arXiv Science⌕ Search

arXiv subjects

Alexei Oblomkov

Publications and source records attributed to Alexei Oblomkov.

At least 19 recordsLinked to original sources

Honesty Is a Race: Predicting When Verifiable-Reward RL Teaches a Model to Bluff

Two rewards with the same expected payoff at the start of training can end a reinforcement-learning run in opposite places, with an honest model or one that bluffs. We study rewards that pay $b$ for a verdict that a checker cannot verify ("this program is correct") and $r$ for a certificate that it accepts (an input revealing a bug). To first order, a GRPO step makes the model more honest exactly when $\tilde{p}\,r/b>1+κ$. Here $\tilde{p}$ is the rate at which the checker accepts the model's certificates, weighted toward the items on which the model still gives both verdicts, and $κ$ is the pull of the items on which the unverifiable verdict is right. Both terms move during training. A checker raises $\tilde{p}$, so near the threshold honesty is a race, which a run just below the threshold can win and a run under a lottery (random pay with the same mean) cannot. The term $κ$ starts large and falls as the other items settle, so a run just above the threshold first dips toward bluffing, on code nearly to zero, before climbing back. Paying the honest verdict also pulls the other items toward it, which predicts a ceiling on the payoff ratio, above which the model gives the honest verdict where it is wrong. On propositional Datalog, two code models, and abstention on grade-school math, with $p$ the model's success rate before RL, no run above $pr/b=1$ ends collapsed, and runs under the checker can survive slightly below it. At $pr/b=3$, above the ceiling, both code models call about half of correct programs buggy. On depth-3 Datalog, training generously under the checker and then cutting the payoff keeps the model at least 67% honest in three seeds, at a ratio where the same model trained from the start falls to 4-15% and is at most 47% honest after 160 steps. On depth-2 Datalog, when honesty collapses, a linear probe fit before RL still reads the truth.

cs.LG↗

Sheaves on nodal singularities I

We study two families of moduli spaces of sheaves attached to the node, namely the punctual Hilbert schemes of the non-reduced plane curve $x^uy^v=0$, and the punctual Quot schemes of the module $k[[x]]^u\oplus k[[y]]^v$ over the reduced node $xy=0$. We compute their motivic generating functions for all $u,v$ and over every field. Both are specializations of a single rational function whose numerator counts matrices by rank, and they are related by the substitution $\mathbb{L}\mapsto Q\mathbb{L}$. This proves the Hilb--Quot correspondence for the node with independent ranks on the two branches. The main geometric ingredient is an elementary modification of an ideal along a branch of the curve, whose fibres are affine spaces of dimension equal to a contact length. We also construct stratifications of the Hilbert schemes into products of affine spaces and tori, and we formulate a branchwise Hilb--Quot--HOMFLY conjecture, which we test on the colored Hopf link.

math.AG↗

Asymptotic $q,t$-Fuss--Catalan numbers for type $B$

Let $W=W(B_n)$ act diagonally on $\mathfrak{h}\oplus\mathfrak{h}^*$, let $S=\mathbb{C}[\mathfrak{h}\oplus\mathfrak{h}^*]$, let $J\subset S$ be the ideal generated by the $W$-alternating polynomials and $\mathfrak{m}_S$ is the maximal ideal of the origin. For sufficiently large $m$ we compute $q,t$-Fuss-Catalan polynomial $Cat^{(m)}(B_n;q,t):=Hilb(\frac{J^m}{\mathfrak{m}_S J^m})_{det-part}$ and imply $Cat^{(m)}(B_n;1,1)=\binom{n(m+1)}{n}$. For proofs, we work with the $Γ$-equivariant Hilbert scheme $Y_n=nΓ$-$Hilb(\mathbb{C}^2)$, $Γ=μ_2$ and Haiman-type Koszul complex that defines the punctual locus of $Y_n$. Our formula for $Cat^{(m)}(B_n;q,t)$ is derived from a localization computaion for the Haiman-type Koszul complex.

