Distribution Control via Optimal Transport

• Distribution control via OT is a framework that manipulates probability measures using cost-optimal transport maps and couplings.
• It employs scalable algorithms such as the Sinkhorn method, block-coordinate Frank–Wolfe, and ADMM-based approaches for robust and distributed control.
• Advanced extensions include game-theoretic, neural, and flow-based models that address adversarial challenges and high-dimensional data.

Distribution control via optimal transport (OT) refers to the systematic steering, alignment, or manipulation of probability distributions using the mathematical structures, algorithms, and theory of OT. OT provides a rigorous, geometry-aware framework for quantifying, mapping, and actively controlling the transformation of one probability measure into another in a manner that is cost-minimizing with respect to a prescribed ground metric or cost. This paradigm underpins a broad spectrum of scientific advances—including robust and distributed control, data-driven system identification, adversarial resilience, generative modeling, and scalable machine learning—by harnessing the fundamental variational and duality principles of OT for both descriptive and prescriptive purposes.

1. Core Mathematical Formulations of Distribution Control via OT

The OT paradigm embodies two principal frameworks for distribution control: the classical Monge and Kantorovich formulations, and their variational, dynamic, or regularized counterparts:

• Monge Formulation: Seeks a measurable map $T: \mathbb{R}^d \to \mathbb{R}^d$ such that $T_\#\mu=\nu$ and minimizes

$\int_{\mathbb{R}^d} c(x,T(x))\,d\mu(x)$

where $c$ is the cost function and $T_\#\mu$ denotes the pushforward measure.

• Kantorovich Formulation: Relaxes the problem to couplings $\pi$ between $\mu$ and $\nu$:

$\inf_{\pi\in\Pi(\mu,\nu)}\int c(x,y)\,d\pi(x,y)$

with $\Pi(\mu,\nu)$ the set of joint measures with marginals $T_\#\mu=\nu$0 (Peyré, 10 May 2025).

• Dynamic Formulation (Benamou–Brenier):

$T_\#\mu=\nu$1

subject to the continuity equation $T_\#\mu=\nu$2, $T_\#\mu=\nu$3, $T_\#\mu=\nu$4 (Peyré, 10 May 2025).

These formulations ground a hierarchy of algorithms and control mechanisms, extending to semi-relaxed, entropic, and regularized variants for tractability and robustness (Fukunaga et al., 2021).

2. Algorithmic and Distributed Methods for Distribution Control

OT-based distribution control is realized via scalable algorithms adapted to discrete, continuous, and networked agent settings:

• Sinkhorn Algorithm: Implements entropic regularization for large-scale transport, iteratively scaling dual potentials to enforce marginal constraints, and provides geometric convergence guarantees; central in sample-based and image-to-image translation contexts (Peyré, 10 May 2025).
• Block-Coordinate Frank–Wolfe (BCFW): Facilitates projection-free optimization for semi-relaxed OT. Each iteration updates a single marginal/column, leveraging curvature constants to achieve sublinear or linear convergence and sparse transport plans essential for real-time or large-scale distribution steering (Fukunaga et al., 2021).
• ADMM-based Distributed OT: For networked multi-agent systems and adversarially robust settings, agents solve target/source-local subproblems and communicate only over permitted graph neighborhoods, enabling scalable, decentralized computation of resilient transport strategies (Hughes et al., 2021, Krishnan et al., 2018).

Table: Key Algorithmic Approaches

Method	Setting	Convergence/Guarantees
Sinkhorn	Dense discrete/cont.	Geometric (entropic regular.), O($T_\#\mu=\nu$5)
BCFW	Semi-relaxed, sparse	$T_\#\mu=\nu$6, dual certs.
ADMM/distributed	Networked/discrete	Saddle-point/Nash, distributed convergence

3. Game-Theoretic and Adversarial Extensions

Recent work extends OT distribution control to adversarial and uncertain environments:

• Resilient OT with Deceptive Adversaries: Models the OT allocation as a two-person zero-sum game in which the adversary perturbs utilities at selected targets, subject to stealth and magnitude constraints. The equilibrium (saddle-point) is guaranteed by the minimax theorem and computed via distributed interleaving of attacker best responses and planner ADMM steps (Hughes et al., 2021).
• Distributionally Robust Optimization via OT (OT-DRO): Unifies divergence-based and Wasserstein-DRO by lifting to an augmented outcome–likelihood space with conditional moment constraints. Allows simultaneous adversarial perturbation of outcomes and likelihoods, yielding a worst-case coupling as a transport plan between nominal and perturbed distributions, and admits tractable dual and conic programming reformulations (Blanchet et al., 2023).

This game-theoretic paradigm encompasses strategic data poisoning, robust resource allocation, and resilient consensus under network compromise.

