Many tasks need the "average" of several random processes rather than of several static distributions. Examples include pooling brain-signal dynamics across subjects, combining forecasters that each output a law over future paths, and building a consensus market model. The usual tool is the Wasserstein barycenter, the distribution that minimizes average transport cost to the inputs. For Gaussians it has a well-known closed form. The problem is that applying it to a whole time series treats the path as one long vector. That lets the transport "look ahead": it can decide how to move the value at time $t$ based on what happens later. The adapted Wasserstein distance fixes this. It allows only couplings in which each time step's assignment depends on the past alone. This is the distance under which optimal stopping, hedging and control are continuous. Adapted barycenters have been defined abstractly, but no one can compute them beyond tiny horizons.
This note proposes to show that, for Gaussian processes, the adapted barycenter is again a Gaussian process, and that it can be computed by an alternating algorithm. Each input's lower-triangular covariance factor (its Cholesky factor) is rotated by one small orthogonal matrix per time step to line up with the current estimate, and the aligned factors are then averaged. The result is causal by construction: the transport map from each input to the barycenter can be applied online, one time step at a time. That property is what makes it useful for real-time normalization and for sequential decisions.
Core Claim
Gaussian closed forms for adapted transport between two processes already exist. Entropic adapted Wasserstein distance on Gaussians proves that the entropically regularized bicausal coupling between discrete-time Gaussian processes stays Gaussian and unique. It gives the distance through Cholesky factors and a shrinkage function $f_\lambda$ applied to singular values of cross terms. The geometry of the adapted Bures–Wasserstein space describes the unregularized space as a quotient of block-lower-triangular square roots by time-consistent block-diagonal rotations. It proves non-negative curvature and gives exponential and logarithmic maps. Multicausal Transport: Barycenters and Dynamic Matching defines bicausal barycenters of $K$ processes, proves a dynamic programming principle and duality, and warns that computation suffers badly as the horizon grows. None of these papers gives the Gaussian barycenter in usable form. That missing piece is the gap we fill.
For regular Gaussian processes, meaning processes with full-rank conditional variance at each step, we read the quotient description as saying that the squared adapted distance between laws with means $a,b\in\mathbb{R}^{dT}$ and block-lower-triangular factors $L,M$ is
$$AW_2^2(\mu,\nu)=\|a-b\|^2+\min_{O\in\mathcal{O}}\|L-MO\|_F^2 .$$
Here $\mathcal{O}$ is the set of block-diagonal matrices with one $d\times d$ orthogonal block per time step. Establishing this reading rigorously is part of the work.
Our main conjecture has two parts. First, for weights $w_k$ and Gaussian inputs $(a_k,L_k)$, every minimizer of $\sum_k w_k AW_2^2(\rho,\mu_k)$ over all filtered processes $\rho$, not only over Gaussian ones, is Gaussian. Second, that minimizer has mean $\sum_k w_k a_k$ and factor $L^\star=\sum_k w_k L_k O_k^\star$, where each $O_k^\star$ is the optimal alignment of $L_k$ to $L^\star$. Each block of $O_k^\star$ is the polar factor of a $d\times d$ cross product of matching block columns. The factor $L^\star$ is automatically block-lower-triangular, because right-multiplying by a block-diagonal matrix and averaging both preserve that structure.
In the entropic version, the polar factor is replaced by the $f_\lambda$-shrunk alignment from Acciaio et al. The difference from existing work is that we answer the $K$-process question in closed algebraic form. Acciaio–Krsek–Pammer leave it abstract. Pammer et al. and Acciaio et al. solve only the two-process problem.
Research Program
The theoretical core has three parts.
The first part proves that the barycenter is Gaussian. The strategy combines the multicausal dynamic programming principle with quadratic dual potentials. This is the same mechanism that keeps entropic plans between Gaussians Gaussian in Janati et al. and keeps worst-case adversaries Gaussian in Lotidis et al.. We would construct a Gaussian candidate from the fixed point, then build a certificate with quadratic potentials that matches the lower bound from multicausal duality.
The second part studies the algorithm. Each alternating step, first align then average, cannot increase the objective, so the iteration decreases monotonically, as in generalized Procrustes analysis. Uniqueness is less clear. The space has non-negative curvature, and in such spaces barycenters need not be unique. We will characterize when they are unique, for example when the inputs lie in a small enough ball. We will also say what the algorithm returns when they are not.
The third part handles boundary cases. These include the entropic version, where strict convexity should restore uniqueness as it does in the two-process case. They also include irregular inputs with rank-dropping conditional variances. We treat these by regularizing with $\lambda\to 0$ and checking whether the limit sits on the tangent-cone boundary described by Pammer et al.
