Convergence of position based dynamics for first-order particle systems with volume exclusionSteffen PlunderKyoto UniversityICIAM 2023 TokyoMinisymposium on Evolution Equations for Interacting Species: Applications and Analysis
Particle systems with volume exclusion Feasible states: "swiss cheese set"With penalty terms: Going non-smooth: set of orthogonally outward pointing vectors
Hard spheres (second order dynamics) Overdamped hard spheres (first-order dynamics)
Comparing position based dynamics (PBD) and a penalty method Models: PBD is fast, stable and easy to implement! 🥳... there must be a catch?! 🤨
Motivation: Agent-based modelling in mathematical biology Typical requirements:Numerical implications:large number of cellscomputationally taxingvolume exclusionstiffness issuesvery uncertain parameterssimulation runtime impactsquality of parameter fitting
History of PBD(2006) Müller, Heidelberger, Hennix, Ratcliff Position Based Dynamics (2013) Macklin, Müller: Position Based Fluids (2015) Wang et al: Chebyshev Semi-Iterative Approach for Accelerating Projective and Position-based Dynamics (2016) Macklin, Müller: XPBD(2019) Macklin et al: Small Steps in Physics Simulation
PBD is used in movies, animation, games... But ignored by math community.Publications on PBD per domain
OutlineMathematical foundation: Differential inclusions on uniformly prox-regular setsPosition based dynamics: A rigorous method or just heuristic show-off?Convergence of PBD for first-order systemsApplication to particle systems with volume exclusion
Preliminaries from variational analysis Example
Well-posednesshas a unique, absolutely continuous solution. (2010) Frédéric Bernicot, Juliette Venel: Differential inclusions with proximal normal cones in Banach spaces, J. of Convex Analysis(2006) Jean Fenel Edmond, Lionel Thibault:BV solutions of nonconvex sweeping process differential inclusion with perturbation, J of Diff. Eqs.(2020) Bernard Brogliato, Aneel Tanwani:Dynamical Systems Coupled with Monotone Set-Valued Operators: Formalisms, Applications, Well-Posedness, and Stability, SIAM Review
Classical numerical methods for non-smooth dynamical systems Moreau's catch-up scheme: Typical approach for time-stepping schemes: Use Lagranian multipliersTaylor (+ index reduction) Solve for multipliers with iterative methods. Projected Nonlinear Gauss-Seidel (PNGS):
First-order position-based dynamics (PBD) Classical convergence proofs do not apply. Seemingly just a "termination of PNGS without error bounds".However, it is extremely easy to implement!
Large, accurate time-steps (PNGS) vs small, inaccurate time-steps (PBD)
Convergence resultThm. [SP, Sara Merino (2023); "soon on arxiv"]
Sketch of the proof (1/2)
Need notion of convergence for Answer: Scalary upper semicontinuity!Sketch of the proof (1/2): Consistency(2006) Jean Fenel Edmond, Lionel Thibault:BV solutions of nonconvex sweeping process differential inclusion with perturbation, J of Diff. Eqs.
Consistency in a scalarly upper semicontinuous senseDef. (SP, Sara Merino, 2023) Thm. (SP, Sara Merino): PBD is consistent in a susc sense!
Sketch of the proof (2/2)How can we avoid that projection errors accumulate?1. Metric calmness of intersection (subregularity): 2. Hypomonotonicity of normal cones for unif. prox-reg. sets:: "direction of projection cannot diverge too much"Main technical lemma. (SP, Sara Merino, 2023) cp. (2013) Robert Hesse, Russell D. Luke:Nonconvex Notions of Regularity and Convergence of Fundamental Algorithms for Feasibility ProblemsSIAM Journal on Optimization
(2006) Jean Fenel Edmond, Lionel Thibault:BV solutions of nonconvex sweeping process differential inclusion with perturbation, J of Diff. Eqs.By adapting the existence proof fromConsistency (in a susc sense) + stability = convergence.we get: Cor. (SP, Sara Merino, 2023).Hence, PBD convergeces for first-order dynamics!Wrapping things up...
Application to particle systems Thm. [SP, Sara Merino (2023)]cp. (2011) Bertrand Maury, Juliette Venel: A discrete contact model for crowd motion, ESAIMAnalysis of intersection of sets and sum of gradients kind of dual to each other: Hint: RHS is nicer (equal for volume exclusion)!computational cost
Suitable for complex models
New Julia package: github.com/SteffenPL/DifferentialInclusions.jlWork in progress!Supports:- Sparse constraints- Integration into SciMLJulia's ecosystemAlternative projects: https://github.com/siconos/siconos
Next steps?Order of convergence; stochastic differential inclusions; sweeping processes...For particles/macroscopic models:PBD particle method for PDE models with volume exclusion? How to fix PBD failure of convergence for second-order systems?PBD might be the only explict scheme for DAEs.Numerical analysis
Thanks for your attention!Sara Merino-AceitunoUniversity ViennaSeirin-Lee, BiMed-Math LabKyoto University, Institute for Advanced Study of Human Biology