Null- and positivstellensätze for rationally resolvable ideals (Q2359567)

Let \(\underline{X}=\{X_1,\dots,X_g\}.\) Hilbert's classical Nullstellensatz tells us that given an ideal \(I\) in the commutative polynomial algebra \(\mathbb{C}[\underline{X}],\) a polynomial \(f\) in \(\mathbb{C}[\underline{X}]\) vanishes on the zero set in \(\mathbb{C}^g\) common to all the polynomials in \(I\) iff \(f\) lies in the radical of \(I.\) The authors seek to find analogues in the noncommutative case (vanishing on matrix \(g\)-uples) for the special class of ideals they call rationally resiolvable. Many of their results give size bounds for the matrices involved; hence steps towards effectivity are made. Earlier noncommutative Nullstellensaetze are surveyed in [\textit{M. Brešar} and \textit{I. Klep}, in: Notions of positivity and the geometry of polynomials. Dedicated to the memory of Julius Borcea. Basel: Birkhäuser. 79--101 (2011; Zbl 1247.14062)]. The present setup is as follows: Let \(\mathbb{K}\) be a field and let \(<\!\underline{X}\!>\) be the family of words on the alphabet \(\underline{X}.\) Denote by \(\mathbb{K}\!<\!\underline{X}\!>\) the free associative algebra over \(\mathbb{K}\) of polynomials in noncommuting variables i.e. the family of \(\mathbb{K}\)-linear combinations of words over \(\underline{X}\); and by \(\mathbb{K}\, (\!\!\!< \underline{X}> \!\!\!\! )\) the associated free skew field of rational expressions. Ideals in \(\mathbb{K} \! <\!\underline{X}\!>\) are assumed twosided: An ideal \(\mathcal{I}\) generated by \(S\) consists of all finite linear combinations of polynomials of form \(fsg\) with \(s\in S,\) \(f,g\in \mathbb{K}\! <\!\underline{X}\!>.\) Evaluation of polynomials or rational functions is understood on \(g\)-uples of square matrices of arbitrary size i.e. on elements of \(\bigcup_n M_n(\mathbb{K})^g.\) The \textit{zero set} of \(\mathcal{I}\) is defined by \(\mathcal{Z}(\mathcal{I}):=\bigcup_n \{\underline A\in M_n(\mathbb{K})^g: p(\underline{A})=0 \text{ for all } p\in \mathcal{I} \}. \) Ideal \(\mathcal{I}\) has the \textit{Nullstellensatz property} if a polynomial \(f\in \mathbb{K}\!<\!\underline{X}\!>\) pertains to \(\mathcal{I}\) if and only if \(f_{| \mathcal{Z}(\mathcal{I})} \equiv 0.\) The aim is to give conditions in order that ideal \(\mathcal{I}\) has the Nullstellensatz property. The two class of ideals for which results are obtained are tentatively called formally or geometrically rationally resolvable (frr) or (grr) ideals. For their definition, \(g''\)-uples of rational expressions in \(g'\) variables and accompanying partitions \(\underline{X}=\{X_1',\dots,X_{g'}'\}\uplus\{X_1'',\dots,X_{g''}''\}\) have to be considered. If \(r=(r_1(\underline{X}'),\dots,r_{g''}(\underline{X}'))\) is such a \(g''\)-uple of rational expressions, the domain \(\text{dom } r\) of \(r\) is naturally defined as the family of all \(g'\)-uples of matrices \(\underline{A}\) such that the matrices \(r_1(\underline{A}),\dots, r_{g''}(\underline{A})\) all exist. Roughly speaking, the quantity \(h(r),\) the \textit{height} of \(r\), is defined via the maximal number of nested inverses in the \(r_j\); and the \textit{graph} of \(r\) is \(\Gamma(r)= \{(\underline{A},r(\underline{A})): \underline{A}\in \text{dom}(r)\}.\) \(\mathbb{K}\,(\!\!\!\!<\underline{X}>\!\!\!\!)_{\; h}\) is the subring of \(\mathbb{K}\,(\!\!\!\!<\underline{X}>\!\!\!\!)\) of rational expressions of height at most \(h.\) One now defines \(R_r\) as the subring of \(\mathbb{K}\,(\!\!\!\!<\underline{X}>\!\!\!\!)\) generated by \( \mathbb{K}\!<\!\underline{X}\!> \cup\, \mathbb{K}\,(\!\!\!\!<\underline{X}'>\!\!\!\!)_{\; h(r)}.