patternMinor
What is the significance of the vector dimension in semidefinite programming relaxations?
Viewed 0 times
significancetherelaxationssemidefinitewhatprogrammingdimensionvector
Problem
Let's say that we want to design a semi-definite programming approximation for an optimization problem such as MAX-CUT or MAX-SAT or what have you.
So, we first write down an integer quadratic program that solves the problem exactly. For example, in the case of MAX-CUT we would write:
$$
\max \sum_{\{u,v\} \in E}{\frac{1 - x_u x_v}{2}}
$$
$$
x_v \in \{-1,+1\}.
$$
Then we relax the program by allowing the variables to take on vector values in the sphere $S^{n-1}$, and replacing products with scalar products, obtaining a semi-definite program that can be solved precisely.
Our approximation proceeds by solving the SDP and rounding the solution to obtain a legal solution for the original problem whose distance from the SDP solution can be bounded.
My question is: What is the importance of the dimension of the vectors? In all the cases I have seen, the analysis would stay the same if we assume that the vectors are all unit vectors in $\mathbb{R}^2$. So why do people always take vectors in $S^{n-1}$?
So, we first write down an integer quadratic program that solves the problem exactly. For example, in the case of MAX-CUT we would write:
$$
\max \sum_{\{u,v\} \in E}{\frac{1 - x_u x_v}{2}}
$$
$$
x_v \in \{-1,+1\}.
$$
Then we relax the program by allowing the variables to take on vector values in the sphere $S^{n-1}$, and replacing products with scalar products, obtaining a semi-definite program that can be solved precisely.
Our approximation proceeds by solving the SDP and rounding the solution to obtain a legal solution for the original problem whose distance from the SDP solution can be bounded.
My question is: What is the importance of the dimension of the vectors? In all the cases I have seen, the analysis would stay the same if we assume that the vectors are all unit vectors in $\mathbb{R}^2$. So why do people always take vectors in $S^{n-1}$?
Solution
Vector programs are solved by solving semidefinite programs. The basic idea is that if $v_1,\ldots,v_n$ are vectors, then the matrix $V_{ij} = \langle v_i,v_j \rangle$ is positive semidefinite:
$$ \alpha' V \alpha = \sum_{i,j} \alpha_i \alpha_j \langle v_i, v_j \rangle = \left\|\sum_i \alpha_i v_i\right\|^2. $$
The converse is also true: a positive semidefinite $n \times n$ matrix $V$ of rank $r$ can be written as $V = LL'$, where $L$ is $n \times r$ (this is known as the Cholesky decomposition). Furthermore, $L$ can be found effectively given $V$. If $\ell_1,\ldots,\ell_n$ are the $r$-dimensional rows of $L$, then $V_{ij} = \ell_i \ell'_j = \langle \ell_i, \ell_j \rangle$. We can therefore optimize a vector program involving inner products of vectors $v_i$ by solving a semidefinite program. Rank constraints are not convex, and so we cannot really control the dimension of the vectors other than the trivial bound $n$.
As you mention, the analysis in the particular case of MAX-CUT only uses the two-dimensional projections. This shows that even if you could somehow guarantee that $\operatorname{rank} V = 2$ (recall that the rank corresponds to the dimension), the analysis wouldn't get better. On the other hand, if you could guarantee than $\operatorname{rank} V = 1$, you will have solved MAX-CUT exactly. This shows that the latter constraint is NP-hard to realize. It is probably the case that any non-trivial rank constraint cannot be enforced in polynomial time, and any constant rank constraint is NP-hard.
$$ \alpha' V \alpha = \sum_{i,j} \alpha_i \alpha_j \langle v_i, v_j \rangle = \left\|\sum_i \alpha_i v_i\right\|^2. $$
The converse is also true: a positive semidefinite $n \times n$ matrix $V$ of rank $r$ can be written as $V = LL'$, where $L$ is $n \times r$ (this is known as the Cholesky decomposition). Furthermore, $L$ can be found effectively given $V$. If $\ell_1,\ldots,\ell_n$ are the $r$-dimensional rows of $L$, then $V_{ij} = \ell_i \ell'_j = \langle \ell_i, \ell_j \rangle$. We can therefore optimize a vector program involving inner products of vectors $v_i$ by solving a semidefinite program. Rank constraints are not convex, and so we cannot really control the dimension of the vectors other than the trivial bound $n$.
As you mention, the analysis in the particular case of MAX-CUT only uses the two-dimensional projections. This shows that even if you could somehow guarantee that $\operatorname{rank} V = 2$ (recall that the rank corresponds to the dimension), the analysis wouldn't get better. On the other hand, if you could guarantee than $\operatorname{rank} V = 1$, you will have solved MAX-CUT exactly. This shows that the latter constraint is NP-hard to realize. It is probably the case that any non-trivial rank constraint cannot be enforced in polynomial time, and any constant rank constraint is NP-hard.
Context
StackExchange Computer Science Q#26262, answer score: 4
Revisions (0)
No revisions yet.