🔢 mathematics

Collective Optimization on Riemannian Manifolds with Bounded Curvature

This paper introduces an intrinsic consensus-based optimization framework for Riemannian manifolds with bounded curvature, proving the global well-posedness of its particle system and mean-field dynamics while demonstrating its effectiveness in finding global minimizers for nonconvex problems through numerical experiments on various manifolds.

Original authors: Hui Huang, Dohyun Kim, Hansol Park

Published 2026-06-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hui Huang, Dohyun Kim, Hansol Park

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find the absolute lowest point in a vast, foggy, and incredibly complex landscape. In the world of computers and data science, this is called global optimization. Usually, we try to solve this by sending out a swarm of tiny "robots" (particles) that wander around, talk to each other, and slowly move toward the lowest spot they can find.

This paper introduces a new, smarter way to guide these robots, specifically when the landscape isn't flat like a sheet of paper, but curved like the surface of a ball, a saddle, or even a spinning top.

Here is the breakdown of their discovery using simple analogies:

1. The Problem: Flat Maps vs. Curved Worlds

Most computer algorithms assume the world is flat (like a standard map of a city). They calculate distances by drawing straight lines. But in many real-world problems—like figuring out the orientation of a robot arm, analyzing 3D shapes, or handling complex data structures—the "ground" is actually curved.

The Old Way (Extrinsic): Imagine trying to walk on a globe, but you are forced to stay inside a giant glass box surrounding it. To move, you have to walk in straight lines inside the box, then get "projected" back onto the globe. This is clumsy. It distorts your path and wastes energy because you aren't respecting the curve of the Earth.
The New Way (Intrinsic): This paper proposes walking on the globe itself. You use the natural curves of the surface to move. You don't need the glass box; you just use the geometry of the sphere. This is faster, more accurate, and respects the true shape of the problem.

2. The Solution: A "Swarm" That Knows Geometry

The authors created a mathematical framework for a Consensus-Based Optimization (CBO) system. Think of this as a flock of birds trying to find the best nesting spot.

The Drift (The Pull): The birds look at where the "best" food is (the lowest energy state). In the old flat-world model, they would just average their positions. On a curved world, you can't just "add" positions together. Instead, the authors use Logarithmic Maps.
- Analogy: Imagine standing on a hill. To tell a friend where the valley is, you don't say "go 5 miles North." You say, "Walk along the path that leads down the steepest slope." The "Logarithmic Map" is the instruction that tells a particle exactly which path to take to get to a specific point on the curve.
The Diffusion (The Exploration): The birds also need to wander randomly to avoid getting stuck in a small, shallow dip (a local minimum) that looks like the bottom but isn't. The paper adds a "wandering" factor that gets stronger the further you are from the consensus, helping the swarm explore the whole landscape before settling down.

3. The Safety Net: The "Cut-Locus" and Cutoffs

Curved spaces have tricky spots. On a sphere, if you are at the North Pole, the "South Pole" is the same distance away in every direction. This creates a mathematical singularity (a point where the math breaks down).

The Fix: The authors put up "fences" (mathematical cutoffs). They ensure the robots only operate in a safe, well-behaved zone where the math works perfectly. If a robot gets too close to a confusing edge, the algorithm gently nudges it back or stops the wandering to prevent errors. This ensures the system never crashes or gets confused.

4. The Proof: It Actually Works

The paper doesn't just guess; it proves three big things:

It won't break: They proved that no matter how you start the swarm, the math guarantees the robots will keep moving and won't disappear or explode into chaos.
It finds the best spot: They proved that if you let the swarm run long enough, and if the "wandering" is tuned correctly, the entire group will eventually collapse onto the single, true lowest point of the landscape, ignoring all the fake dips along the way.
It works on different shapes: They tested this on three very different "worlds":
- The Sphere ( $S^2$ ): Like the Earth.
- Hyperbolic Space ( $H^2$ ): A saddle-shaped world that expands infinitely (like a Pringles chip that keeps getting bigger).
- The Rotation Group ($SO(3)$): The space of all possible 3D rotations (like a spinning top).

