Abstract Lagrangian

ML

An optimization method to find local minima or maxima of integer/combinatorial problems efficiently.

1. Summary

Let us consider an example of integer problem

\[ \max_{x} f(x) \text{ s.t. } x\in \{0, 1 \} \]

Then, the above problem can be rewritten as \(\max_{x} f(x), s.t. x\in[0,1], x-x^2\le0\) without loss of generality.

The abstract Lagrangian of the original problem can be defined by

\[ \max_{x} L(x,\mu)=f(x)-\mu(x-x^2), x\in[0,1] \]

2. Principal of the abstract Lagrangian Duality

For large values of \(\mu_i\), the optimization problem of abstract duality is equivalent to the original problem,
which means \(d^* = \min_{\mu}\max_{x}L(x,\mu)=\max_{x}\min_{\mu} L(x,\mu)=p^*.\)