Isolated hypersurface singularities may be stably degenerate
Why this paper
A fifty-year-old question of Arnold closed in the unexpected direction, by importing the birational geometry of moduli spaces into local singularity theory — the singularity remembers an abelian variety. The kind of cross-field plenary worth attending whatever one's own session is.
The question (Arnold, 1975)
Newton non-degenerate singularities are the good ones: their invariants (Milnor number, spectrum, …) can be read combinatorially off the Newton polyhedron, à la Kouchnirenko — but the property depends strongly on coordinates. Arnold asked whether the good form is always reachable up to stabilisation: is every polynomial with an isolated singularity at $0$ stably equivalent — after adding a nondegenerate quadratic form $Q(w_1,\dots,w_k)$ in fresh variables and a formal change of coordinates — to a Newton non-degenerate one? Without stabilisation the answer was known to be no (e.g. $f(x,y) = x^2y^2(x+y)^2 + x^7 + y^7$ is degenerate in every coordinate system); whether new variables always rescue the situation stayed open for fifty years.
The result
No: stably degenerate isolated hypersurface singularities exist — singularities that remain Newton-degenerate in every coordinate system after every stabilisation. The combinatorial normal form is genuinely unreachable, not merely hard to find. Combined with a result from Stevens' survey of the problem, some cubic polynomial (in perhaps very many variables) is stably degenerate. The argument is an existence proof; the authors note, almost apologetically, that finding a concrete example algorithmically would be wildly impractical.
The example's shape
Take a very general principally polarised abelian variety $A$ of dimension $\ge 16$, realize it birationally as a hypersurface $\{F = 0\}$ in projective space ($F$ homogeneous of degree $d$; every variety is birational to a hypersurface), and set $f = F + \sum_j z_j^D$ with $D > d$ to isolate the singularity at the origin. "Very general" does real work: $A$ is chosen so that everything isogenous to its dual avoids all rational subvarieties of the moduli space, avoids a countable bad set, and has Néron–Severi rank one.
The proof in five moves
(1) The $d$-contact locus of a singularity — the Ein–Lazarsfeld–Mustață space of jets $u(t)$ with $u(0)=0$ and $g(u(t)) = at^d$ mod $t^{d+1}$ — is an invariant of formal equivalence and carries a mixed Hodge structure. (2) For this $f$, a weight-graded piece of that MHS is the $H^1$ of a polarised abelian variety isogenous to $A$: the singularity remembers the abelian variety. (3) If some stabilisation of $f$ were Newton non-degenerate after a coordinate change, $f$ would sit in the family of all Newton non-degenerate polynomials with the same Newton polyhedron and bounded degree — a rational parameter variety $T$. (4) A simultaneous toric log resolution over $T$ makes the contact-locus cohomology locally constant, so the graded MHS pieces trace a map from $T$ to a moduli space of polarised abelian varieties. (5) For dimension $\ge 16$ that moduli space is of general type (Mumford; Tai), so the rational $T$ cannot map to it non-constantly through a very general point; the map is constant, hence hits only countably many abelian varieties — contradicting the choice of $A$.