arXiv subjects

Anurag Rao

Publications and source records attributed to Anurag Rao.

10 recordsLinked to original sources

arXiv2026

Zero-one laws for uniform approximation via Gaussian and Eisenstein integers

We establish two distinct zero-one laws for the uniform Diophantine approximation of complex numbers by quotients of Gaussian integers and by quotients of Eisenstein integers. Using tools from homogeneous dynamics, we study this problem by reducing to a shrinking target problem on certain homogeneous spaces of $\mathrm{SL}_2(\mathbb{C})$. The main novel ingredients include measure estimates on a certain family of neighborhoods of the corresponding critical loci, as well as new disjointness statements to control the short-range mixing contribution. Due to the different nature of the critical loci in the Gaussian and Eisenstein cases, these measure estimates are obtained by rather different arguments.

math.DS↗

arXiv2026

Constructing bounded orbits of special types on homogeneous spaces

Let $X = G/Γ$ be a quotient of a real Lie group by a non-uniform lattice. Consider a one-parameter subgroup $F$ of $G$ that is $\operatorname{Ad}$-diagonalizable over $\mathbb{C}$ and whose action on $(X,m_X)$ is mixing. In this dynamical system we study the set of points $x \in X$ with a precompact orbit, written as $E(F,\infty)$, which is known to be a dense subset of $X$ of full Hausdorff dimension. We prove that $E(F,\infty)$ is indecomposable in the following sense: given any $y \in E(F,\infty)$, the set of $x \in E(F,\infty)$ for which $y \in \overline{F_+x}$, where $F_+$ denotes the positive ray in $F$, is uncountable and dense in $E(F,\infty)$. When the dimension of the neutral subgroup of $G$ with respect to $F$ is $1$ we demonstrate, for any $\varepsilon>0$, the existence of many points $x \in X$ whose orbit closure $\overline{F_+x} \subset X$ is compact and has Hausdorff dimension at least $\dim X - \varepsilon$.

math.DS↗

arXiv2024

Model Editing at Scale leads to Gradual and Catastrophic Forgetting

Editing knowledge in large language models is an attractive capability to have which allows us to correct incorrectly learnt facts during pre-training, as well as update the model with an ever-growing list of new facts. While existing model editing techniques have shown promise, they are usually evaluated using metrics for reliability, specificity and generalization over one or few edits. We argue that for model editing to have practical utility, we must be able to make multiple edits to the same model. With this in mind, we evaluate the current model editing methods at scale, focusing on two state of the art methods: ROME and MEMIT. We find that as the model is edited sequentially with multiple facts, it continually forgets previously edited facts and the ability to perform downstream tasks. This forgetting happens in two phases -- an initial gradual but progressive forgetting phase followed by abrupt or catastrophic forgetting phase. Both gradual and catastrophic forgetting limit the usefulness of model editing methods at scale -- the former making model editing less effective as multiple edits are made to the model while the latter caps the scalability of such model editing methods. Our analysis also highlights other key limitations of ROME and MEMIT at scale. With our work, we push for the development and evaluation of model editing methods keeping scalability in mind.

cs.CL↗

arXiv2024

Badly approximable grids and k-divergent lattices

For an m by n real matrix A, we investigate the set of badly approximable targets for A as a subset of the m-torus. It is well known that this set is large in the sense that it is dense and has full Hausdorff dimension. We investigate the relationship between its measure and Diophantine properties of A. On the one hand, we give the first examples of a non-singular matrix A such that the set of badly approximable targets has full measure with respect to some non-trivial algebraic measure on the torus. For this, we use transference theorems due to Jarnik and Khintchine, and the parametric geometry of numbers in the sense of Roy. On the other hand, we give a novel Diophantine condition on A that slightly strengthens non-singularity, and show that under the assumption that A satisfies this condition, the set of badly approximable targets is a null-set with respect to any non-trivial algebraic measure on the torus. For this we use naive homogeneous dynamics, harmonic analysis, and a novel concept we refer to as mixing convergence of measures.