The computational output is a solver costing about $O(K T d^3)$ per iteration, plus causal transport maps from each input to the barycenter. Being causal means the map at time $t$ uses only inputs up to $t$.
Baselines
The naive reference point pools all samples, or averages the input covariance matrices entrywise. It then fits one Gaussian process. This is what practitioners do by default, and any structured barycenter has to beat it on downstream tasks to be worth the effort.
The strongest existing method is the static Bures–Wasserstein barycenter of the flattened $dT\times dT$ covariances. This is the closed-form object used for cross-subject EEG normalization by Flamary et al. and for modality fusion by Qiu et al.. It is the strongest competitor because it is exact, cheap and Gaussian-preserving, and differs from our method only in ignoring information flow over time. It will receive the same inputs and the same covariance estimates. It will be solved by its standard fixed-point iteration to the same tolerance, and its transport maps will be used for normalization. No reported result was found on Lacuna for either static or adapted barycenters on the tasks below, so every number will be produced in-house under matched conditions. We beat this baseline if we achieve lower error on the sequential tasks, and if we can do online (causal) normalization, which its non-causal maps cannot.
For small instances, the ground-truth reference is the discretized multicausal dynamic program of Acciaio–Krsek–Pammer, run on binomial-tree approximations of the Gaussian inputs. Two ablations isolate the key ingredient. The first averages Cholesky factors without alignment. The second computes per-time-step marginal barycenters and ignores cross-time structure.
Why not just…?
Why not just use the static Bures–Wasserstein barycenter on the flattened sequence? It allows anticipative couplings. The literature shows that models close in static Wasserstein distance can have very different optimal-stopping values (Backhoff-Veraguas et al., Bartl et al.). Lotidis et al. report that clairvoyant, non-causal uncertainty sets degrade estimators. Our stopping-value experiment tests this directly. If static and adapted barycenters give indistinguishable stopping values and online normalization accuracy, the proposal has little practical value, and we will report that.
Why not just average the positive-diagonal Cholesky factors? This is a real threat when $d=1$. Each alignment block is then just a sign, and positive diagonals often already fix the sign, so this shortcut may frequently coincide with our barycenter. We will measure how often it fails to align for $d=1$ and how far off it is when it does. The case for our method rests on $d>1$, where the alignments are genuine rotations, and on the guarantee that the answer is optimal over all processes. The ablation will show how large the gap is.
Why not just run a numerical causal-transport solver on samples? Empirical measures do not converge under the adapted distance without smoothing (Blanchet et al., Estimating causal distances with non-causal ones). Multicausal computation also scales poorly with the horizon. We use such solvers only as small-scale ground truth, and we report the horizon at which they become impractical.
Proposed Experiments
The first experiment checks exactness. We draw random Gaussian processes, including AR-1 processes like those used by COT-GAN, with $T\le 5$ and $d\le 2$. We compare our barycenter's objective value with the tree-based multicausal solver as the tree is refined. This includes non-Gaussian candidate barycenters, which directly tests the claim that the optimum is Gaussian.
The second experiment is the decisive one: stability of sequential decisions. We compute optimal-stopping and American-put values under the inputs, under the adapted barycenter and under the static barycenter. We then test whether the barycenter's value is controlled by $\sum_k w_k AW_2(\rho,\mu_k)$, as continuity under the adapted distance predicts. The same test is repeated on Gaussian fits to the Stock series from DMD-GEN.
The third experiment is causal cross-subject EEG normalization. Following Flamary et al.'s protocol, we fit per-subject multichannel Gaussian processes over windows and map each subject to the barycenter causally in streaming mode. We compare cross-subject decoding accuracy against static-barycenter normalization, both offline and restricted to online use.
The fourth experiment is forecast combination on ETTh (loader from SCINet). We combine the predictive path laws of several Gaussian forecasters and score held-out paths by log-likelihood and multi-step calibration.
Across all experiments we also sweep $\lambda$ to trace the entropic barycenter as it approaches the unregularized one, and we record runtime against $T$ and $d$.
References
- Entropic adapted Wasserstein distance on Gaussians
- The geometry of the adapted Bures--Wasserstein space
- Multicausal Transport: Barycenters and Dynamic Matching
- Entropic Optimal Transport between Unbalanced Gaussian Measures has a Closed Form
- Wasserstein Distributionally Robust Linear-Quadratic Estimation under Martingale Constraints
- Convolution Monge Mapping Normalization for learning on sleep data
- Multimodal Variational Autoencoder: A Barycentric View
- Causal Transport in Discrete Time and Applications
- The Wasserstein space of stochastic processes
- Bounding adapted Wasserstein metrics
- Estimating causal distances with non-causal ones
- COT-GAN: Generating Sequential Data via Causal Optimal Transport
- A Geometry-Aware Metric for Mode Collapse in Time Series Generative Models