Computing Periodic Billiard Orbits in L p L^{p} Balls via Newton’s Method and Smale’s α \alpha -Criterion

Igor Rivin · 2025

Read Computing Periodic Billiard Orbits in L p L^{p} Balls via Newton’s Method and Smale’s α \alpha -Criterion by Igor Rivin (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-22251.txt.

Introduction

Igor Rivin computes and certifies bouncing-ball orbits in curved balls in a 2025 paper. The arena is an Lp ball, a rounded shape set by a power-law norm, for powers above two. A billiard ball inside travels straight and reflects off the wall, and periodic paths that repeat are the quarry.

Newton's method, fed by a variational setup, hunts the paths. Smale's alpha-criterion then certifies each find as a true orbit, not a numerical mirage. Batched runs in the JAX framework swept many powers and bounce counts. Patterns emerged in the orbits' signatures and rotation numbers, tied to parity and primality of the bounce count. One large run certified nearly nine thousand distinct orbits with five bounces.

Catalog record

CREATORIgor Rivin
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE12 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-22251

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Computing Periodic Billiard Orbits in L p L^{p} Balls via Newton’s Method and Smale’s α \alpha -Criterion Thanks: with much help from Claude Code and ChatGPT Igor Rivin Abstract . We present a computational method for finding and verifying periodic billiard orbits in L p L^{p} balls ( p > 2 p>2 ) using Newton’s method applied to a variational formulation. The orbits are verified with Smale’s alpha-criterion, which provides a rigorous certificate of existence. We implement efficient batched computations in JAX and present systematic results for various p p and bounce counts N N . Our experiments reveal striking patterns in the critical-point structure, including a predominance of specific Morse signatures and rotation numbers that depend on the parity and primality of N N . Notably, our method routinely finds many more than the two periodic orbits per rotation number guaranteed by Birkhoff’s theorem—a large-scale run with five bounces in the L 3 L^{3} ball produced 8,927 distinct certified orbits from 30,000 random seeds, uncovering power-law growth and intricate clustering visualised with UMAP. 1991 Mathematics Subject Classification 37D50, 37M21, 65P10, 70H12, 49M15 1. Introduction The study of billiard dynamics in convex domains has a rich history connecting dynamical systems, symplectic geometry, and variational methods [ Bir27 , Tab05 ] . For the L^{p} ball defined by \mathcal{B}_{p}=\{(x,y)\in\mathbb{R}^{2}:|x|^{p}+|y|^{p}\leq 1\}, periodic billiard orbits correspond to critical points of the perimeter functional on the space of inscribed polygons. This paper presents: (1) A robust computational method using Newton’s method with careful parametrization (2) Rigorous verification via Smale’s \alpha -criterion (3) Systematic computational results revealing patterns in orbit signatures and rotation numbers (4) Open-source implementations for reproducibility (5) A breakthrough in computational billiards : While Birkhoff’s theorem guarantees at least two periodic orbits for each rational rotation number p/q , our method routinely finds orders of magnitude more orbits, revealing a rich landscape of critical points of the perimeter functional (6) Large-scale exploration and visualization : Using UMAP dimensionality reduction on thousands of orbits reveals hierarchical clustering by Morse signature, rotation number, and continuous parameter families To our knowledge, this represents the first systematic computational exploration of periodic billiard orbits in L^{p} balls that goes significantly beyond the minimal existence guaranteed by classical theorems. In our largest experiment with N=5 bounces in the L^{3} ball, we discovered and rigorously verified 8,927 distinct periodic orbits from 30,000 random seeds, exhibiting power-law growth with exponent \approx 0.714 and rich clustering structure when visualized using dimensionality reduction techniques. 2. Mathematical Formulation 2.1. Parametrization of the Boundary We parametrize the boundary of \mathcal{B}_{p} using the angular parameter t\in[0,1) : \gamma(t)=\left(\text{sgn}(\cos(2\pi t))|\cos(2\pi t)|^{2/p},\text{sgn}(\sin(2\pi t))|\sin(2\pi t)|^{2/p}\right) To avoid numerical issues at the singular points (where the boundary meets the axes), we add a small regularization \epsilon=10^{-14} to the absolute values before taking powers. 2.2. Variational Formulation A periodic billiard orbit with N bounces corresponds to a critical point of the perimeter functional: L(\theta_{1},\ldots,\theta_{N})=\sum_{i=1}^{N}\|\gamma(\theta_{i})-\gamma(\theta_{i+1})\| where indices are taken modulo N . The critical point equation is \nabla L=0 , which we solve using Newton’s method: \theta^{(k+1)}=\theta^{(k)}-H^{-1}\nabla L where H is the Hessian of L . 