math.CO↗

Quot scheme of points on torus knot singularities

For $\gcd(a,b)=1$, we show that the moduli space of $m$-codimensional $\Bbbk[\![T^a,T^b]\!]$-submodules of $\Bbbk[\![T]\!]^n$ is paved by affine cells, by proving that each Białynicki-Birula stratum of a closed related moduli space with respect to the natural $\mathbb{G}_m$-action is an affine bundle over the fixed point locus and that the fixed point locus is an iterated Grassmannian bundle. As an application, we determine the motive of this moduli space in the Grothendieck ring of varieties in terms of an explicit two-variable series $N_{a,b;n}(q,t)$, and use it to explicit compute the groupoid volume of the category of finite modules over $\mathbb{F}_q[\![T^a,T^b]\!]$. The series $N_{a,b;n}$ carries the conjectures we then formulate. At $n=\infty$ we conjecture a bi-infinite family of Rogers--Ramanujan type identities by specializing the $t$-variable; we identify their product side with the normalized character of a module over the $\mathcal{W}$-algebra minimal model $\mathcal{W}_a(a,a+b)$, and observe a connetion to colored Jones tails. At $n<\infty$ we conjecture that $N_{a,b;n}$ is computed by the bottom $α$-row of the trigraded $S^n$-colored HOMFLY homology of the torus knot $T(a,b)$, and that this same bottom row also computes the Quot schemes of finite codimensional $\Bbbk[\![T^a,T^b]\!]$-submoudles of $\Bbbk[\![T^a,T^b]\!]^n$ and the punctual Hilbert schemes of the non-reduced curve $(Y^a-X^b)^n=0$; the three quantities are special values at three points of the trigrading, and when $n=1$ they recover both the conjectures of Oblomkov--Rasmussen--Shende and of Kivinen--Trinh. Finally we conjecture that the one direction of the trigrading these three points do not see is a perverse filtration on the moduli spaces themselves, and we verify its prediction for a smooth germ at $n=2$ by computing the decomposition theorem for the $\mathrm{GL}_2$ spectral-curve family.

math.AG↗

Wall-crossing for Hilbert schemes

The goal of the minimal model program for the Hilbert scheme of points on a surface aims is to describe the (stable) base loci of all divisors, their associated birational models, and the maps between them. We answer all of these questions for the Hilbert scheme of points on the blowup of the affine plane at the origin. The birational models are Brill-Noether loci in a larger Hilbert scheme, and nested variants thereof, and the wall-crossing maps are described as explicit projections. We also establish several new facts about the homogeneous coordinate ring of this Hilbert scheme, including finding the minimal set of line bundles whose sections generate the ring.

math.AG↗

Affine Schubert calculus and double coinvariants

We define an action of the double coinvariant algebra $DR_n$ on the equivariant Borel-Moore homology of the affine flag variety $\widetilde{Fl}_n$ in type $A$, which has an explicit form in terms of the left and right action of the (extended) affine Weyl group and multiplication by Chern classes. Up to first order in the augmentation ideal, we show that it coincides with the action of the Cherednik algebra on the equivariant homology of the homogeneous affine Springer fiber $\widetilde{S}_{n,n+1} \subset \widetilde{Fl}_n$ due to Yun and the second author, and therefore preserves the non-equivariant Borel-Moore homology groups $H_*(\widetilde{S}_{n,n+1})\hookrightarrow H_*(\widetilde{Fl}_n)$. We then define a geometric filtration $F_{a} H_*(\widetilde{S}_{n,n+1})=H_*(\widetilde{S}(a))$ by closed subspaces $\widetilde{S}(a)\subset \widetilde{S}_{n,n+1}$, which we prove recovers the Garsia-Stanton descent order on $DR_n$. We use this to deduce an explicit monomial basis of $DR_n$, as well as an independent proof of the (non-compositional) Shuffle Theorem.

math.CO↗

A categorification of representations of $U_q(\mathfrak{gl}_{1|1})$

We categorify the action of $U_q(\mathfrak{gl}_{1|1})$ on the tensor product of its vector representations $(\mathbb{C}^{1|1})^{\otimes N}$. The generators $E$ and $F$ are represented by Fourier-Mukai functors between the derived categories of coherent sheaves on the total spaces of "semi-parabolic" vector bundles over the Grassmannians $Gr(k,N)$.

math.RT↗

Newton-Okounkov bodies for nested Hilbert schemes

We study sections of line bundles on the nested Hilbert scheme of points on the affine plane. We describe the spaces of sections in terms of certain ideals introduced by Haiman, and find explicit bases for them by analyzing the trailing terms in some monomial order. As a consequence, we compute the Newton-Okounkov bodies for nested Hilbert schemes.

math.AG↗

Symplectic involutions of $K3^{[n]}$ type and Kummer $n$ type manifolds

In this paper we describe the fixed locus of a symplectic involution on a hyperkähler manifold of type $K3^{[n]}$ or of Kummer $n$ type. We prove that the fixed locus consists of finitely many copies of Hilbert schemes of $K3$ surfaces of lower dimensions and isolated fixed points.