5. The Results: Robots That Learn to Dance

In their computer simulations, they watched the swarm of particles start scattered all over the place (maximum confusion).

On the Sphere: The swarm started as a messy cloud, then slowly tightened up, avoiding the fake valleys, and finally snapped into a tight cluster right on top of the true lowest point.
On the Saddle and the Spinning Top: The same thing happened. Even though the math for these shapes is much harder, the "intrinsic" method (walking on the curve) worked perfectly.

Summary

This paper is like giving a GPS system a new operating system. Instead of forcing a curved world into a flat map (which causes errors), it teaches the GPS to understand curves natively. They proved mathematically that this new system is stable, reliable, and will always find the true "bottom" of the hill, even if the hill is a sphere, a saddle, or a spinning wheel. They showed this works in theory and confirmed it with computer experiments.

Technical Summary: Collective Optimization on Riemannian Manifolds with Bounded Curvature

Problem Statement

Global optimization in high-dimensional spaces is a central challenge in scientific computing and machine learning. While Consensus-Based Optimization (CBO) has emerged as a robust, derivative-free paradigm in Euclidean spaces, existing extensions to manifolds largely rely on extrinsic formulations. These approaches embed the manifold into an ambient Euclidean space and utilize the ambient metric and projections.

The authors identify two primary limitations of extrinsic methods:

Geometric Distortion: Extrinsic dynamics depend on the specific embedding and ambient metric, which may not align with the intrinsic geodesic distance or natural diffusion properties of the manifold. This is particularly problematic for general Riemannian manifolds where no canonical ambient space exists.
Computational Inefficiency: Embedding an $n$ -dimensional manifold into a higher-dimensional Euclidean space (e.g., $SO(n)$ embedded in $\mathbb{R}^{n^2}$ ) artificially inflates the state space dimension, increasing computational costs.

The paper addresses the problem of developing an intrinsic CBO framework that operates directly on a Riemannian manifold $(M, g)$ using only geometric primitives (geodesic distance, exponential/logarithmic maps, and the Laplace–Beltrami operator), specifically for manifolds with bounded sectional curvature.

Methodology

1. Intrinsic Model Formulation

The authors propose a particle system and a corresponding mean-field limit defined entirely intrinsically.

Consensus Mechanism: Instead of Euclidean vector subtraction ( $x - x_\alpha$ ), the drift is constructed using the Riemannian logarithmic map ( $\log_x y$ ). The consensus direction $u_\alpha(\rho; x)$ at a base point $x$ is defined as a weighted average of logarithmic vectors:
$u_\alpha(\rho; x) \propto \int_M w_\alpha(y) \log_x y \, d\rho(y)$
where $w_\alpha(y) = \exp(-\alpha E(y))$ is the Gibbs weight.
Geometric Cutoffs: To handle structural difficulties such as the cut-locus (where the logarithmic map is not well-defined) and curvature-induced dispersion, the model employs smooth cutoff functions $h_{R, R+\delta}$ $h_{R, R + δ}$ .
- The drift term is localized to a geodesic ball $B_{R+\delta}(o)$ to ensure $\log_x y$ is single-valued and regular.
- The diffusion term is similarly localized to prevent unbounded growth of moments in negatively curved spaces (Cartan–Hadamard manifolds), where standard diffusion can inflate moments more strongly than in Euclidean space.
Dynamics:
- Particle System (SDE): $N$ particles evolve via a system of SDEs driven by the intrinsic drift and a state-dependent diffusion term involving a Riemannian Brownian motion.
- Mean-Field Limit: As $N \to \infty$ , the empirical measure converges to a law $\rho_t$ satisfying a nonlinear Fokker–Planck equation on the manifold.

2. Theoretical Framework

The analysis relies on the assumption that the manifold $M$ is complete, connected, and has bounded sectional curvature ( $-\kappa^- \le K \le \kappa^+$ ).