2.3. Smale’s \alpha -Criterion To verify that a numerical critical point corresponds to a genuine periodic orbit, we use Smale’s \alpha -criterion. Define: (1) \displaystyle\beta \displaystyle=\|H^{-1}\nabla L\| (2) \displaystyle\gamma \displaystyle=\frac{1}{2}\|H^{-1}\|\|H\| (3) \displaystyle\alpha \displaystyle=\beta\cdot\gamma Theorem 1 (Smale [ Sma80 ] ) . If \alpha<\sqrt{3}-1\approx 0.732 , then Newton’s method converges quadratically to a unique critical point near the current iterate. In practice, we use a more conservative threshold \alpha<0.15767 based on numerical experience. 3. Implementation Details 3.1. Canonical Representatives Due to symmetries, many parameter vectors \theta represent the same geometric orbit. We implement a canonicalization procedure: (1) Apply modulo 1 to all parameters (2) Cyclically permute to put the smallest parameter first (3) Choose between forward and reflected orientation 3.2. Coalescing Near-Duplicates After finding many candidate orbits, we coalesce near-duplicates: if \|\theta_{1}-\theta_{2}\|<\alpha_{1}+\alpha_{2} , we keep only the one with smaller \alpha . 3.3. Computational Signatures For each orbit, we compute: • Morse signature (n_{+},n_{-},n_{0}) : eigenvalue counts of the Hessian • Rotation number r/s : topological winding of the orbit • Perimeter : total length of the orbit 4. Experimental Results 4.1. Overview of Patterns Our experiments reveal several striking patterns: N Signatures Found Rotation Numbers Special Properties 2 (0,2,0) , (1,1,0) 0/1 , 1/2 Oscillatory orbits possible 3 (0,3,0) only 1/3 only All orbits are triangular 4 (0,4,0) , (1,3,0) 0/1 , 1/4 D_{4} symmetry effects 5 (0,5,0) , (1,4,0) , (2,3,0) 1/5 , 2/5 , 3/5 , 4/5 Star patterns dominate 7 (0,7,0) , (1,6,0) , (2,5,0) Various k/7 Rich rotation diversity Table 1 . Summary of orbit types for p=3 with various N 4.2. Key Observations Proposition 2 . For prime N , orbits can have various rotation numbers k/N where \gcd(k,N)=1 . Star patterns (with k>1 ) often dominate. Proposition 3 . For even N , oscillatory orbits with rotation number 0/1 are possible. Remark 4 . Interestingly, saddle points (signatures with n_{+}>0 ) often have longer perimeters than local maxima, contrary to naive expectation. 4.3. Large-Scale Exploration To understand the true richness of the orbit landscape, we conducted a large-scale exploration with N=5 bounces in the L^{3} ball, using 30,000 random seeds. 4.3.1. Discovery Statistics Signature Count Percentage Rotation Numbers Perimeter Range (0,5,0) 7,378 82.6% 1/5,2/5,3/5,4/5 [6.43,10.27] (1,4,0) 1,486 16.7% 2/5,3/5 [10.15,10.25] (2,3,0) 63 0.7% 2/5,3/5 [10.13,10.13] Table 2 . Statistics for 8,927 certified orbits with p=3 , N=5 from 30,000 random seeds. The orbit count follows a power law: N_{\text{orbits}}\propto N_{\text{seeds}}^{0.714} , suggesting we are far from saturating the orbit landscape. The discovery rate remains substantial even at 30,000 seeds (Figure 1 ). Figure 1 . Orbit discovery analysis showing: (a) cumulative unique orbits found, (b) discovery rate per batch, (c) power-law scaling on log-log plot, and (d) projected saturation curves. The power-law exponent of 0.714 indicates we are discovering fundamentally new orbit families, not just variations. 4.3.2. Rotation Number Distribution Within the dominant (0,5,0) signature: • 2/5 rotation: 3,149 orbits (42.7%) • 3/5 rotation: 3,135 orbits (42.5%) • 1/5 rotation: 551 orbits (7.5%) • 4/5 rotation: 543 orbits (7.4%) The star patterns ( 2/5 and 3/5 ) strongly dominate, while simple polygons ( 1/5 ) and their duals ( 4/5 ) are much rarer. 4.4. UMAP Visualization and Clustering To understand the structure of the orbit landscape, we applied UMAP (Uniform Manifold Approximation and Projection) [ MHSG18 ] to the 8,927 orbits, using their parameter vectors as features. Figure 2 . UMAP visualization of 8,927 periodic orbits colored by: (a) Morse signature, (b) rotation number, and (c) perimeter. The projection reveals hierarchical clustering: orbits first separate by signature, then by rotation number within each signature, with continuous variation in perimeter creating gradients within clusters. Key insights from the UMAP analysis (Figure 2 ): (1) Hierarchical structure : The primary clusters correspond to Morse signatures, with sub-clusters for rotation numbers. (2) Continuous families : Within each signature-rotation combination, orbits form continuous manifolds in parameter space. (3) Perimeter gradients : The perimeter varies smoothly within clusters, suggesting continuous deformation families. (4) Signature isolation : Different signatures occupy distinct regions of the reduced space, indicating fundamentally different orbit geometries. 4.5. Comparison with

Read the complete text