math.AG↗

Discriminants and motivic integration

We study invariants of a plane cuve singularity $(f,0)$ coming from motivic integration on symmetric powers of a formal deformation of $f$. We show that a natural discriminant integral recovers the motivic classes of the principal Hilbert schemes of points on $f$, while the orbifold integral gives the plethystic exponential of the motivic Igusa zeta function of $f$. The latter result also holds in higher dimemsions. Combined with results of Gorsky and Némethi we obtain an interpretation of the discriminant integrals in terms of knot Floer homology, which is reminiscent of the relation between the cohomology of contact loci and fixed point Floer homology proven by de la Bodega and Poza.

math.AG↗

Fixed loci of symplectic automorphisms of $K3^{[n]}$ and $n$-Kummer type manifolds

The aim of this paper is to give an explicit description of the fixed loci of symplectic automorphisms for certain hyperkahler manifolds, namely for Hilbert schemes on K3 surfaces and for generalized Kummer varieties. Here we extend our previous results from the case of involutions to more general groups. In particular, under some conditions on the dimension, we give the full answer for finite group actions of symplectic automorphisms coming from K3 surfaces. We prove that the all irreducible components of the fixed loci are of $K3^{[k]}$ type of lower dimensions or isolated points.

math.AG↗

The affine Springer fiber-sheaf correspondence

Given a semisimple element in the loop Lie algebra of a reductive group, we construct a quasi-coherent sheaf on a partial resolution of the trigonometric commuting variety of the Langlands dual group. The construction uses affine Springer theory and can be thought of as an incarnation of 3d mirror symmetry. For the group $GL_n$, the corresponding partial resolution is $\mathrm{Hilb}^n(\mathbb{C}^\times\times \mathbb{C})$. We also consider a quantization of this construction for homogeneous elements.

math.AG↗

Generic curves and non-coprime Catalans

We compute the Poincaré polynomials of the compactified Jacobians for plane curve singularities with Puiseaux exponents $(nd,md,md+1)$, and relate them to the combinatorics of $q,t$-Catalan numbers in the non-coprime case. We also confirm a conjecture of Cherednik and Danilenko for such curves.

math.AG↗

3D TQFT and HOMFLYPT homology

We describe a family of 3d topological B-models whose target spaces are Hilbert schemes of points in $\mathbb{C}^2$. The interfaces separating theories with different numbers of points correspond to braid strands. The Hilbert space of the picture of a closed braid is the HOMFLY-PT homology of the corresponding link.

math.GT↗

HOMFLYPT homology of Coxeter links

A Coxeter link is a closure of a product of two braids, one being a quasi-Coxeter element and the other being a product of partial full twists. This class of links includes torus knots \(T_{n,k}\) and torus links \(T_{n,nk}\). We identify the knot homology of a Coxeter link with the space of sections of a particular line bundle on a natural generalization of the punctual locus inside the flag Hilbert scheme of points in \(\mathbb{C}^2\).

math.AG↗

Categorical Chern character and braid groups

To a braid $β\in Br_n$ we associate a complex of sheaves $S_β$ on $Hilb_n(C^2)$ such that the previously defined triply graded link homology of the closure $L(β)$ is isomorphic to the homology of $S_β$. The construction of $S_β$ relies on the Chern functor $CH: MF_n^{st}\to D^{per}_{C^*\times C^*}(Hilb_n(C^2))$ defined in the paper together with its adjoint functor $HC$. We prove a formula for the closure of sufficiently positive elements of the Jucys-Murphy algebra previously conjectured by Gorsky, Negut and Rasmussen.

math.GT↗

Matrix factorizations and $gl(m|k)$-quantum invariants

In our previous papers we used the Hilbert scheme of points on $C^2$ in order to construct a triply graded link homology and its $gl(m)$ version. Here we extend the $gl(m)$ construction to super-algebras $gl(m|k)$.

math.GT↗

Soergel bimodules and matrix factorizations

We establish an isomorphism between the Khovanov-Rozansky triply graded link homology and the geometric triply graded homology due to the authors. Hence we provide an interpretation of the Khovanov-Rozansky homology of the closure of a braid $β$ as the space of derived sections of a $\mathbb{C}^*\times \mathbb{C}^*$- equivariant sheaf $Tr(β)$ on the Hilbert scheme $Hilb_n(\mathbb{C}^2)$, thus proving a version of Gorsky-Negut-Rasmussen conjecture \cite{GorskyNegutRasmussen16}. As a consequence we prove that Khovanov-Rozansky homology of knots satisfies the $q\to t/q$ symmetry conjectured by Dunfield-Gukov-Rasmussen \cite{DunfieldGukovRasmussen06}. We also apply our main result to compute the Khovanov-Rozansky homology of torus links.

math.GT↗