arXiv subjects

François Clément

Publications and source records attributed to François Clément.

At least 19 recordsLinked to original sources

arXiv2026

On the optimization of discrepancy measures

Numerical optimizations can find point sets with lower discrepancies than prior methods based on algebra, and $L_2$ discrepancies are computationally attractive for this. The $L_2$ star discrepancy unreasonably emphasizes the origin and Matoušek pointed to a serious drawback. He showed that the $L_2$ star discrepancy of $n$ points at $(1,\dots,1)\in[0,1]^d$ is less than the root mean squared $L_2$ star discrepancy of $n$ random points until $n>(3/2)^d-1$. We introduce the \emph{average squared discrepancy} (ASD) which averages $2^d$ versions of the squared $L_2$ star discrepancy, one for each vertex of $[0,1]^d$. We show that it can be computed in $O(dn^2)$ work and that it does not suffer from Matoušek's repeated-point pathology. We inspect several $L_2$ discrepancies that do not favor the origin over other vertices. Only two of them avoid the repeated-point pathology: a symmetric discrepancy and a periodic discrepancy. The symmetric discrepancy has a twin-point pathology where point sets with only $2$ distinct points can be better than random points until $n\ge 3^d$. We also study weighted $L_2$ discrepancies introduced by Hickernell that use some product weights. Weighting lower dimensional projections reduces but does not eliminate the repeated-point pathologies. A range of numerical experiments showcase recent progress in the construction of sets with low $L_2$ discrepancies for many different measures. In our examples, optimizing the ASD also gives good results for \emph{the other} discrepancies we consider, especially in lower dimensions. This version corrects some of the threshold sample sizes n at which random points become better than n repeated points for some discrepancy measures. In particular, the periodic L_2 discrepancy has no repeated point pathology.

math.NA↗

arXiv2026

Gradient descent with exponentially increasing stepsizes and restarts

Let $f:\mathbb{R}^d \rightarrow \mathbb{R}$. We consider gradient descent $x_{n+1} = x_n - τ_n \nabla f(x_n)$, where the stepsize $τ_n = τ\cdot e^{rn}$ is exponentially growing (with $τ> 0$ and $0 < r \ll 1$). This diverges for almost all initial values. We show that restarting the algorithm whenever $\|x_{n+1} - x_n\| \geq e^r\|x_n - x_{n-1}\|$ has good properties: it works very well in practice; we determine the limiting convergence rate in the case of convergence to a non-degenerate local minimum: it improves on classic gradient descent even though computational cost is comparable. The precise choice of $0 < r \ll 1$ does not matter much and the method is virtually independent of an initial stepsize $τ$ that is too small: while the convergence rate for gradient descent decays linearly as $τ\rightarrow 0$, it decays as $1/\log(1/τ)$ in this modified version; numerical examples illustrate the results.

math.OC↗

arXiv2026

An effective variant of the Hartigan $k$-means algorithm

The k-means problem is perhaps the classical clustering problem and often synonymous with Lloyd's algorithm (1957). It has become clear that Hartigan's algorithm (1975) gives better results in almost all cases, Telgarsky-Vattani note a typical improvement of $5\%$ -- $10\%$. We point out that a very minor variation of Hartigan's method leads to another $2\%$ -- $5\%$ improvement; the improvement tends to become larger when either dimension or $k$ increase.

cs.LG↗

arXiv2026

A Rocq Formalization of Simplicial Lagrange Finite Elements

Formalization of mathematics is a major topic, that includes in particular numerical analysis, towards proofs of scientific computing programs. The present study is about the finite element method, a popular method to numerically solve partial differential equations. In the long-term goal of proving its correctness, we focus here on the formal definition of what is a finite element. Mathematically, a finite element describes what happens in a cell of a mesh. It notably includes the geometry of the cell, the polynomial approximation space, and a finite set of linear forms that computationally characterizes the polynomials. Formally, we design a finite element as a record in the Rocq proof assistant with both values (such as the vertices of the cell) and proofs of validity (such as the dimension of the approximation space). The decisive validity proof is unisolvence, that makes the previous characterization unique. We then instantiate this record with the most popular and useful, the simplicial Lagrange finite elements for evenly distributed nodes, for any dimension and any polynomial degree, including the difficult unisolvence proof. These proofs require many results (definitions, lemmas, canonical structures) about finite families, affine spaces, multivariate polynomials, in the context of finite or infinite-dimensional spaces.

