There are significantly more nonnegative polynomials than sums of squares (Q2480563): Difference between revisions

The paper gives quantitative estimates between the cones of nonnegative polynomials, cones of sums of squares and cones of sums of even powers of linear forms in the form of bounds of volumes of compact sections of these cones (or more precisely closely related ones). These are asymptotically exact as the degree is fixed and the variable number tends to infinity. They show that there are significantly more nonnegative polynomials than sums of squares which in turn are much more numerous than sums of even powers of linear forms. The paper can be seen as continuing research begun by \textit{G. Blekherman} [Discrete Comput. Geom.32, 345-371, (2004; Zbl 1071.22011)]. Fix positive integers \(n,\) and \(k\) and let \(P_{n,2k}; Pos=Pos_{n,2k}, Sq=Sq_{n,2k}, Lf=Lf_{n,2k},\) be respectively the vectorspace of real polynomials, and the cones of nonnegative polynomials, sums of squares of polynomials, and \(2k\)-th powers of linear forms in \(\mathbb R[x_1,x_2,...,x_n]\) that are homogeneous of degree \(2k.\) Let \(\sigma\) be the rotation invariant probability measure on the unit sphere \(S^{n-1}.\) Define \(M=M_{n,2k}=\{p\in P_{n,2k}: \int_{S^{n-1}} p d\sigma =0\}.\) This is a subspace of codimension 1 in \(P_{n,2k}.\) Let \(B_M\) be unit ball in \(M\) with respect to the \(L^2\)-norm (defined by \(\sigma\)), and \(D_M={n+2k-1\choose 2k}-1\) be its dimension. For \(X\) any one of the sets \(Pos, Sq, Lf,\) define \(\tilde{X}=\{p\in M: p+(x_1^2+...+x_n^2)^k \in X\}.\) The main theorem is: There exist constants \(c_X(k), c_X'(k)\) only dependent on \(k\) and reals \(l_X,l_X'\) such that for any \(\epsilon>0\) and all sufficiently large \(n\) there holds: \(c_X n^{l_X} \leq (vol(\tilde{X})/vol(B_M))^{1/D_M} \leq c_X' n^{l_X'}.\) Here \(l_X= -1/2, -k/2, -k+(1/2),\) and \(l_X'= -1/2, -k/2, -k+(1/2)+\epsilon\) for the cases \(X=Pos, Sq, Lf, \) respectively. The proofs give even more explicit estimates. The lower bound in the \(\tilde{X}=\widetilde{Pos}\) case uses a general relationship between the gauge and the volume of a convex body found in [\textit{G. Pisier}, The volume of convex bodies and Banach Space Geometry. Cambridge Tracts in Mathematics, 94. (Cambridge) etc.: Cambridge University Press. (1989; Zbl 0698.46008)], estimates of \textit{A. Barvinok} [Found. Comput. Math. 2, No. 4, 393-412 (2002; Zbl 1049.90072)], and estimates for Gegenbauer's ultraspherical polynomials. Along the way more standard estimates (Hölder, Jensen) as well as observations on the invariance of \(M\) under a natural action of SO\((n)\) are used in certain integrations. The upper bound for the \(\widetilde{Pos}\) case uses a version of the Blaschke-Santaló inequality, see \textit{M. Meyer} and \textit{A. Pajor} [Arch. Math. 55, 82--93 (1990; Zbl 0718.52011)], namely \(\text{vol} K\cdot \text{vol} K^o \leq (\text{vol} B)^2\) (\(K^o\) being the polar of \(K\)) for the case \(K=Pos\) (where the Santaló point is the origin), and a classical result of \textit{O. D. Kellogg} [Math. Z. 27, 55--64 (1927; JFM 53.0082.03)] according to which for \(f\in P_{n,2k},\) \(| | \langle\nabla f, \nabla f\rangle| | _{\infty} =4k^2 | | f| | ^2_{\infty},\) where \(| | \cdot| | _{\infty}\) is to be understood the sup-norm on unit sphere \(S^{n-1}.\) The bounds for deriving the inequalities for the cases \(\tilde{X}= \widetilde{Sq}, \widetilde{Lf}\) are also proved using duality and convexity inequalities, but as a further ingredient involve the introduction of an additional metric on \(P_{n,2k}\) which the author calls differential metric. \quad Lots of interesting and advanced mathematics have to be learned by anyone desiring to understand the details of the pillars on which this well written paper rests. Reviewer's remark: A typo: p. 359, definition of \(G_K(v)\) should read `\(\dots v\in \lambda K.\)'