\) In \(R_r\) the ideal \(\mathcal{I}_r= \text{ideal}(\{X_j''-r_j(\underline{X}')\}_{j=1}^{g''})\) is considered. An ideal \(\mathcal{I}\) is \textit{formally rationally resolvable} (frr) (by means of the \textit{resolvent} \(r\)) if its defining relations permit to express some of its variables, say \(\underline{X}''\) rationally by means of the other variables, say \(\underline{X}'\). In ideal theoretic language this is expressed by requiring that \(\mathcal{I}\cap K\! <\!\underline{X}'\! >=\{0\}\) and that \(\mathcal{I}_r\) is generated by \(\mathcal{I}\) in \(R_r.\) An ideal \(\mathcal{I}\) is \textit{geometrically rationally resolvable} (grr) (by means of the resolvent \(r\)) if \(\Gamma(r)\subseteq Z(\mathcal{I})\) and every polynomial vanishing on \(\Gamma(r)\) vanishes on \(Z(\mathcal{I}).\) The formulation of the first main result involves the categorical construction of a coproduct: Theorem 2.5: Let \(\mathcal{I}\) be a frr ideal. Then if the coproduct \( \mathbb{K}\!<\!\underline{X}\!>\!\!/\mathcal{I} *_{\mathbb{K}\!<\!\underline{X}'\!>} \mathbb{K}\,(\!\!\!\!<\underline{X}'>\!\!\!\!)_{\;h(r)}\) is faithful, \(\mathcal{I}\) is grr and has the Nullstellensatz property. Furthermore if an ideal \(\mathcal{I}\) has these latter two properties, then \(\mathbb{K}\!<\!X\!>\!\!/\mathcal{I}\) is imbeddable into a free skew field. The authors then give some more easily checkable sufficient conditions for an frr ideal to have a faithful coproduct; e.g. it is sufficient that \(\mathbb{K}\! <\!X\!>\!\!/\mathcal{I}\) be a Sylvester domain (see e.g. \textit{W. Dicks} and \textit{E. D. Sontag} [J. Pure Appl. Algebra 13, 243--275 (1978; Zbl 0393.16002)]) or in particular a free ideal ring. To convey the flavour of the results obtained, consider an \(f \in\mathbb{K} < X_1,\dots,X_g,Y_1,\dots,Y_g >\) so that \(f(\underline{A},\underline{B})=0\) for all \(2g\)-uples of matrices \((\underline{A},\underline{B})\) of arbitrary size for which conditions \(A_jB_j=I_n,\) \(j=1,\dots,g\) hold. Then, as a consequence to Corollary 2.8, \(f\in \text{ideal}(\{1-X_jY_j,1-Y_jX_j: j=1,\dots,g \}).\) Section 2 concludes with a number of examples showing that the assumptions of Theorem 2.5 cannot be weakened. Section 3 applies realization theory for noncommutative polynomial functions -- see e.g. [\textit{J. A. Ball} et al., SIAM J. Control Optim. 44, No. 4, 1474--1528 (2005; Zbl 1139.93006)] -- to prove bounds \(N\) for the Nullstellensatz of the following form: Let \(\mathcal{I}\) be a grr ideal. Then \(f\in\mathcal{I}\) vanishes on \(Z(\mathcal{I})\) iff \(f\) vanishes on all matrices of size \(N\) in \(Z(\mathcal{I}).\) The main point is to find bounds for testing when a rational expression is a rational identity. The path to such results is as follows. Let \(\mathcal{A}=M_m(\mathbb{K})\) and let \(\underline{Y}=\{Y_1,\dots,Y_g\}.\) Then let \(\mathcal{A}\!<\!\underline{Y}\!>=M_m(\mathbb{K}) *_\mathbb{K} <\!\underline{Y}\!>\) be the generalized noncommutative polynomials over \(M_m(\mathbb{K}).\) In this structure exist the free bimodules \(\mathcal{A}^{Y_i}= \sum \mathcal{A}Y_i \mathcal{A}. \) More generally for a word \(w \in <\!\underline{Y}\!>\) define \(\mathcal{A}^w=\mathcal{A}^{w_1}\cdots \mathcal{A}^{w_{|w|}}.\) An element in \( \mathcal{A}^w\) is of the form \([S,w]:=\sum_{i=1}^{n_w} a_{i,w}^{(0)}w_1 a_{i,w}^{(1)} w_2 \dots w_{|w|} a_{i,w}^{(|w|)},\) so abbreviated with the aim to create a \textit{generalized formal series} \(S=\sum_{w} [S,w]\) in the completion \(\mathcal{A}\ll \underline{Y}\gg.\) Series \(S\) is called \textit{recognizable} of dimension \(n\) if one can find an \(A\in (\mathcal{A}^w)^{n\times n}\) (this being hence an \(n \times n\) matrix whose entries are \(m\times m\) matrices) and vectors \({\mathbf c}\in \mathcal{A}^{1\times n}, {\mathbf b}\in \mathcal{A}^{n \times 1}\) such that for all words \(w\) we have \([S,w]={\mathbf c} A^w {\mathbf b}.