Weighted, Multiphase, Volume-Preserving Mean Curvature Flow as Limit of the MBO Scheme on Manifolds

Fabius Krämer · 2025

Read Weighted, Multiphase, Volume-Preserving Mean Curvature Flow as Limit of the MBO Scheme on Manifolds by Fabius Krämer (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-2508-09064.txt.

Introduction

Fabius Krämer wrote a 2025 mathematics paper on a fast simulation method for moving interfaces. When several materials meet, their boundaries evolve to reduce total surface energy while keeping each region's volume fixed. A famous thresholding scheme simulates this motion cheaply: blur the picture, then cut it at a threshold, and repeat. It works on flat space; this paper extends the theory to curved, weighted spaces, the kind that arise in data science.

The main result proves the scheme's output converges to the true volume-preserving flow on a smooth closed surface. The proof is conditional: it assumes the scheme's energy approaches the right perimeter measure, a standard and natural step in this literature. The argument runs through an energy-dissipation inequality inherited from the flow's gradient structure. The work widens a practical algorithm's theoretical footing, from image processing roots toward geometry and data applications on manifolds.

Catalog record

CREATORFabius Krämer
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE130 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-09064

Browse the full library · About the Tower of Babel

Full text

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

Weighted, Multiphase, Volume-Preserving Mean Curvature Flow as Limit of the MBO Scheme on Manifolds Fabius Krämer Abstract . The famous thresholding scheme by Merriman, Bence, and Osher (Motion of multiple junctions: A level set approach. Journal of Computational Physics 112.2 (1994): 334-363.) proved itself as a very efficient time discretization of mean curvature flow. The present paper studies a multiphase, volume constrained version on a weighted manifold that naturally arises in the context of data science. The main result of this work proves convergence to multiphase, weighted, volume-preserving mean curvature flow on a smooth, closed manifold. The proof is only conditional in the sense that convergence of the approximating energy to the weighted perimeter is assumed. These type of assumptions are natural and common in the literature. The proof shows convergence in the energy dissipation inequality, induced by the underlying gradient flow structure, similar to the work of Laux and Otto (The thresholding scheme for mean curvature flow and De Giorgi’s ideas for minimizing movements. The Role of Metrics in the Theory of Partial Differential Equations 85 (2020): 63-94.). This leads to a new De Giorgi solution concept that describes weakly the limiting flow. Estimates regarding the derivatives of the heat kernel are developed in order to pass to the limit in the variations of the energy and metric. Keywords: Mean Curvature flow, gradient flows, heat kernel, clustering Mathematical Subject Classification (MSC2020) : 53E10, 53Z50, 49Q20 (Primary), 35D30, 58J35, 49Q05 (Secondary) 1. Introduction Mean curvature flow traditionally arose in material science, for example as flow describing the grain boundaries of metal in the annealing process [ 33 ] . But nowadays, its natural property to (locally) minimize the perimeter as gradient flow, motivates its use in data science tasks [ 6 , 14 , 30 , 31 ] . One prime example for this is the MBO scheme on graphs by Bertozzi et al. [ 31 ] ; it is a space and time discretization of mean curvature flow based on the famous MBO scheme [ 32 ] . The scheme is remarkable as, although computationally cheap, it inherits the gradient flow structure to iteratively minimize (locally), over all subsets \Omega of the data, the energy \sum_{x}\sum_{y}p(h,x,y)\mathbbold{1}_{\Omega}(x)\mathbbold{1}_{\Omega^{c}}(y). The heat kernel p(h,x,y) depends on the weights of the underlying graph and is an over all paths from node x to node y averaged weight. Therefore, a minimizing set \Omega of the energy minimizes the by the heat kernel given weights of edges that go from the set \Omega to its complement \Omega^{c} . Remember, the classical partition problem aims for minimizing the sum of weights from \Omega to its complement. Hence, the MBO scheme solves, locally and up to reweighing, the partition problem. Empirical studies [ 17 ] have shown that the scheme has even better