Forms derived from the arithmetic-geometric inequality (Q1098872)

A real homogeneous polynomial (a form) \(p\) is called positive semidefinite (or psd) if \(p(x)\geq 0\) for all \(x\in\mathbb R^ n\); it is called sos if it is a sum of squares of forms. For \(x\in\mathbb R^ n\) and \(u\in\mathbb Z^ +_ n\), let \(x^ u=\prod x_ j^{u_ j}\). Let \(\mathcal U=\{u_ i\}\) be a set of even \(n\)-tuples satisfying \(\sum _{j}u_{ij}=2d\) for each \(i\) (\(x^{u_ i}\geq 0\) has degree \(2d\)) and let \(C(\mathcal U)=cvx(\mathcal U)\cap\mathbb Z^ n\). For \(w=\sum \lambda _ iu_ i\in C(\mathcal U)\), \(\lambda _ i\geq 0\), \(\sum \lambda _ i=1\), \(f(\mathcal U,\lambda,w)(x):=\sum \lambda _ ix^{u_ i}-x^ w\) is psd by the arithmetic-geometric inequality (AGI). A positive multiple of \(f\) is called an agiform. Hilbert proved that not every psd form is sos. The first explicit example, produced by Motzkin in 1967, is the agiform \(M(x,y,z)=x^ 4y^ 2+x^ 2y^ 4+z^ 6-3x^ 2y^ 2z^ 2\). Other non-sos agiforms have been found by Choi, Lam and the author. Hurwitz proved that every agiform \(\sum c_ ix_ i^{2d}-2d\prod x_ i^{c_ i}\) is sos. In this paper we define \(\mathcal U^ *\), a computable subset of \(C(\mathcal U)\), which has the property that, if the agiform \(f=f(\mathcal U,\lambda,w)\) is sos, then \(w\in\mathcal U^ *\). Conversely, if \(\mathcal U\) is the vertex-set of a simplex (a ``trellis'') and \(w\in\mathcal U^ *\), then \(f\) is sos; the proof is an algorithm for representing \(f\) as a sum of squares. As a result, if \(f\) is an agiform and \(k\geq \max (2,n-2)\), then \(f(x_ 1^ k,\ldots,x_ n^ k)\) is sos. A psd form \(p\) is called extremal if \(p=\sum h_ i\), \(h_ i\) psd, implies \(h_ i=\alpha _ ip\). Choi and Lam proved that \(M(x,y,z)\) and two other agiforms they introduced, \(S(x,y,z)\) and \(Q(x,y,z,w)\), are extremal. (Most other known extremal forms are squares of indefinite forms.) If \(f(\mathcal U,\lambda,w)\) is extremal, then \(\mathcal U\) is a trellis. We derive a necessary and sufficient condition for an agiform to be extremal, involving the congruence properties (mod 2) of \(w\) and \(C(\mathcal U)\). The extremal agiforms \(M\), \(S\) and \(Q\) generalize to families of extremal agiforms in many variables; the constructions are built on those of Motzkin, Choi and Lam.