\) Theorem 3.3 shows that if \(S_1,S_2,S\) are recognizable series with \(S\) being invertible then series composed from these under basic arithmetic operations are recognizable. This has an important consequence for rational expressions \(r.\) If \(r\) is such an expression in \(\underline{X}\) and \(\underline{P}\in \text{dom } r \) then it can be formally expanded around \(\underline{P}\) to yield a series in \(r(\underline{Y}+\underline{P})\in \mathcal{A}\!\ll \!\underline{Y}\!\gg\) which is recognizable; an associated triple \(({\mathbf c},A,{\mathbf b})\) is called a \textit{realization of \(r\) about \(\underline{P}.\)} Then Theorem 3.8 says that if \(r\) is an identity on matrices of size \(m\lceil \frac{mn}{2}\rceil,\) then \(r\) is a rational identity. From this then for example the following is derived. Assume \(r=(r_1,\dots,r_{g''})\) is a rational resolvent for grr ideal \(\mathcal{I}\) in the sense of the previous section and that rational functions \(r_j\) can be defined around a point in \(\text{dom } r\) and that the dimensions of these realizations are \(\leq n.\) Corollary 3.10: Under these conditions, if \(f\in \mathbb{K}\!<\!\underline{X}\!>\) is of degree \(d\) and vanishes on \(Z(\mathcal{I})\cap M_N(\mathbb{K})^g\) with \(N=m\lceil \frac{1}{2} md(g+1)^d \max(n,2)\rceil\) then \(f\) vanishes in \(Z(\mathcal{I}).\) The theme of Section 4 is to adapt and then apply previous results to obtain results for polynomials mainly in \(\mathbb{C}\!<\!\underline{X},\underline{X}^{\mathsf T}\! >,\) where in polynomial evaluations \(X_i\) and \(X_i^{\mathsf T}\) have to be substituted by matrices \(A_i\) and \(A_i^*.\) An ideal \(\mathcal{I}\) in \(\mathbb{C}\!<\underline{X},\underline{X}^{\mathsf T} \!>\) so that \(\mathcal{I}^{\mathsf T} =\mathcal{I}\) is a \(*\)-ideal. If \(\mathcal{I}\) is a \(*\)-ideal then it satisfies the \textit{\(*\)-Nullstellensatz} provided \(f\in \mathcal{I}\) if and only if \(f|_{Z_*(\mathcal{I})}=0,\) where \(Z_*(\mathcal{I})=\{\underline{A}\in M_n(\mathbb{C}): n\in \mathbb{N} \text{ and } \forall g\in \mathcal{I} \; g(\underline{A},\underline{A}^*)=0 \} .\) Consider the ideal \(\mathcal{S}=\text{ideal}(1-X_1X_1^{\mathsf T}-\cdots -X_gX_g^{\mathsf T}). \) Let \(p \in \mathbb{C}\!<\!\underline{X},\underline{X}^{\mathsf T}\!>\) be a polynomial of degree \(d\) with \(v\) terms. It is proved for example in Theorem 4.8 that if \(p\in \mathbb{C}\!<\!\underline{X},\underline{X}^{\mathsf T}\!>\) of degree \(d\) vanishes on all \(g\)-tuples \((A_1,\dots,A_g)\) of matrices of size \(\lceil \frac{g+1}{2} d v\rceil\) satisfying \(A_1A_1^*+\cdots +A_gA_g^*=I_n,\) i.e. \textit{spherical isometries}, then \(f\in \mathcal{S}.\) Similar results are presented for noncommutative trigonometric polynomials and unitary matrices. Near to the end some positivstellensaetze are also proved following work of \textit{J. W. Helton} et al. [J. Reine Angew. Math. 568, 71--80 (2004; Zbl 1039.47004)]. If \(\mathcal{Z}\) is a set of \(g\)-tuples of matrices, then a polynomial \(f\in \mathbb{C}\!<\!\underline{X},\underline{X}^{\mathsf T}\!>\) is positive on \(\mathcal{Z}\) if \(f(\underline{A})\) is positive definite for all \(\underline{A}\in \mathcal{Z}.\) Let \(\mathcal{P}_d \subseteq \mathbb{C}\!<\!\underline{X},\underline{X}^{\mathsf T}\!>\) be the subspace of polynomials of degree\(\leq d.\) Let \(\mathcal{C}_{2d}= \text{cone}\{pp^{\mathsf T}:p\in \mathcal{P}_d\}\) be the associated convex cone of sums of Hermitian squares. Corollary 4.14: If \(f\in \mathbb{C}\!<\!\underline{X},\underline{X}^{\mathsf T}\!>\) of degree \(d-1\) is positive on all spherical isometries of size \((2g+1)^d,\) then \(f\in \mathcal{C}_{2d}+\mathcal{S}.\) In an appendix, some \(*\)-Nullstellensaetze are reproved (without size bounds) by ideas inspired from functional analysis.