4. Neural and Flow-based OT for Continuous and High-Dimensional Distribution Control

Advanced neural OT frameworks enable distribution control with exact marginal preservation, monotonic cost decrease, and dynamic path modeling:

• c-Rectified Flow: Constructs a sequence of neural-parameterized ODE flows that always respect marginal constraints, reduce expected cost monotonically, and reduce the optimization to unconstrained regression over time-dependent potentials. Any fixed point is a cost-optimal coupling, and the flow is interpreted as a monotonic cost-decreasing path inside the transport polytope (Liu, 2022).
• Displacement Interpolation OT Models: Such as DIOTM, leverage the entire trajectory of Wasserstein geodesic interpolants between source and target distributions, enforcing the time-consistency (HJB) PDE structure. This approach leads to improved training stability, enhanced empirical fidelity, and enables the learning of time-dependent control policies for full geodesic distribution steering (Choi et al., 2024).
• Diffusion Probabilistic Models with OT: DPM-OT computes a semi-discrete Brenier OT map as a cost-minimizing “expressway” between white noise and data latents in a diffusion model, dramatically accelerating sampling and mitigating mode mixture due to intrinsic discontinuities in the OT map. The resulting frameworks admit explicit ELBO error bounds and outperform multi-step samplers in both speed and empirical quality on generative benchmarks (Li et al., 2023).

5. Structural, Robust, and Scalable Distribution Control Extensions

Several extensions improve reliability, interpretability, and applicability in the presence of noise and high dimensionality:

• Latent-Anchor OT (LOT): Imposes low-rank structure by routing mass through a small number of “anchor” points, efficiently solving OT optimization via alternating Sinkhorn projections and closed-form anchor updates. LOT regularizes against outliers, denoises domain shifts, and achieves quantifiable sample-rate improvements over unstructured OT (Lin et al., 2020).
• Displacement-Reshaped OT (ReshapeOT): Integrates known displacement statistics into the ground metric via a Mahalanobis-like distance, reshaping the cost surface to promote alignment along empirically validated transport directions. This produces more reliable solutions, aligns with geometric priors, and is directly compatible with any OT solver, including kernelized variants for high-dimensional feature spaces (Naumann et al., 6 May 2026).

6. Applications: Robust Control, Multi-Agent Coordination, and Scientific Data Analysis

OT-based distribution control has transformative applications across multiple domains:

• Data Compression in Distributionally Robust Control: OT compresses large behavioral datasets to compact, synthetic representative atoms via Wasserstein barycenters. Robust DeePC controllers built on compressed supports maintain out-of-sample guarantees while allowing substantially faster computation and longer receding-horizon operation (Fabiani et al., 2020).
• Distributed Multi-Agent Swarm Control: Agents implement local primal-dual potential estimation and proximal descent steps, steering their empirical configuration to track desired target distributions with mean-field guarantees. The schemes admit distributed communication and convergence analysis in both static and time-varying graph topologies (Krishnan et al., 2018, Duan et al., 8 Oct 2025).
• Cosmological Matching and Sensitivity Analysis: SHAM-OT recasts subhalo abundance matching as an OT matching on discretized mass functions, supporting exact, regularized, and multi-marginal couplings at $T_\#\mu=\nu$7100$T_\#\mu=\nu$8 lower computational cost and permitting rapid sensitivity studies to cosmological parameter variation (Fischbacher et al., 24 Feb 2025).

7. Theoretical Guarantees and Limitations

OT-based distribution control mechanisms are equipped with explicit optimality, convergence, and robustness guarantees:

• Existence and Uniqueness: Guaranteed by convexity-compactness arguments (Brenier, von Neumann) for static, dynamic, and game-theoretic OT problems. Saddle points always exist; uniqueness may fail in absence of strict convexity (Peyré, 10 May 2025, Hughes et al., 2021).
• Convergence: All major algorithmic schemes (Sinkhorn, block-coordinate Frank–Wolfe, distributed ADMM) have sublinear or geometric convergence under standard conditions (Fukunaga et al., 2021, Krishnan et al., 2018).
• Sample-Efficient Structure: Anchor-point and displacement-informed variants (LOT, ReshapeOT) offer improved sample efficiency, strict upper bounds, and reliability, as formally shown in sampling rate results and “expressway” inequalities (Lin et al., 2020, Naumann et al., 6 May 2026).

A limitation common to all OT-based control methods is sensitivity to misspecification of cost, regularization, or anchor/metric choices, especially in high dimensions or highly degenerate target structures. Careful calibration, adaptive selection, and physical knowledge integration are essential for robust deployment.

---

References:

• (Peyré, 10 May 2025) Optimal Transport for Machine Learners
• (Hughes et al., 2021) Resilient and Distributed Discrete Optimal Transport with Deceptive Adversary: A Game‐Theoretic Approach
• (Fukunaga et al., 2021) Fast block-coordinate Frank-Wolfe algorithm for semi-relaxed optimal transport
• (Liu, 2022) Rectified Flow: A Marginal Preserving Approach to Optimal Transport
• (Choi et al., 2024) Improving Neural Optimal Transport via Displacement Interpolation
• (Lin et al., 2020) Making transport more robust and interpretable by moving data through a small number of anchor points
• (Naumann et al., 6 May 2026) Reliable Modeling of Distribution Shifts via Displacement-Reshaped Optimal Transport
• (Krishnan et al., 2018) Distributed Online Optimization for Multi-Agent Optimal Transport
• (Fischbacher et al., 24 Feb 2025) SHAM-OT: Rapid Subhalo Abundance Matching with Optimal Transport
• (Blanchet et al., 2023) Unifying Distributionally Robust Optimization via Optimal Transport Theory
• (Li et al., 2023) DPM-OT: A New Diffusion Probabilistic Model Based on Optimal Transport
• (Duan et al., 8 Oct 2025) Trajectory-Optimized Density Control with Flow Matching
• (Fabiani et al., 2020) The optimal transport paradigm enables data compression in data-driven robust control