest image. In this case it would be sort of the furtherest the image $x$ would go and then change label.

The reason I am asking is because I am interested is comparing the difference between the two cases (a max and a min). I thought that a natural way to design the second one would be with the standard auxiliary equation:

$$ \min \| r \| + \lambda f(x+r,w) $$

(lets ignore the $x+r \in [0,1]^m$ constraint for a moment). Since one can just let $c = \frac{1}{\lambda}$ I assume its equivalent to solve:

$$ \min c\| r \| + Loss(f(x+r,w),y) $$

however what is not clear to me is what if we want to solve the maximization problem:

$$ \max\| r \| \text{ subject to } f(x+r,w) = l$$

do we just do:

$$ \min -\| r \| \text{ subject to } f(x+r,w) = l$$

and solve:

$$ \min -\| r \| + c Loss(f(x+r,w),l)$$


$$ \min -c \| r \| + Loss(f(x+r,w),l)$$

what feels odd to me is that in this case we can just arbitrarily make $\| r\|$ large so that the above is minimized at minus infinity. Is that right? I am confused about that. Or perhaps we just consider:

$$ L(r) = -\| r \| + c Loss(f(x+r,w),l) $$

and use gradient ascent instead of gradient descent. Is the algorithm Im using wrong or the auxiliary equation I'm trying to design is wrong...what is wrong?

Note: I am ok with local maximums or minimums.

  • $\begingroup$ Apriori you cannot distinguish in the same way that the unconstrained $\max$ or $\min$ of a function both have zero slope. You need additional information to determine if a stationary point is a $\min$, $\max$ (or saddle). $\endgroup$ – copper.hat Apr 20 '18 at 3:41
  • $\begingroup$ @copper.hat thats why I suggested Gradient Descent vs Gradient Ascent. What I am confused is, do I just optimize $L_1$ with Gradient Descent (GD) for a solution with a minimization and optimize $L_1$ with Gradient Descent (GD) for the maximization solution or do I just do Gradient Descent on $L_2$? I don't care about getting stuck in local solution (or even the risk of saddle points) $\endgroup$ – Pinocchio Apr 20 '18 at 3:44
  • $\begingroup$ I'm not exactly sure what you are asking. In general, there is no guarantee that you will reach an extremum. One generally tries to solve the Lagrange equations explicitly only in cases where it is easy to do so. In other cases you need to look for a numerical solution in which case you might be better off looking for an algorithm that explicitly handles equality constraints (reduced gradients, for example). $\endgroup$ – copper.hat Apr 20 '18 at 3:54
  • $\begingroup$ @copper.hat I thought my question was clear, my apologies if its not. I am interested in an algorithms that approximately solves (1) and (2). I know I am using a non convex objective because I am using Neural Nets so I never expected any sort of Guarantees. For (1) the literature seems to use Gradient Descent with unconstrained problem $f(x) + \lambda g(x)$. So if I solve that same problem but use Gradient Ascent do I get a solution more in the spirit of (2)? Notice that I am ok with any approximation. $\endgroup$ – Pinocchio Apr 20 '18 at 3:58
  • $\begingroup$ @copper.hat Notice that in principle GD could get stuck in a local max. Thats the type of approximation algorithm I expect, some algorithm that can get stuck in saddle points but won't get stuck in maximas if its looking for minimas (even though it could in principle but its seems ridiculously unlikely) or the other way round. $\endgroup$ – Pinocchio Apr 20 '18 at 4:00

Your Answer

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

Browse other questions tagged or ask your own question.