Sufficient conditions for a real polynomial to be a sum of squares of polynomials (Q309665)

\textit{J. B. Lasserre}'s paper [Arch. Math. 89, No. 5, 390--398 (2007; Zbl 1149.11018)] was the first to give nontrivial sufficient conditions, solely using the coefficients of a polynomial in several variables, in order for it to be a sum of squares. A number of such conditions have since been found and the present paper contributes with conditions encompassing about all of its predecessors. These latter suffer from the fact that they involve the coefficients of the one-variable-monomials of a polynomial too decisively. Also formulated are geometric programs to find lower bounds for \(f_*=\inf\{f(x): x\in \mathbb{R}^n\}\) replacing earlier such programs which also suffer from the referred handicap. Let the homogeneous polynomial \(f\in \mathbb{R}[x]=\mathbb{R}[x_1,...,x_n]\) of degree \(2d\) be written as \(\sum_\alpha f_\alpha x^\alpha,\) with \(x^\alpha=x_1^{\alpha_1} \cdots x_n^{\alpha_n}.\) Let \(\Delta=\{\alpha: f_\alpha<0 \text{ or } \alpha \text{ has odd entries} \}.\) Let \(\Gamma(f)\) be the Newton polytope of \(f\) and \(V\subset (2\mathbb{Z})^n\) be its vertices; further put \(\mathcal{C}= \Gamma(f)\cap (2\mathbb{Z})^n;\) \(\mathcal{A}:=\{\frac{1}{2}(s+t): s,t\in \mathcal{C} \};\) and \(\mathcal{V}:=\) the points of \(\mathcal{A}\) that cannot be written as average of distinct points of \(\mathcal{A}.\) Theorem 2.6: If there exists a framework \(\mathcal{U}\) in \textit{B. Reznick}'s sense [Math. Ann. 283, No. 3, 431--464 (1989; Zbl 0637.10015)] with \(\mathcal{U}\)-mediated set \(\mathcal{U}^* \) so that the conditions \(V\subseteq \mathcal{U} \subseteq \mathcal{V};\) \(\Delta \subseteq \mathcal{U}^*;\) \(\min_{u\in \mathcal{U}} f_u \geq \sum_{\alpha\in \Delta} |f_\alpha|\) are satisfied, then \(f\) is a sum of squares of polynomials. The point of this theorem is that it replaces the conditions given by Lassere [loc. cit], \textit{C. Fidalgo} and the reviewer [Math. Z. 269, No. 3--4, 629--645 (2011; Zbl 1271.11045)], and \textit{M. Ghasemi} and \textit{M. Marshall} [Arch. Math. 95, No. 4, 343--353 (2010; Zbl 1205.12001)] and \textit{S. K. Khanduja} and \textit{S. Kumar} [J. Algebra Appl. 12, No. 1, Paper No. 1250125, 10 p. (2013; Zbl 1271.12004)] which involve lower bounds on the coefficients \(f_{2de_i}\) of the pure powers \(x_i^{2d}\) by the weaker ones given above. (The family \(\{2de_i\}_{i=1}^n\) (\(e_i\) standard vectors) is an example of a framework.) In a similar vein, the other main result generalizes Ghasemi and Marshall 2's [loc. cit, Theorem 2.3]: Theorem 2.12: Let \(f\) be a form and suppose that \(\Gamma(f)\) is a simplex with vertex set \(V=\{u^1,...,u^m\}\) and \(\Delta \subseteq V^*.\) A sufficient condition for \(f\) to be sos is that there exist nonnegative reals \(a_{\alpha,i}\) for each \(\alpha \in \Delta,\) \(i=1,...,m,\) such that \(a_{\alpha}^{\lambda_\alpha} =|f_\alpha| \lambda_\alpha^{\lambda_\alpha}\) for all \(\alpha \in \Delta\), and \(f_{u^i}\geq \sum_{\alpha \in \Delta} a_{\alpha,i},\) \(i=1,...,m.\) Here \(\lambda_\alpha =(\lambda_{\alpha,1},...,\lambda_{\alpha,m})\) are the barycentric coordinates of \(\alpha\) with respect to \(V.\) From these two theorems a host of corollaries is spelled out, showing that the conditions on the coefficients of polynomials to be sos found by the cited previous workers are consequences of above theorems. In a fourth section this theorem is used, similarly as in Ghasemi Marshall 2[loc.cit], to find lower bounds for polynomials by means of geometric programming. The main difference is that to find lower bounds it is now not anymore necessary to consider the family \(\{f_{2de_i}\}_{i=1}^n\) of coefficients. In particular even if some of the \(f_{2d e_i}\) are zero, their programs will work. It should be said that a recent article by \textit{S. Iliman} and \textit{T. de Wolff} [SIAM J. Optim. 26, No. 2, 1128--1146 (2016; Zbl 1380.12001)] replaces the use of sums of squares for finding lower bounds for \(f_*\) by the use of sums of nonnegative circuit polynomials. It also gets rid of the relevance of the family \(\{f_{2de_i}\}_{i=1}^n.\) There is some overlap between these articles which have apparently been written within the same year 2014, without the authors being aware of each other.