accuracy in semi-supervised clustering tasks if one adds a volume constraint to the scheme. Often the number of data points in every set is (roughly) known in advance such that one can predescribe it in the algorithm. The aim of this paper is to show that this volume-constrained MBO scheme on graph is a space and time discretization of volume-preserving mean curvature flow. To that aim Laux and the author [ 18 ] have already proven that the scheme converges to its mean field limit if the number of data points goes to infinity. This result is obtained by using the \Gamma -convergence [ 19 ] of a suitable rescaling of the previously mentioned energy together with convergence of the Lagrange multipliers [ 18 ] of the volume constraints to obtain convergence of the iterates as minimizers in the minimizing movement interpretation of the MBO scheme [ 11 ] . The setting for the big-data limit is given by the “manifold assumption”; it states that the data points are sampled from a probability measure that is concentrated on a lower-dimensional manifold. The idea behind the assumption is that although the data may be embedded in a high dimensional space, points in the same class or cluster are similar and thus live in a lower dimensional structure. So far the mean field limit and its partition properties are not well understood. One is especially interested in its long-time behavior, as this describes approximately the properties of the result of the discrete clustering algorithm mentioned in the beginning. This paper is filling this gap by proving in the main Theorem 1 that the mean field limit of the multiphase volume constrained MBO scheme on the weighted data-manifold (see ( 1 ) below) conditionally convergences to multiphase, weighted, volume-preserving mean curvature flow on the manifold. The evolution law of the limiting flow on the interface of phase i is (formally) given by V_{i}=-H_{i}-\langle\mathbf{n}_{i},\nabla\log(\rho)\rangle+\Lambda_{i} where V_{i} is the (outer) normal velocity, H_{i} is the mean curvature, \mathbf{n}_{i} is the outer unit normal on the boundary of phase i , respectively, and \rho is the probability density on the manifold while \Lambda_{i} is the Lagrange multiplier ensuring the preservation of volume. The parameter for the convergence is the diffusion time h of the heat kernel (compare with the definition of the MBO scheme ( 1 )), which is actually the step-size for the time discretization of the gradient flow. The setting yields four major difficulties: The lack of the comparison principle as there are more than two phases, the curvature of the manifold, the inhomogeneity due to the weight, and the Lagrange multiplier arising from the volume constraint. There have been already several works regarding the convergence of the MBO scheme to mean curvature flow [ 3 , 8 , 12 , 16 , 20 , 21 , 22 , 23 , 24 , 26 , 38 ] . The only work on a weighted manifold so far is by Laux and Lelmi [ 21 ] where they use the powerful concept of viscosity solutions for mean curvature flow. Viscosity solutions are based on the comparison principle for mean curvature flow which is only valid for the case of two phases. The present paper is the first that proves the convergence of the scheme in the multiphase case on manifolds. Therefore, it introduces a new distributional solution concept for (weighted, multiphase,) volume-preserving mean curvature flow on a manifold. This weak solution idea is a natural extension of the ones introduced in [ 20 , 24 ] and based on De Giorgi’s energy dissipation inequality for gradient flows. That is, one encodes all information of the gradient flow equation \partial_{t}u=-\nabla E(u) in the inequality E(u(0))+\frac{1}{2}\int_{0}^{T}|\partial_{t}u(t)|^{2}+|\nabla E(u(t))|^{2}\,dt\leq E(u(T)) stating sharp energy dissipation. In our case, the energy reads as the weighted perimeter, i.e., E(u)=\int_{M}\rho|\nabla u| where M is the manifold, \rho is the weight and |\nabla u| is the total variation of the Gauss-Green measure of u . The L^{2} -gradient (on the boundary and weighted by \rho ) of the weighted perimeter used in Definition 1 is, at least formally (one can make this rigorous by looking at inner variations, see [ 29 , Chapter 17] ), \nabla E(\mathbbold{1}_{\Omega})=H+\langle\mathbf{n},\nabla\log(\rho)\rangle where H and \mathbf{n} are the mean curvature and normal of \Omega , respectively. Similar to the weight on the manifold, the work [ 10 ] regards an MBO scheme for an

Read the complete text