Traditional Bayesian computation and generative modeling often rely on a patient wait. In Markov Chain Monte Carlo (MCMC) or Langevin dynamics, we design a process that eventually reaches a target distribution after an "infinite" amount of time. This requirement for ergodicity—the guarantee that a system will eventually explore all states—comes with a heavy price: slow mixing times, difficult-to-diagnose burn-in periods, and the constant threat of mode collapse.
A growing body of research suggests a more direct alternative: instead of waiting for equilibrium, we can explicitly construct a "bridge" that transports a simple reference distribution to a complex target over a fixed, finite time interval.
The Geometry of the Path
The core intuition behind this shift is the construction of a geometric path. Rather than letting a particle wander aimlessly until it finds the target, we define a sequence of intermediate distributions that smoothly interpolate between a tractable starting point (like a Gaussian) and the objective.
In the context of Bayesian model selection, Iwona Hawryluk, Thomas A. Mellan, and colleagues demonstrate this in Application of referenced thermodynamic integration to Bayesian model selection. They use a power parameter to scale the influence of the reference and target densities. By navigating this geometric path, they can calculate normalizing constants—the "evidence" for a model—with orders of magnitude fewer samples than standard methods.
This interpolation is often achieved by linearly combining the log-densities of the two distributions. As Omar Chehab and Anna Korba explore in Provable Convergence and Limitations of Geometric Tempering for Langevin Dynamics, this "tempering" allows gradient-based samplers to follow a curriculum, gradually introducing the complex features of the target. However, they also warn that this path is not a panacea; if the interpolation creates artificial bottlenecks, convergence can still slow down exponentially.
From Ergodicity to Transport
If we view the transition as a transport problem rather than a search for equilibrium, the requirement for ergodicity disappears. In Schr{ö}dinger-F{ö}llmer Sampler: Sampling without Ergodicity, Yuling Jiao, Lican Kang, and their co-authors frame sampling as moving a point mass at time zero to a target distribution at time one.
Because the process is defined over a fixed unit interval, every generated path results in a valid sample. There is no "burn-in" because the system is never intended to reach a steady state; it is always in motion along a predefined schedule. This mechanism is particularly powerful for global optimization. Yin Dai and Jerry Zhijian Yang show in Global Optimization via Schr{ö}dinger-F{ö}llmer Diffusion that this finite-horizon approach can escape local minima and find global optima of non-convex functions with polynomial complexity.
Schrödinger Bridges and Stochastic Interpolants
The most mathematically rigorous way to define these paths is through the lens of Schrödinger bridges. Originally a problem in statistical mechanics, a Schrödinger bridge finds the "most likely" path between two distributions given a certain level of background noise.
Jeremy Heng and Randal Douc adapt this to Bayesian computation in Diffusion Schrödinger Bridges for Bayesian Computation. By solving a constrained optimization problem, they minimize the divergence between the transport path and a reference process. This allows them to "transport" a simple normal distribution to a complex posterior in finite time, bypassing the infinite-time requirements of standard diffusion models.
This idea is further generalized by the Stochastic Interpolants framework. Michael S. Albergo and Eric Vanden‐Eijnden provide a unified theory in Stochastic Interpolants: A Unifying Framework for Flows and Diffusions. Their framework allows practitioners to bridge any two arbitrary distributions. Once the interpolant is trained, one can choose to sample using either deterministic Ordinary Differential Equations (Flows) or stochastic Stochastic Differential Equations (Diffusions).
Deterministic Paths in Unit Time
While many bridging methods are stochastic, some recent advances focus on purely deterministic transport. Aimee Maurais and Youssef M. Marzouk developed Kernel Fisher–Rao Flow (KFRFlow), detailed in Sampling in Unit Time with Kernel Fisher-Rao Flow.
KFRFlow moves particles along a geometric mixture path by solving a Poisson equation within a Reproducing Kernel Hilbert Space. The result is a closed-form, deterministic trajectory that reaches the target in exactly one unit of time. By removing the randomness of the path, these methods offer a more predictable and often more efficient route to the target distribution, rivaling the performance of even the most sophisticated gradient-based samplers.
By shifting the focus from "waiting for equilibrium" to "engineering a path," this research direction provides a robust toolkit for high-dimensional sampling, global optimization, and generative modeling that is faster, more flexible, and theoretically grounded in finite-time transport.