Towards a classification of modular compactifications of \(\mathcal {M}_{g,n}\) (Q2377345)

Moduli spaces, arguably, are among the most fundamentally important developments in modern algebraic geometry. One of the paramount interests in, and applications of, them is the intersection theory that results from a suitable compactification. Although a compactification is far from unique, often the most natural ones are modular in the sense that the limiting points one adds to compactify the space represent degenerations, usually singular, of the objects being parameterized. The absolutely canonical example of this idea is Deligne-Mumford's stable curve compactification of the moduli space of curves, \(\mathcal{M}_g\), and its pointed variants, \(\mathcal{M}_{g,n}\). Here a smooth curve degenerates to a nodal curve, and the crucial point is that any punctured 1-parameter family of smooth curves admits a unique nodal limit (this translates to the valuative criterion for properness of the moduli space). Despite its ubiquity, there have been a few other instance of modular compactifications of the moduli space of curves; most notably, perhaps, is that all the known steps in the log minimal model program for \(M_g\) appear to be modular, and this behavior is expected to persist: one believes (hopes?) that as one approaches closer and closer to the canonical model, by way of log canonical models, that one is re-compactifying the moduli space in a way that corresponds to replacing the ``boundary'' nodal curves in Deligne-Mumford's compactification with curves of increasingly severe singularities. Well, this isn't quite true as geometry in the interior of the moduli space may change as well (for instance, the hyper-elliptic locus should be flipped at a certain point), but the key is that the surgery performed downstairs on the coarse moduli space seems to naturally lift to the stack-theoretic level in a geometrically natural way. Before venturing down the path of the minimal model program in its entirely, one can quite reasonably restrict attention to compact birational models of this form which preserve the interior. The present paper, therefore, initiates a marvelous investigation into the moduli space of curves: to classify all modular compactifications. In a more technical sense, what this means is the following. The stack of all (pointed) curves has an irreducible component corresponding to smoothable curves (this is the component containing \(\mathcal{M}_{g,n}\)), and Smyth defines a modular compactification to be an open substack of this that is proper over Spec \(\mathbb{Z}\). His main theorem, in a very rough sense, is that there is a natural combinatorial construction that yields a significant fraction of all modular compactifications, and which in genus zero yields all of them. The ``missing'' compactifications that begin to appear in genus one are pursued in the author's follow-up papers in Compositio. The basic idea of the combinatorial construction is, for each DM-stable curve, to choose a collection of irreducible components that in essence are closed under specialization and satisfy a couple other natural axioms. One can contract these components down to points that are singularities of the same type (genus and multiplicity) as the sub-curves being contracted. This can be stated entirely in terms of dual graphs, and hence is combinatorially described and in particular yields a finite number of possibilities, for fixed \(g\) and \(n\). A natural example of this construction is the Schubert moduli space of pseudo-stable curves, where one contract elliptic tails down to cusps. Another example is to contract all unmarked components of a DM-stable curves; the author says this has not appeared in the literature previously, but at least in genus zero it has in Boggi's thesis work providing an algebraic construction of a topological construction Kontsevich suggested in all genera. Let us consider for a moment the special case of genus zero (here the singularities that can arise are locally the union of coordinate axes, and as mentioned above Smyth's result says that all modular compactifications arise from the above combinatorial setup). A corollary is that every modular compactification receives a regular morphism from \(\mathcal{M}_{0,n}\). We expect there to be many flips of this moduli space, yet a flip by definition is a small modification that does not receive a morphism, so in every single case the flip of \(\mathcal{M}_{0,n}\) must, by Smyth's result, be a birational model that is not modular in his sense of the word. This raises a fascinating question of what these flipped spaces could possibly parameterize, if they are indeed moduli spaces. In summary, this paper introduces a remarkably new and powerful perspective into the study of moduli spaces of curves: to define, and attempt to classify, all modular compactifications. On the one hand, this seminal approach may one day be attempted for moduli spaces of higher-dimensional objects, and it begins to illuminate another facet of the general story of stability conditions defining moduli spaces that we have witnessed in Mumford's GIT (more appropriately, the variational aspects of GIT) and Bridgeland's work in derived categories (although a theory of wall-crossing is not present here, though I strongly suspect it is lurking in the background). On the other hand, this approach immediately yields several fascinating results about the moduli space of curves itself, so the reader need not wait in suspense for subsequent vast generalizations of the ideas contained here (rich as they likely will be): this is a paper about the moduli space of curves, and an influential and strikingly original one at that.