arXiv subjects

Dmitry Babichev

Publications and source records attributed to Dmitry Babichev.

5 recordsLinked to original sources

arXiv2026

Counting Lattice Rectangles in $O(n\log n)$ Operations

Let $F(n)$ be the number of rectangles, not necessarily axis-parallel, whose vertices belong to the $n\times n$ square grid of lattice points. We give an exact algorithm that computes one prescribed value $F(n)$ in $O(n\log n)$ arithmetic operations and $O(n^{3/4})$ arithmetic words of working memory. The algorithm decomposes the count into M\"obius divisor layers, partitions weighted floor-moment queries by a truncated Euclidean coefficient-cone recursion, and reuses uniform marker grids along common coefficient paths. Each marker requires only its uniform cell and constant-size corrections at nearby boundaries, which select an exact precompiled cell operator. All integer operands have $O(\log n)$ bits. An exact 128-bit C++ implementation for the reported input range is compared experimentally with the previous $O(n\log^2 n)$ algorithm.

cs.CG↗

arXiv2026

Computing All Lattice-Rectangle Counts by Rational Staircase Sums

Let $F(n)$ be the number of rectangles, not necessarily axis-parallel, whose vertices belong to the $n\times n$ square grid of lattice points. We compute the complete table $F(1),\ldots,F(N)$ exactly in $O(M(N)\log N)$ coefficient-ring operations and $O(N\log N)$ ring elements of working memory, where $M(N)$ is a regular bound for multiplying degree-$N$ polynomials. The ring-level statement assumes that $6$ is invertible; over $\mathbb Z$ the only division is instead performed exactly in the elementary boundary term. With quasi-linear polynomial multiplication the arithmetic bound is $O(N\log^2 N)$. The algorithm applies a square-root cover before coefficient extraction and evaluates the resulting rational wedge and triangular sums by a local-denominator divide-and-conquer recursion. Primitive directions are recovered coefficientwise by M\"obius inversion, followed by five prefix sums. A modular number-theoretic-transform (NTT) implementation with certified Chinese-remainder (CRT) recovery is evaluated experimentally against the $O(N^{3/2})$ all-values algorithm.

cs.CG↗

arXiv2026

Counting All Lattice Rectangles in the Square Grid in Near-Linear Time

We study the exact counting problem for all lattice rectangles contained in the square $[0,n)\times[0,n)$, including non-axis-parallel ones. Starting from the standard parametrization by a primitive direction $(u,v)$ and two side lengths, we derive several exact algorithms: the classical $O(n^2)$ sweep, decompositions of complexity $O(n^{3/2}\log n)$ and $O(n^{4/3}\log n)$, a ten-moment weighted-floor-sum reduction of complexity $O(n\log^3 n)$, and a divisor-layer algorithm with the complexity $O(n\log^2 n)$. We also give an all-values algorithm that computes $F(1),\ldots,F(N)$ in $O(N^{3/2})$ arithmetic operations. The main idea behind the near-linear one-value algorithms is to reduce the geometric summation to constant-size families of weighted floor sums closed under Euclidean-style affine and reciprocal transformations. Besides the exact algorithmic results, we derive a two-term asymptotic expansion, $F(n)=\frac{4\log 2-1}{\pi^2}n^4\log n+B\,n^4+o(n^4)$ with the explicit formula for $B$, which provides an independent consistency check for the large-$n$ numerical data produced by the algorithms.

cs.CG↗

arXiv2019

Efficient Primal-Dual Algorithms for Large-Scale Multiclass Classification

We develop efficient algorithms to train $\ell_1$-regularized linear classifiers with large dimensionality $d$ of the feature space, number of classes $k$, and sample size $n$. Our focus is on a special class of losses that includes, in particular, the multiclass hinge and logistic losses. Our approach combines several ideas: (i) passing to the equivalent saddle-point problem with a quasi-bilinear objective; (ii) applying stochastic mirror descent with a proper choice of geometry which guarantees a favorable accuracy bound; (iii) devising non-uniform sampling schemes to approximate the matrix products. In particular, for the multiclass hinge loss we propose a \textit{sublinear} algorithm with iterations performed in $O(d+n+k)$ arithmetic operations.

stat.ML↗

arXiv2018

Constant Step Size Stochastic Gradient Descent for Probabilistic Modeling

Stochastic gradient methods enable learning probabilistic models from large amounts of data. While large step-sizes (learning rates) have shown to be best for least-squares (e.g., Gaussian noise) once combined with parameter averaging, these are not leading to convergent algorithms in general. In this paper, we consider generalized linear models, that is, conditional models based on exponential families. We propose averaging moment parameters instead of natural parameters for constant-step-size stochastic gradient descent. For finite-dimensional models, we show that this can sometimes (and surprisingly) lead to better predictions than the best linear model. For infinite-dimensional models, we show that it always converges to optimal predictions, while averaging natural parameters never does. We illustrate our findings with simulations on synthetic data and classical benchmarks with many observations.

stat.ML↗