Kobayashi-Hitchin correspondence for tame harmonic bundles. II. (Q1006129)

Assume \(X\) is a smooth irreducible complex projective variety with ample line bundle \(L\), and \(D\) a simple normal crossing hypersurface. In this impressive work, the author proves in this very general setting a Kobayashi-Hitchin type correspondence (between harmonic bundles on \(X\setminus D\) and stable parabolic flat bundles) and derives from it various important corollaries. The paper under review extends his previous work [Asterisque No. 309 (2006; Zbl 1119.14001)], where the parameter \(\lambda\) was considered as trivial (see below for details). A particular case of this result is a quasi-projective version of the Corlette-Simpson correspondence between Higgs bundles and flat bundles. Twenty years ago, \textit{C. Simpson} showed an existence result of Hermitian-Einstein metrics for a stable Higgs bundle on an open manifold under the assumption of the existence of an initial metric with special curvature properties [J. Am. Math. Soc. 1, 867--918 (1988; Zbl 0669.58008)]. This allowed him to solve the one-dimensional case. Later, for a smooth divisor, \textit{O. Biquard} extended Simpson's work [Ann. Sci. Éc. Norm. Supér. (4) 30, No. 1, 41--96 (1997; Zbl 0876.53043)]. First of all, we state the main results of the paper.{\parindent=6mm\begin{itemize}\item[1.] Let \((E_*,\mathbb{D}^\lambda)\) be a regular filtered \(\lambda\)-flat bundle on \((X,D)\). Then \((E_*,\mathbb{D}^\lambda)\) is \(\mu_L\)-polystable with trivial characteristic numbers if and only if there exists a unique pluriharmonic metric \(h\) on \((E_{X\setminus D}, \mathbb{D}^\lambda)\) adapted to the parabolic structure. Here, to be pluriharmonic means to be harmonic when restricted to each curve. \item[2.] There is an equivalence of categories between the category of \(\mu_L\)-polystable regular filtered Higgs bundles on \((X,D)\) with trivial characteristic numbers and the category of \(\mu_L\)-polystable regular filtered flat bundles on \((X,D)\) with trivial characteristic numbers.\item[3.] For \((E_*,\mathbb{D}^\lambda)\) a regular filtered \(\lambda\)-flat bundle on \((X,D)\), there is a Bogomolov-Gieseker type inequality for the parabolic characteristic numbers for \(E_*\). \end{itemize}} Note that one recovers from these results the existence of a Corlette-Jost-Zuo metric for a simple flat bundle on \(X\setminus D\), i.e. when the relative representation of \(\pi_1(X\setminus D)\) is assumed to be simple. We now explain some of the notions used in the paper. Here a subtlety is to define the right compatibility conditions that have to be satisfied by all the filtrations which come into play. First of all, a regular filtered \(\lambda\)-flat sheaf \(E_*=\{\mathbf{E},\{ \,_cE\}\}\) is a filtered sheaf on \((X,D)\) such that:{\parindent=6mm\begin{itemize}\item[(a)] \(\mathbf{E}\) is a torsion free coherent \(\mathcal{O}_X(*D)\)-module,\item[(b)] \(\{\,_cE\}\) is an increasing filtration by coherent submodules of \(\mathbf{E}\) such that \(\mathbf{E}_{X\setminus D}=\,_cE_{|X\setminus D}\) for any index \(c\in \mathbb{R}^S\), \(_aE = \bigcap_{a<b} E\), \(\mathbf{E}=\bigcup_{a} \,_aE\), \item[(c)] \(\,_{a'}E=\,_aE\otimes \mathcal{O}_X(-\sum_j n_j\cdot D_j)\) as submodules of \(\mathbf{E}\) where \(a'=a-(n_j|j\in S), \) \item[(d)] for any index \(c\), the filtration \[ ^i \mathcal{F}_d=\bigcup_{a_i\leq d, \mathbf{a}\leq c} \,_{\mathbf{a} } E \] induces a \(c\)-parabolic sheaf \(\,_cE_*=\{\,_cE, \{^i \mathcal{F}\, |i\in S\}\}\), i.e. the sets \[ \{a|\,^i Gr_a^\mathcal{F}(\,_cE):= \,^i \mathcal{F}_a(\,_cE)/ \,^i \mathcal{F}_{<a}(\,_cE)\} \} \] are finite. \end{itemize}} This notion of regular filtered sheaf is essentially equivalent to the notion of parabolic sheaf that was described in [Asterisque No. 309 (2006; Zbl 1119.14001)], where the notion of slope stability \(\mu_L\) is also introduced. On another hand, a flat logarithmic \(\lambda\)-connexion is a map \(\mathbb{D}^\lambda: E\rightarrow E\otimes \Omega^{1,0}_X(\log D)\) satisfying:{\parindent=6mm\begin{itemize}\item[(i)] the Leibniz rule \[ \mathbb{D}^\lambda(f\cdot s) = f\cdot \mathbb{D}^\lambda(s) + \lambda \cdot d_X(f)\cdot s \] where \(f\) is a holomorphic section of \(\mathcal{O}_X\) and \(s\) of \(E\), \item[(ii)] the flatness condition \(\mathbb{D}^\lambda \circ \mathbb{D}^\lambda = 0 \), \item[(iii)] \(\mathbb{D}^\lambda( \,^i \mathcal{F}_a ) \subset \,^i \mathcal{F}_a \otimes \Omega^{1,0}_X(\log D)\). \end{itemize}} A tuple \((\mathcal{E},\mathbb{D}^\lambda)\) formed by a \(c\)-parabolic structure on \(X\setminus D\) and a flat logarithmic connexion is called a regular parabolic \(\lambda\)-flat sheaf. There is one major difficulty to extend Simpson and Biquard's results, due to the non compatibility of the nilpotent parts of the residues on the graded pieces. Thus, the author introduces a perturbation method of the parabolic structure to obtain a parabolic structure which is graded semi-simple, that is the residues are essentially semisimple. For this new structure, it is possible to construct an initial metric with the required curvature properties, and if this deformed structure is \(\mu_L\)-stable, one obtains the existence of a Hermitian-Einstein metric adapted to the parabolic structure. The problem is now to derive the existence of a pluriharmonic metric at the limit (i.e when doing a perturbation as ``small'' as possible). The convergence is delicate and coming from a variant of the work of S.K. Donaldson. Firstly, the author studies this problem over a curve and proves a \(L^2_1\) convergence result using Donaldson's functionals. For the general case, he proves the convergence when the metrics are restricted to an ample curve in \(X\). One can remark at that stage that in parallel, the slope stability holds when restricting the base to such a curve (hence there is a Mehta-Ramanathan type theorem for the considered structures). Finally, using the first two steps, the convergence of the sequence of metrics is guaranteed almost everywhere on \(X\setminus D\) and with some extra argument, the limit is plurisubharmonic. Another important feature in the proof is the notion of `Kashiwara-Malgrange-Sabbah-Simpson' spectrum (KMS spectrum). The residue of a harmonic bundle on a component of \(D\) breaks up into pieces corresponding to the different elements of this spectrum. We believe that for regular filtered flat-bundle, the results contained in the paper are optimal. Without any doubt, they will have in a long term deep consequences in both geometry and topology. Finally, the non regular case is studied in a recent preprint of the author [``Wild harmonic bundles and wild pure twistor \(D\)-module'', \url{arXiv:0803.1344}].