cs.LO↗

arXiv2025

Many critical points for discrete Riesz energy on $\mathbb{T}^2$

It is widely believed that the energy functional $E_p:(\mathbb{S}^2)^n \rightarrow \mathbb{R}$ $$ E_p = \sum_{i,j=1 \atop i \neq j}^{n} \frac{1}{\|x_i-x_j\|^p}$$ has a number of critical points, $\nabla E(x) = 0$, that grows exponentially in $n$. Despite having been extensively tested and being physically well motivated, no rigorous result in this direction exists. We prove a version of this result on the two-dimensional flat torus $\mathbb{T}^2$ and show that there are infinitely many $n \in \mathbb{N}$ such that the number of critical points of $E_p: (\mathbb{T}^2)^n \rightarrow \mathbb{R}$ is at least $\exp(c \sqrt{n})$ provided $p \geq 5 \log{n}$. We also investigate the special cases $n=3,4,5$ which turn out to be surprisingly interesting.

math.CA↗

arXiv2025

A Rocq Formalization of Monomial and Graded Orders

Even if binary relations and orders are a common formalization topic, we need to formalize specific orders (namely monomial and graded) in the process of formalizing in Rocq the finite element method. This article is therefore definitions, operators, and proofs of properties about relations and orders, thus providing a comprehensive Rocq library. We especially focus on monomial orders, that are total orders compatible with the monoid operation. More than its definition and proved properties, we define several of them, among them the lexicographic and grevlex orders. For the sake of genericity, we formalize the grading of an order, a high-level operator that transforms a binary relation into another one, and we prove that grading an order preserves many of its properties, such as the monomial order property. This leads us to the definition and properties of four different graded orders, with very factorized proofs. We therefore provide a comprehensive and user-friendly library in Rocq about orders, including monomial and graded orders, that contains more than 700 lemmas.

cs.LO↗

arXiv2025

Balanced Stick Breaking

Consider an infinite sequence $(x_k)_{k=1}^{\infty}$ on the unit circle $\mathbb{S}^1$. We may interpret the first $n$ elements $(x_k)_{k=1}^{n}$ as places where the `circular stick' $\mathbb{S}^1$ is broken into a total of $n+1$ pieces. It is clear that they cannot all be the same length all the time. de Bruijn and Erdős (1949) show that the ratio of the largest to the smallest has to be arbitrarily close to 2 infinitely many times which is sharp. They also consider the problem of balancing the length of $r$ consecutive intervals and prove $$ \frac{\max \mbox{length of}~r~\mbox{consecutive intervals}}{\min \mbox{length of}~r~\mbox{consecutive intervals}} \geq 1 + \frac{1}{r}.$$ We prove that this ratio can be as small as $1 + c \log{r}/ r$. This is done by means of refined discrepancy estimates for the van der Corput sequence over very short intervals and proves a conjecture of Brethouwer.

math.CO↗

arXiv2025

Low-Discrepancy Set Post-Processing via Gradient Descent

The construction of low-discrepancy sets, used for uniform sampling and numerical integration, has recently seen great improvements based on optimization and machine learning techniques. However, these methods are computationally expensive, often requiring days of computation or access to GPU clusters. We show that simple gradient descent-based techniques allow for comparable results when starting with a reasonably uniform point set. Not only is this method much more efficient and accessible, but it can be applied as post-processing to any low-discrepancy set generation method for a variety of standard discrepancy measures.

