Generative optimal transport via forward-backward HJB matching
A new mathematical framework lets you steer a chaotic physical system toward a target state using only forward simulations — no backward guessing required.
The Thesis
Moving a disordered system — say, a cloud of molecules or a noisy data distribution — into a precise target configuration is a core problem in both physics and machine learning. The catch has always been circular: computing the optimal steering strategy requires knowing trajectories that already reach the target, which is exactly what you're trying to build. This paper cuts that knot by proving a 'time-reversal duality': the equations governing the hard backward problem are equivalent to a forward equation whose solution can be read off from easy, natural simulations of the system relaxing on its own. The result is a principled, sample-efficient method for what's called 'stochastic optimal transport' — moving probability distributions from one shape to another at minimum energetic cost. Practical impact would show up first in molecule generation, protein design, and generative AI pipelines where controlling the path of diffusion matters as much as the endpoint.
Catalyst
Diffusion-based generative models (systems that learn to reverse a noisy 'melting' process to produce structured outputs like images or molecules) have become a dominant paradigm in machine learning over the last three years, creating strong demand for theoretical frameworks that make their training more principled and their outputs more controllable. Simultaneously, the Schrödinger bridge problem — finding the most likely stochastic path between two probability distributions — has attracted serious computational attention, with new solvers appearing since 2021. This paper arrives at the intersection of both trends, offering a unifying theory grounded in classical physics (non-equilibrium statistical mechanics) that the community now has the numerical tools to actually implement.
What's New
Earlier approaches to the Schrödinger bridge and stochastic optimal control problems (such as IPFP — iterative proportional fitting procedure — and more recent score-based methods) required alternating between forward and backward simulations, or knowing the target distribution analytically. Those iterative methods can be expensive and unstable when the two distributions are very different. This paper shows that the Hamilton-Jacobi-Bellman (HJB) equation — the central equation of optimal control theory — governing the backward problem can be exactly rewritten as a forward equation, whose solution is a 'free energy' (a physics concept measuring accessible work) computed by averaging over the easy forward trajectories. The claimed advantage is eliminating backward simulation entirely while retaining a physically interpretable, theoretically grounded framework.
The Counter
The paper's core contribution is a theoretical equivalence — a proof that two equations describe the same object — illustrated with small numerical toy examples. There are no benchmarks against state-of-the-art generative models, no protein or molecule generation results, and no comparison to the iterative Schrödinger bridge solvers it implicitly competes with. The Cole-Hopf transformation and Feynman-Kac representation used here are classical tools; showing they apply in this setting is elegant but not obviously sufficient to produce a practical algorithm at the scale modern generative models require. The free-energy average over forward trajectories could have high variance in high dimensions, which is precisely the regime that matters for drug discovery or image generation — and the paper does not address this. Finally, the connection to Fermat's Principle, while evocative, is an analogy, not a scaling guarantee. The gap between a clean theoretical framework and a competitive implementation is large, and the paper does not close it.
Longs
- RXRX (Recursion Pharmaceuticals) — molecule generation and biological transport problems are direct use cases
- SCHD/XBI (biotech ETF) — broad exposure to computational drug discovery beneficiaries
- IONQ — quantum simulation of many-body stochastic systems is a downstream application
- NVDA — diffusion model training infrastructure benefits from more sample-efficient methods
Shorts
- Vendors of iterative Schrödinger bridge solvers — if forward-only computation proves robust, the iterative alternating approach loses its justification
- Flow-matching startups (e.g., those building on CNF/normalizing flow architectures) — a more physically principled diffusion framework could shift practitioner preference
Enablers (Picks & Shovels)
- JAX and PyTorch autodiff ecosystems — required for implementing HJB solvers at scale
- OpenFold / AlphaFold infrastructure — protein structure datasets that would serve as target ensembles
- DiffSBDD and related molecular diffusion open-source codebases — existing pipelines this theory could plug into
- arXiv math.PR and cs.LG communities — cross-pollination between control theory and ML that made this synthesis possible
Private Watchlist
- Isomorphic Labs — Google DeepMind spinout focused on ML-driven molecular design
- Insilico Medicine — generative chemistry pipelines that use diffusion-style methods
- Genesis Therapeutics — structure-based drug design using learned molecular dynamics
Resources
The Paper
Controlling the evolution of a many-body stochastic system from a disordered reference state to a structured target ensemble, characterized empirically through samples, arises naturally in non-equilibrium statistical mechanics and stochastic control. The natural relaxation of such a system - driven by diffusion - runs from the structured target toward the disordered reference. The natural question is then: what is the minimum-work stochastic process that reverses this relaxation, given a pathwise cost functional combining spatial penalties and control effort? Computing this optimal process requires knowledge of trajectories that already sample the target ensemble - precisely the object one is trying to construct. We resolve this by establishing a time-reversal duality: the value function governing the hard backward dynamics satisfies an equivalent forward-in-time HJB equation, whose solution can be read off directly from the tractable forward relaxation trajectories. Via the Cole-Hopf transformation and its associated Feynman-Kac representation, this forward potential is computed as a path-space free energy averaged over these forward trajectories - the same relaxation paths that are easy to simulate - without any backward simulation or knowledge of the target beyond samples. The resulting framework provides a physically interpretable description of stochastic transport in terms of path-space free energy, risk-sensitive control, and spatial cost geometry. We illustrate the theory with numerical examples that visualize the learned value function and the induced controlled diffusions, demonstrating how spatial cost fields shape transport geometry analogously to Fermat's Principle in inhomogeneous media. Our results establish a unifying connection between stochastic optimal control, Schrödinger bridge theory, and non-equilibrium statistical mechanics.