math.NT↗

arXiv2023

A dichotomy phenomenon for Bad minus normed Dirichlet

Given a norm $ν$ on $\mathbb{R}^2$, the set of $ν$-Dirichlet improvable numbers $\mathbf{DI}_ν$ was defined and studied in the papers of Andersen-Duke (Acta Arith. 2021) and Kleinbock-Rao (Internat. Math. Res. Notices 2022). When $ν$ is the supremum norm, $\mathbf{DI}_ν= \mathbf{BA}\cup \mathbb{Q}$, where $\mathbf{BA}$ is the set of badly approximable numbers. Each of the sets $\mathbf{DI}_ν$, like $\mathbf{BA}$, is of measure zero and satisfies the winning property of Schmidt. Hence for every norm $ν$, $\mathbf{BA} \cap \mathbf{DI}_ν$ is winning and thus has full Hausdorff dimension. In the present article we prove the following dichotomy phenomenon: either $\mathbf{BA} \subset \mathbf{DI}_ν$ or else $\mathbf{BA} \smallsetminus \mathbf{DI}_ν$ has full Hausdorff dimension. We give several examples for each of the two cases. The dichotomy is based on whether the critical locus of $ν$ intersects a precompact $g_t$-orbit, where $\{g_t\}$ is the one-parameter diagonal subgroup of $\operatorname{SL}_2(\mathbb{R})$ acting on the space $X$ of unimodular lattices in $\mathbb{R}^2$. Thus the aforementioned dichotomy follows from the following dynamical statement: for a lattice $Λ\in X$, either $g_\mathbb{R} Λ$ is unbounded (and then any precompact $g_{\mathbb{R}_{>0}}$-orbit must eventually avoid a neighborhood of $Λ$), or not, in which case the set of lattices in $X$ whose $g_{\mathbb{R}_{>0}}$-trajectories are precompact and contain $Λ$ in their closure has full Hausdorff dimension.

math.DS↗

arXiv2023

K-divergent lattices

We introduce a novel concept in topological dynamics, referred to as $k$-divergence, which extends the notion of divergent orbits. Motivated by questions in the theory of inhomogeneous Diophantine approximations, we investigate this notion in the dynamical system given by a certain flow on the space of unimodular lattices in $\mathbb{R}^d$. Our main result is the existence of $k$-divergent lattices for any $k\geq 0$. In fact, we utilize the emerging theory of parametric geometry of numbers and calculate the Hausdorff dimension of the set of $k$-divergent lattices.

math.DS↗

arXiv2022

Weighted uniform Diophantine approximation of systems of linear forms

Following the development of weighted asymptotic approximation properties of matrices, we introduce the analogous uniform approximation properties (that is, study the improvability of Dirichlet's Theorem). An added feature is the use of general norms, rather than the supremum norm, to quantify the approximation. In terms of homogeneous dynamics, the approximation properties of an $m \times n$ matrix are governed by a trajectory in $\mathrm{SL}_{m+n}({\mathbb R})/\mathrm{SL}_{m+n}({\mathbb Z})$ avoiding a compact subset of the space of lattices called the critical locus defined with respect to the corresponding norm. The trajectory is formed by the action of a one-parameter diagonal subgroup corresponding to the weights. We first state a very precise form of Dirichlet's theorem and prove it for some norms. Secondly we show, for these same norms, that the set of Dirichlet-improvable matrices has full Hausdorff dimension. Though the techniques used vary greatly depending on the chosen norm, we expect these results to hold in general.

math.NT↗

arXiv2021

Abundance of Dirichlet-improvable pairs with respect to arbitrary norms

In a recent paper of Akhunzhanov and Shatskov the two-dimensional Dirichlet spectrum with respect to Euclidean norm was defined. We consider an analogous definition for arbitrary norms on $\mathbb{R}^2$ and prove that, for each such norm, the set of Dirichlet improvable pairs contains the set of badly approximable pairs, hence is hyperplane absolute winning. To prove this we make a careful study of some classical results in the geometry of numbers due to Chalk--Rogers and Mahler to establish a Hajós--Minkowski type result for the critical locus of a cylinder. As a corollary, using a recent result of the first named author with Mirzadeh, we conclude that for any norm on $\mathbb{R}^2$ the top of the Dirichlet spectrum is not an isolated point.

math.NT↗

arXiv2021

Critical loci of convex domains in the plane

Let $K$ be a bounded convex domain in $\mathbb{R}^2$ symmetric about the origin. The critical locus of $K$ is defined to be the (non-empty compact) set of lattices $Λ$ in $\mathbb{R}^2$ of smallest possible covolume such that $Λ\cap K= \lbrace 0\rbrace$. These are classical objects in geometry of numbers; yet all previously known examples of critical loci were either finite sets or finite unions of closed curves. In this paper we give a new construction which, in particular, furnishes examples of domains having critical locus of arbitrary Hausdorff dimension between $0$ and $1$.

math.MG↗

arXiv2020

A zero-one law for uniform Diophantine approximation in Euclidean norm

We study a norm sensitive Diophantine approximation problem arising from the work of Davenport and Schmidt on the improvement of Dirichlet's theorem. Its supremum norm case was recently considered by the first-named author and Wadleigh, and here we extend the set-up by replacing the supremum norm with an arbitrary norm. This gives rise to a class of shrinking target problems for one-parameter diagonal flows on the space of lattices, with the targets being neighborhoods of the critical locus of a suitably scaled norm ball. We use methods from geometry of numbers and dynamics to generalize a result due to Andersen and Duke on measure zero and uncountability of the set of numbers for which Minkowski approximation theorem can be improved. The choice of the Euclidean norm on $\mathbb{R}^2$ corresponds to studying geodesics on a hyperbolic surface which visit a decreasing family of balls. An application of a dynamical Borel-Cantelli lemma of Maucourant produces a zero-one law for improvement of Dirichlet's theorem in Euclidean norm.

math.NT↗