math.OC↗

arXiv2025

Optimizing Kernel Discrepancies via Subset Selection

Kernel discrepancies are a powerful tool for analyzing worst-case errors in quasi-Monte Carlo (QMC) methods. Building on recent advances in optimizing such discrepancy measures, we extend the subset selection problem to the setting of kernel discrepancies, selecting an m-element subset from a large population of size $n \gg m$. We introduce a novel subset selection algorithm applicable to general kernel discrepancies to efficiently generate low-discrepancy samples from both the uniform distribution on the unit hypercube, the traditional setting of classical QMC, and from more general distributions $F$ with known density functions by employing the kernel Stein discrepancy. We also explore the relationship between the classical $L_2$ star discrepancy and its $L_\infty$ counterpart.

stat.ML↗

arXiv2025

Robust Online Sampling from Possibly Moving Target Distributions

We suppose we are given a list of points $x_1, \dots, x_n \in \mathbb{R}$, a target probability measure $μ$ and are asked to add additional points $x_{n+1}, \dots, x_{n+m}$ so that $x_1, \dots, x_{n+m}$ is as close as possible to the distribution of $μ$; additionally, we want this to be true uniformly for all $m$. We propose a simple method that achieves this goal. It selects new points in regions where the existing set is lacking points and avoids regions that are already overly crowded. If we replace $μ$ by another measure $μ_2$ in the middle of the computation, the method dynamically adjusts and allows us to keep the original sampling points. $x_{n+1}$ can be computed in $\mathcal{O}(n)$ steps and we obtain state-of-the-art results. It appears to be an interesting dynamical system in its own right; we analyze a continuous mean-field version that reflects much of the same behavior.

math.OC↗

arXiv2025

carps: A Framework for Comparing N Hyperparameter Optimizers on M Benchmarks

Hyperparameter Optimization (HPO) is crucial to develop well-performing machine learning models. In order to ease prototyping and benchmarking of HPO methods, we propose carps, a benchmark framework for Comprehensive Automated Research Performance Studies allowing to evaluate N optimizers on M benchmark tasks. In this first release of carps, we focus on the four most important types of HPO task types: blackbox, multi-fidelity, multi-objective and multi-fidelity-multi-objective. With 3 336 tasks from 5 community benchmark collections and 28 variants of 9 optimizer families, we offer the biggest go-to library to date to evaluate and compare HPO methods. The carps framework relies on a purpose-built, lightweight interface, gluing together optimizers and benchmark tasks. It also features an analysis pipeline, facilitating the evaluation of optimizers on benchmarks. However, navigating a huge number of tasks while developing and comparing methods can be computationally infeasible. To address this, we obtain a subset of representative tasks by minimizing the star discrepancy of the subset, in the space spanned by the full set. As a result, we propose an initial subset of 10 to 30 diverse tasks for each task type, and include functionality to re-compute subsets as more benchmarks become available, enabling efficient evaluations. We also establish a first set of baseline results on these tasks as a measure for future comparisons. With carps (https://www.github.com/automl/CARP-S), we make an important step in the standardization of HPO evaluation.

cs.LG↗

arXiv2025

Regular Structures in Kronecker Permutations

Kronecker sequences $(k α\mod 1)_{k=1}^{\infty}$ for some irrational $α> 0$ have played an important role in many areas of mathematics. It is possible to associate to each finite segment $(k α\mod 1)_{k=1}^{n}$ a permutation $π\in S_n$ associated with the canonical lifting to two dimensions. We show that these permutations induced by Kronecker sequences based on irrational $α$ are extremely regular for specific choices of $n$ and $α$. In particular, all quadratic irrationals have an infinite number of choices of $n$ that lead to permutations where no cycle has length more than 4.

math.CO↗

arXiv2025

Monotone Paths on Acyclic 3-Regular Graphs

Motivated by trying to understand the behavior of the simplex method, Athanasiadis, De Loera and Zhang provided upper and lower bounds on the number of the monotone paths on 3-polytopes. For simple 3-polytopes with $2n$ vertices, they showed that the number of monotone paths is bounded above by $(1+φ)^n$, with $φ$ being the golden ratio. We improve the result and show that for a larger family of graphs the number is bounded above by $c \cdot 1.6779^n$ for some universal constant $c$. Meanwhile, the best known construction and conjectured extremizer is approximately $φ^n$.

