Research map / Probabilistic and scientific modeling / Variational and generative sampling
PDPO: Parametric Density Path Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PDPO, which recasts action-minimizing density path problems as optimization over cubic Hermite spline control points in the parameter space of Neural ODE pushforward maps, with alternating path and boundary/coupling optimization.
Problem
Computing action-minimizing (least-action) paths between probability densities with potential terms (obstacles, entropy/Fisher information, mean-field interactions) is infeasible in high dimension when solved directly as PDEs over densities.
Approach
Represent density paths as pushforwards of a reference density through a parametric Neural ODE map, and optimize a cubic Hermite spline through 3-5 control points in parameter space; a static coupling-cost formulation replaces the dynamic density problem, with an error bound on action under regularity assumptions.
Key results
- On SCC, PDPO reaches action 39.90±0.16 versus 40.02±0.75 for GSBM and 39.95±0.25 for APAC-Net, with 11 min runtime versus 21 min for GSBM.
- On VNEFI, PDPO achieves action 148.91±0.72 versus 156.80±4.31 for GSBM, in 70 min versus 115 min.
- On GMM, PDPO achieves action 79.46±0.62 versus 88.77±3.31 for GSBM and 82.06±2.80 for NLOT, in 8 min versus 123 min for GSBM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PDPO instead of GSBM | SCC · SCC | action: 40.02± .75 -> 39.90± 0.16 | Helped |
| Use PDPO instead of GSBM | SCC · SCC | runtime: 21m -> 11m | Helped |
| Use PDPO instead of GSBM | SCC · SCC | boundary W2: (0.21, 0.25) -> (\mathbf{0.028},\mathbf{0.014}) | Helped |
| Use PDPO instead of GSBM | VNEFI · VNEFI | action: 156.80\pm4.31 -> 148.91\pm0.72 | Helped |
| Use PDPO instead of GSBM | VNEFI · VNEFI | runtime: 115m -> 70m | Helped |
| Use PDPO instead of GSBM | GMM · GMM | action: 88.77\pm 3.31 -> 79.46\pm 0.62 | Helped |
| Use PDPO instead of GSBM | GMM · GMM | runtime: 123m -> 8m | Helped |
| Use PDPO instead of NLOT | GMM · GMM | action: 82.06\pm 2.80 -> 79.46\pm 0.62 | Helped |
| Increase time discretization from N=10 to N=50 (ablation) | PDPO · Action convergence with time discretization | action: 66.06 -> 30.10 | Helped |
Surprising findings
- Reported GSBM obstacle-avoidance violations: both GSBM and APAC-Net violate the obstacle constraint in the SCC scenario.
Limitations
- Currently limited to MLP-based Neural ODE architectures, which restricts application to image-based transport problems.
- Lipschitz assumptions for the error bound (especially for F in W2) are strong, and the authors leave relaxing them to future work.
- Computing entropy and Fisher information adds costly ODEs, scaling as O(N*M*(1+d)), so high-dimensional Fisher information cases are costly.
Experimental setup
Datasets: S-curve with congestion (SCC) V-neck with entropy and Fisher information (VNEFI) Gaussian mixture obstacle (GMM) high-dimensional opinion depolarization
Benchmarks: SCC VNEFI GMM high-dimensional opinion depolarization
Metrics: action A[rho_t] expected potential integral Wasserstein-2 boundary error runtime
Baselines: GSBM APAC-Net NLOT
Builds on: Generalized Schrödinger Bridge Method Neural ODE Flow Matching Benamou-Brenier formulation
Closest papers by meaning
- Constrained Diffusers for Safe Planning and Control (NeurIPS 2025 · similarity 0.84)
- Model-based Diffusion for Trajectory Optimization (NeurIPS 2024 · similarity 0.84)
- Deterministic Policy Gradient Primal-Dual Methods for Continuous-Space Constrained MDPs (AAAI 2025 · similarity 0.84)
- DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising (NeurIPS 2026 · similarity 0.84)
- Model-based Reinforcement Learning for Parameterized Action Spaces (ICML 2024 · similarity 0.83)
- Graph Sparse Sampling: Breaking the Curse of the Horizon in Continuous MDP Planning (NeurIPS 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.