This page is created and supported by Lyudmila Chernyshenko, a first-year undergraduate student of the Mathematics and Mechanics Faculty of the Saint Petersburg State University. The materials on this page are aimed at undergraduate students.
We begin with the idea:
In this section we give a brief introduction to sum of squares polynomials technique introduced in Pablo Parrilo’s thesis in 2000 [1], and show how the existence of a sum of squares decomposition can be verified using semidefinite programming.
Note: Semidefinite programming is a subfield of convex optimization concerned with the optimization of a linear objective function (that is, a function to be maximized or minimized) over the intersection of the cone of positive semidefinite matrices with an affine space.
Definition: for х ϵ Rn a multivariate polynomial p(x)is a sum of squares if there exist polynomials fi(x) i=1, … ,M such that
p(x)=
Proposition: A polynomial p(x) of degree 2d is a sum of squares if and only if there exists a positive semidefinite matrix Q and a vector of monomials Z(x), containing monomials of the form ax1r1x2r2...xnrn, where r1+...rn ≤ d such that p(x)=ZT(x)QZ(x).
In general, the monomials in Z(x) are not algebraically independent. Expanding ZT(x)QZ(x) and equating the coefficients of the resulting monomials to the ones in p(x), we obtain a set of affine relations in the elements of Q. The problem of finding a semi-definite matrix Q, the elements of which satisfy the obtained affine relations, can be solved using semi-definite programming. As soon as we find that for p(x) there exists a corresponding positive-semidefinite matrix Q, it is proved that p(x) can be represented as a sum of squares.
хTАх= = +…+ , and +…+ ≥0 is valid for all xif and only if ≥0.
Example (from [2]): Suppose that we want to know whether or not the quartic polynomial in two variables p() =
2+2-+5 is a sum of squares.
For this purpose, define
Z(x)=Т
and consider the following quadratic form:
p() =
2+2-+5
= Z(x)ТZ(x)
= + + +2 +2 .
From this we get the following relations:
Now, using semi-definite programming, we can find
and ,
satisfying the last equation and such that the matrix Q is positive semi-definite. These conditions are satisfied with и .
After that we find a basis in which matrix Qwill be diagonal. Monomials in Z(x) can be expressed as a linear combination of the basis vectors. By the formula p(x)=Z(x)TQZ(x) we obtain the following sum of squares decomposition:
p() = .
Note: p(x) being sum of squares implies that p(x)≥0 for all х ϵ Rn. However, the converse is not always true. Not all nonnegative polynomials can be written as sum of squares, apart from three special cases:
As we have already mentioned, a semidefinite matrix, such that its elements satisfy a given set of affine relations, if exists, can be found using semidefinite programming. This is so because the set of positive-semidefinite matrices is convex.
Info: A set is convex if for every pair of points within the set, every point on the straight line segment that joins the pair of points is also within the set. In other words, a set G ϵ Rn is convex if for every pair of points x1 ϵ G, x2 ϵ G and for any λ, 0≤λ≤1 it is true that λx1+(1-λ)x2 ϵ G. |
For verifying this we take two matrices from this set, and verify that the matrix A=+, where 0≤λ≤1, will also be positive semidefinite. Consider xТAx= xТ +)x, where 0≤λ≤1. Since the matrices are positive-semidefinite and because of the conditions on λ, we obtain that xТAx≥0, which means that the matrix A is positive semidefinite. Hence, the set of such matrices is indeed convex.
We are looking for a matrix from convex set of positive-semidefinite matrices such that the elements of this matrix satisfy a set of affine relations. Since affine relations form a hyperplane, the desired matrix is located in the intersection of a convex set of positive semidefinite matrces and a hyperplane. This intersection will be convex, so the problem of searching for the matrix can indeed be solved using semidefenite programming.
This page is based on [2].
1. P. A. Parrilo. Structured Semidefinite Programs and Semialgebraic Geometry Methods in Robustness and Optimization. PhD thesis, California Institute of Technology, Pasadena, CA, 2000.
2. A. Papachristodoulou and S. Prajna. A Tutorial on Sum of Squares Techniques for Systems Analysis. 2005 American Control Conference June 8-10, 2005. Portland, OR, USA.