Optimal transport (OT) and its dynamic counterpart, the Schrödinger Bridge (SB) problem, are fundamentally challenging because they operate over the infinite-dimensional space of all possible probability measures. To compute an optimal transport plan, a barycenter, or a generative trajectory, researchers typically resort to costly iterative algorithms, discretization grids, or deep neural networks.
However, a remarkable simplification occurs when the underlying distributions are Gaussian: the optimal transport plans, interpolating paths, and barycenters often remain strictly within the Gaussian family. This phenomenon—Gaussian transport plan preservation—collapses infinite-dimensional functional optimization into finite-dimensional matrix algebra, yielding exact, closed-form solutions.
Why Gaussians Stay Gaussian
At its core, this tractability stems from the fact that the log-density of a Gaussian distribution is a quadratic form. When optimization objectives (such as the Wasserstein-2 distance) or regularization terms interact with these quadratic forms, the resulting dual potentials or optimal transport maps remain linear or quadratic.
Consequently, instead of solving complex partial differential equations or functional fixed-point equations over arbitrary probability spaces, we can solve algebraic matrix equations, such as Riccati or Lyapunov equations. This turns a optimization problem that would normally scale exponentially with dimensionality into a highly efficient linear algebra problem.
Static Transport and Wasserstein Barycenters
In static optimal transport, finding the average of multiple distributions—known as the Wasserstein barycenter—is generally computationally intensive. But for multivariate Gaussians, the barycenter is also Gaussian, governed by the elegant geometry of the Bures-Wasserstein space.
Peijie Qiu et al. leverage this closed-form Bures-Wasserstein barycenter to aggregate data modalities efficiently in variational autoencoders without relying on expensive numerical solvers. Similarly, Kevin C Cheng et al. show that modeling discrete time-series states as Gaussians decouples the optimization of mean and covariance parameters, making barycentric time-series modeling tractable.
This structural preservation extends to even more complex geometric settings:
- Subspace Projections: Nathaël Gozlan et al. prove that generalized Wasserstein barycenters for distributions living on entirely different subspaces remain Gaussian, satisfying an explicit matrix fixed-point equation.
- Linear Ensembles: Johan Karlsson et al. reduce the problem of tracking large ensembles of linear systems to a Semidefinite Programming (SDP) problem by exploiting Gaussian properties.
- Signal Normalization: For biological signals, Rémi Flamary et al. utilize element-wise operations in the frequency domain derived from the closed-form Gaussian barycenter to normalize EEG data rapidly across subjects.
- Mixture Models: Even when dealing with complex, multi-modal mixtures, Allen Tannenbaum et al. preserve structural information by treating Gaussian Mixture Models (GMMs) as discrete distributions over the space of Gaussians, computing distances directly via the constituent components.
The Resilience of Entropic Regularization
One might expect that adding entropic regularization—which introduces a smoothing effect via a Kullback-Leibler (KL) divergence term—would break this clean algebraic structure. Remarkably, the Gaussian family is resilient to entropic smoothing.
Hicham Janati et al. demonstrate that the entropic optimal transport plan between unbalanced Gaussian measures is itself a scaled Gaussian measure. By assuming a quadratic form for the dual potentials, the authors simplify functional fixed-point equations into algebraic matrix equations. This preservation ensures that the entropic barycenter remains strictly within the Gaussian family, a property utilized by Ryan Cumings-Menon et al. to combine probability forecasts analytically.
Furthermore, this structural resilience holds under temporal constraints. Beatrice Acciaio et al. show that the optimal bi-causal transport plan for time-series data maintains a Gaussian structure under entropic regularization, allowing for efficient optimization directly over covariance matrices while respecting the forward flow of time.
Closed-Form Schrödinger Bridges and Diffusion Models
The preservation principle shines brightest in dynamic settings, where a distribution evolves over time. The Schrödinger Bridge (SB) problem seeks the most likely stochastic path between two distributions under a reference diffusion process. When the boundary distributions are Gaussian and the reference dynamics are linear stochastic differential equations (SDEs), the entire optimal transport path is a Gaussian process.
Charlotte Bunne et al. derived the first closed-form analytical solutions for this problem, providing explicit kernels for the time-varying mean and covariance. This eliminates the need for iterative numerical solvers or neural network training, a breakthrough also leveraged by Stephen Y. Zhang et al. to enable simulation-free reconstruction of non-equilibrium cellular trajectories using multivariate Ornstein-Uhlenbeck processes.
[Gaussian μ₀, Σ₀] ==================================> [Gaussian μ₁, Σ₁]
Boundary Condition Closed-Form Linear SDE Boundary Condition
(Gaussian Process Path)
In a similar vein, Marc Lambert demonstrates that under a Linear Quadratic Regulator (LQR) cost, the Kantorovich potentials remain Gaussian throughout the process, facilitating efficient closed-form updates. This linearity extends directly to generative diffusion models: Emile Pierret et al. show that for Gaussian data, the score function (the gradient of the log-density) reduces to a simple linear operator, enabling exact closed-form solutions for backward SDEs and probability flow ODEs. Even when moving to non-Euclidean domains, Maosheng Yang shows that the optimal bridge between Gaussian distributions on topological structures can be derived in an exact closed form.
Relational Alignment via Gromov-Wasserstein
When comparing distributions across entirely different metric spaces (where direct point-to-point distance is undefined), researchers turn to the Gromov-Wasserstein (GW) distance. Amazingly, the Gaussian preservation principle still applies.
Khang Le et al. prove that the optimal transport plan for the entropic inner product GW distance between two Gaussians is itself a Gaussian distribution, simplifying the optimization to a restricted parametric space. When looking at the unregularized setting, Antoine Salmona et al. establish that constraining the GW transport plan to be Gaussian makes the problem equivalent to the linear Gromov-Monge problem. The optimal plan simply aligns the principal components of the two distributions before mapping them, revealing a deep connection between optimal transport and Principal Component Analysis (PCA).
Broader Implications for Inference and Sampling
Beyond transport, this algebraic closure underpins adjacent sampling and variational frameworks. For instance, Yingdong Lu demonstrates that Hamiltonian Monte Carlo (HMC) operators applied to Gaussian variables with random Hamiltonians preserve the Gaussian property, allowing the sampler's state to be tracked precisely via its mean and covariance. In variational inference, Michael Ziyang Diao et al. show that the Jordan-Kinderlehrer-Otto (JKO) proximal operator for entropy possesses a closed-form algebraic solution within the Gaussian family, bypassing gradient approximations and accelerating convergence.
Ultimately, while the real world is rarely perfectly Gaussian, these exact solutions provide a rigorous, computationally efficient baseline. As demonstrated by Yulong Lu et al., complex, multimodal target distributions can be systematically approximated by minimizing the KL divergence to optimal Gaussian distributions or mixtures, allowing practitioners to anchor complex non-Gaussian problems around these highly tractable analytical bridges.
Key Foundational Work
State Tracking of Linear Ensembles via Optimal Mass Transport (2018) — contributed 1 concept elements to this direction.
Gaussian Approximations for Probability Measures on Rd (2017) — contributed 1 concept elements to this direction.