Proper explanation of KKT equations and Basic constraint qualification works?

78 Views Asked by At

I am learning Non-Linear optimisation. Last time I read about KKT (Prof notes and some googling) and Basic constraint qualification and somehow tried to convince myself this and that but now when I am revising it, I feel that I am just storing everything without proper understanding of what is going on.

  1. For equality constraint we form a function (Lagrangian function) , and try to optimise it using unconstrained optimisation i.e $ \nabla f(x) $. Ques: Why does this work ? How can i visualise it?

  2. Then for inequality constraints we can write every inequality constraint as equality constraint :

$$ g(x) \ge C $$ is equal to

$$ g(x) + S^2 = C $$

Now equality constraint problem can be solved as we solved 1. We will get following conditions

$$ \nabla f(x) + \sum{\lambda \nabla g(x)} $$ $$ g(x) \ge C $$ $$ h(x) = K $$ $$ \lambda(i) * g(x) = 0 $$

Now till here It made sense to me if I assumed 1 to be correct. After that I read that these are necessary conditions but not sufficient conditions and since then I am scrolling over and over but unable to find a proper place that summarises what is going on.

I read (in prof notes) that BCQ : Basic constraint qualification, says that if BSC is satisfied the KKT multipliers exist. Now what is meaning of they exist ? They exist as in can be found ? That I can know from solving the above 4 equations is not it?

If it means that BCQ is sufficient condition, later he writes that holding BCQ doesn't mean that a given x* is minima. Now if it doesn't hold why we were dealing with BCQ in first place?

I just want to ask :

  1. Why Lagrangian method of optimising works?
  2. How KKT Works, what are necessary and sufficient conditions?
  3. What are we trying to do using BCQ and why even if BCQ is satisfied the multipliers doesn't exist?
  4. What is the meaning of "multipliers exist"?
  5. Something that you learned and feel that this should be known for better understanding of this topic.