Well-posedness: The authors establish the global existence and uniqueness of solutions for both the particle system and the McKean–Vlasov process. This involves proving Lipschitz continuity of the drift and diffusion coefficients with respect to the Wasserstein distance and the base point, utilizing parallel transport to compare vectors in different tangent spaces.
Global Convergence: Under additional assumptions on the objective function $E$ (existence of a unique global minimizer and specific growth conditions near the minimizer), the paper proves that the mean-field distribution concentrates exponentially around the global minimizer as the weight parameter $\alpha \to \infty$ . The proof utilizes a dissipative estimate of a variance functional and a quantitative Laplace principle to bound the drift error.

Key Contributions

Intrinsic CBO Framework: The paper introduces the first CBO formulation that is fully intrinsic, avoiding ambient embeddings. It replaces Euclidean operations with Riemannian analogs (logarithmic averaging and parallel transport).
Rigorous Well-posedness: The authors provide a rigorous proof of global well-posedness for the proposed particle system and McKean–Vlasov dynamics on manifolds with bounded curvature, including explicit curvature-dependent estimates for the logarithmic map.
Global Convergence Proof: The paper establishes the global convergence of the mean-field equation toward a global minimizer. It demonstrates that for sufficiently large $\alpha$ , the law of the process concentrates around the minimizer with an exponential decay rate.
Handling Geometric Singularities: The work addresses the cut-locus problem and curvature effects by introducing localized geometric cutoffs, ensuring the model remains well-defined and tractable even on manifolds where global convexity does not hold.

Results

Theoretical Results

Theorem 3.1 & 3.2: Proved the existence and uniqueness of strong solutions for the particle system and the mean-field process, respectively, under natural regularity assumptions.
Theorem 4.1: Established the global convergence of the mean-field dynamics. Specifically, it showed that the variance functional $V[\rho_t]$ decays exponentially to an arbitrarily small neighborhood of the global minimizer, provided the attraction strength $\lambda$ dominates the diffusion-curvature contribution and $\alpha$ is sufficiently large.

Numerical Experiments

The authors validated the theory through numerical experiments on three distinct manifolds using a Riemannian Euler–Maruyama scheme:

Sphere ( $S^2$ ): A compact manifold with positive curvature. The algorithm successfully optimized a nonconvex "spherical Ackley" function, guiding particles from a uniform distribution to the global minimizer despite the presence of local traps and the cut-locus.
Hyperbolic Space ( $H^2$ ): A non-compact manifold with negative curvature. The algorithm demonstrated exponential convergence on an unbounded domain without the need for geometric cutoffs (as the cut-locus is empty), validating the theory for Cartan–Hadamard manifolds.
Special Orthogonal Group ($SO(3)$): A compact Lie group. The method successfully optimized a rotation-based Ackley function, demonstrating applicability to non-commutative matrix Lie groups.

In all cases, the empirical variance of the particle swarm exhibited a linear decay on a semi-logarithmic scale, consistent with the theoretical exponential convergence rates.

Significance and Claims

The paper claims to provide a rigorous PDE/SDE framework for analyzing consensus-based optimization beyond Euclidean space while preserving the intrinsic geometry of the manifold.

Motivation: The work is motivated by the limitations of extrinsic methods, which distort geometry and increase computational dimensionality.
Scope: The results are presented for manifolds with bounded sectional curvature. The authors note that while previous works studied intrinsic aggregation or self-organization, this is the first to formulate and analyze consensus-based optimization intrinsically.
Modesty: The authors acknowledge that the theoretical guarantees (well-posedness and convergence) rely on specific geometric cutoffs to handle the cut-locus and curvature effects. In numerical experiments, they relax these cutoffs to demonstrate robustness, but the theoretical proofs strictly adhere to the localized definitions. The paper does not claim to solve optimization on all possible manifolds without restriction, but rather establishes a foundational framework for those with bounded curvature.

The significance lies in bridging the gap between the practical success of CBO in Euclidean spaces and the geometric necessities of optimization on curved spaces, offering a mathematically sound alternative to embedding-based approaches.