math.CO↗

arXiv2025

Teaching Divisibility and Binomials with Coq

The goal of this contribution is to provide worksheets in Coq for students to learn about divisibility and binomials. These basic topics are a good case study as they are widely taught in the early academic years (or before in France). We present here our technical and pedagogical choices, the numerous exercises we developed and a small experiment we conducted on two students. As expected, it required additional Coq material such as other lemmas and dedicated tactics. The worksheets are freely available and flexible in several ways.

cs.LO↗

arXiv2025

On the largest singular vector of the Redheffer matrix

The Redheffer matrix $A_n \in \mathbb{R}^{n \times n}$ is defined by setting $A_{ij} = 1$ if $j=1$ or $i$ divides $j$ and 0 otherwise. One of its many interesting properties is that $\det(A_n) = O(n^{1/2 + \varepsilon})$ is equivalent to the Riemann hypothesis. The singular vector $v \in \mathbb{R}^n$ corresponding to the largest singular value carries a lot of information: $v_k$ is small if $k$ is prime and large if $k$ has many divisors. We prove that the vector $w$ whose $k-$th entry is the sum of the inverse divisors of $k$, $w_k = \sum_{d|k} 1/d$, is close to a singular vector in a precise quantitative sense.

math.NT↗

arXiv2025

Small gaps in the Ulam sequence

The Ulam sequence, described by Stanislaw Ulam in the 1960s, starts $1,2$ and then iteratively adds the smallest integer that can be uniquely written as the sum of two distinct earlier terms: this gives $1,2,3,4,6,8,11,\dots$. Already in 1972 the great French poet Raymond Queneau wrote that it `gives an impression of great irregularity'. This irregularity appears to have a lot of structure which has inspired a great deal of work; nonetheless, very little is rigorously proven. We improve the best upper bound on its growth and show that at least some small gaps have to exist: for some $c>0$ and all $n \in \mathbb{N}$ $$ \min_{1 \leq k \leq n} \frac{a_{k+1}}{a_k} \leq 1 + c\frac{\log{n}}{n}.$$

math.CO↗

arXiv2024

Failure of Orthogonality of Rounded Fourier Bases

The purpose of this note is to prove estimates for $$ \left| \sum_{k=1}^{n} \mbox{sign} \left( \cos \left( \frac{2πa}{n} k \right) \right) \mbox{sign} \left( \cos \left( \frac{2πb}{n} k \right) \right)\right|,$$ when $n$ is prime and $a,b \in \mathbb{N}$. We show that the expression can only be large if $a^{-1}b \in \mathbb{F}_n$ (or a small multiple thereof) is close to $1$. This explains some of the surprising line patterns in $A^T A$ when $A \in \mathbb{R}^{n \times n}$ is the signed discrete cosine transform. Similar results seem to exist at a great level of generality.

math.CA↗

arXiv2024

Approximately Jumping Towards the Origin

Given an initial point $x_0 \in \mathbb{R}^d$ and a sequence of vectors $v_1, v_2, \dots$ in $\mathbb{R}^d$, we define a greedy sequence by setting $x_{n} = x_{n-1} \pm v_n$ where the sign is chosen so as to minimize $\|x_n\|$. We prove that if the vectors $v_i$ are chosen uniformly at random from $\mathbb{S}^{d-1}$ then elements of the sequence are, on average, approximately at distance $\|x_n\| \sim \sqrt{πd/8}$ from the origin. We show that the sequence $(\|x_n\|)_{n=1}^{\infty}$ has an invariant measure $π_d$ depending only on $d$ and we determine its mean and study its decay for all $d$. We also investigate a completely deterministic example in $d=2$ where the $v_n$ are derived from the van der Corput sequence. Several additional examples are considered.

math.PR↗