Barycentric Coding Model (BCM) Overview

• BCM is a representation framework where an unknown measure is expressed as the Wasserstein barycenter of a finite set of reference measures.
• It recovers barycentric coordinates by formulating a convex quadratic program based on inner products of tangent-space displacement fields.
• The model is applied in Gaussian covariance estimation, image processing, and NLP, with established statistical consistency and convergence guarantees.

Searching arXiv for recent and foundational papers on the Barycentric Coding Model. Search query: "Barycentric Coding Model Wasserstein barycenter analysis" The Barycentric Coding Model (BCM) is a representation and estimation framework in which an unknown object is assumed to lie in a barycenter family generated by a finite set of known references, and inference is reduced to recovering the corresponding barycentric coordinates. In the formulation developed for probability measures, BCM assumes that an unknown measure belongs to the set of Wasserstein-2 barycenters of known reference measures, so that measure estimation becomes equivalent to estimating a vector of simplex-constrained coordinates (Werenski et al., 2022). The same synthesis–analysis viewpoint also appears in later extensions to Gromov–Wasserstein geometry for finite metric spaces and networks (Martín et al., 14 Jul 2025), and it has conceptual affinities with barycentric subspaces on manifolds, where weighted means of reference points define geometry-aware coordinate systems (Pennec, 2016). In the modern literature, however, the term most commonly denotes the Wasserstein analysis model of measure estimation, whose central result is that barycentric coordinates can be recovered through a convex quadratic program built from inner products of optimal displacement maps (Werenski et al., 2022).

1. Formal definition in Wasserstein space

In the Wasserstein formulation, BCM starts from known reference probability measures $\{\nu_i\}_{i=1}^m$ and assumes that a target measure $\mu$ lies in the set of Wasserstein-2 barycenters of these references. Thus there exists $\theta \in \Delta_m$ such that

$\mu = \mathrm{Bary}_{W_2}(\{\nu_i\};\theta),$

where the probability simplex is

$$\Delta_m = \{\theta \in \mathbb{R}^m : \theta_i \ge 0,\ \sum_{i=1}^m \theta_i = 1\}.$$

The Wasserstein-2 distance is defined by

$$W_2^2(\mu,\nu)=\inf_{\pi\in\Pi(\mu,\nu)} \int \|x-y\|^2\, d\pi(x,y),$$

or, under absolute continuity, through the Monge formulation

$$W_2^2(\mu,\nu)=\min_{T_{\#}\mu=\nu}\int \|T(x)-x\|^2\, d\mu(x).$$

The barycenter associated with weights $\theta$ is

$$\nu_\theta=\arg\min_\nu \frac12\sum_{i=1}^m \theta_i W_2^2(\nu,\nu_i).$$

Under BCM, the analysis problem is to recover $\theta$ from $\mu$0 and the references, ideally by solving

$\mu$1

with minimum $\mu$2 exactly when $\mu$3 belongs to the barycenter set (Werenski et al., 2022).

The geometric theory in the measure-estimation paper is developed under three regularity assumptions. The measures are absolutely continuous with shared support $\mu$4 that is either all of $\mu$5 or a bounded open convex set; their densities are bounded above, with strict positivity for the reference densities; and additional regularity is imposed through local Hölder continuity in $\mu$6 or lower density bounds on bounded supports. These assumptions ensure existence and uniqueness of optimal transport maps and permit the identification of barycenters with Karcher means in the relevant setting (Werenski et al., 2022).

This formulation makes BCM an analysis model: barycenter computation is the synthesis map from coordinates to an object, while BCM inverts that map by estimating coordinates from an observed object. A plausible implication is that BCM occupies, in Wasserstein geometry, a role analogous to sparse coding or dictionary analysis in Euclidean spaces, but with transport-induced geometry replacing linear superposition.

2. Geometric mechanism and the quadratic program

The principal insight of BCM is that coordinate recovery can be expressed in the tangent geometry of Wasserstein space. If $\mu$7 is absolutely continuous, the optimal transport from $\mu$8 to $\mu$9 is uniquely given by $\theta \in \Delta_m$0, and the associated displacement map is

$\theta \in \Delta_m$1

At $\theta \in \Delta_m$2, the tangent-space inner product is the $\theta \in \Delta_m$3 inner product,

$\theta \in \Delta_m$4

For the barycenter variance functional

$\theta \in \Delta_m$5

its gradient at $\theta \in \Delta_m$6 is

$\theta \in \Delta_m$7

Hence, whenever $\theta \in \Delta_m$8 is a barycenter, the weighted displacement fields cancel in the tangent space (Werenski et al., 2022).

This cancellation yields a Gram representation. Define

$\theta \in \Delta_m$9

Then

$\mu = \mathrm{Bary}_{W_2}(\{\nu_i\};\theta),$0

Because $\mu = \mathrm{Bary}_{W_2}(\{\nu_i\};\theta),$1 is a Gram matrix, it is symmetric positive semidefinite. Proposition 1 in the paper establishes this representation, and Theorem 1 shows that under the regularity assumptions, $\mu = \mathrm{Bary}_{W_2}(\{\nu_i\};\theta),$2 belongs to the barycenter set if and only if the convex quadratic program

$\mu = \mathrm{Bary}_{W_2}(\{\nu_i\};\theta),$3

has minimum value $\mu = \mathrm{Bary}_{W_2}(\{\nu_i\};\theta),$4; any minimizer $\mu = \mathrm{Bary}_{W_2}(\{\nu_i\};\theta),$5 then satisfies

$\mu = \mathrm{Bary}_{W_2}(\{\nu_i\};\theta),$6

In the main result, the objective is purely quadratic: the more general template

$\mu = \mathrm{Bary}_{W_2}(\{\nu_i\};\theta),$7

is discussed only as a possible extension, for example under regularization or approximate map estimation, whereas the exact BCM result has $\mu = \mathrm{Bary}_{W_2}(\{\nu_i\};\theta),$8 and $\mu = \mathrm{Bary}_{W_2}(\{\nu_i\};\theta),$9 (Werenski et al., 2022).

The paper further identifies a compatibility regime in which this tangent-space criterion coincides with exact projection onto the barycenter family. If the family $$\Delta_m = \{\theta \in \mathbb{R}^m : \theta_i \ge 0,\ \sum_{i=1}^m \theta_i = 1\}.$$0 is compatible in the sense that optimal maps compose, then the quadratic-program solution also solves

$$\Delta_m = \{\theta \in \mathbb{R}^m : \theta_i \ge 0,\ \sum_{i=1}^m \theta_i = 1\}.$$1

so the BCM analysis problem and the tangent-space surrogate are identical (Werenski et al., 2022).

3. Identifiability, exactness, and statistical estimation

Because the BCM objective is governed by a Gram matrix, identifiability is determined by the linear dependence structure of the displacement fields $$\Delta_m = \{\theta \in \mathbb{R}^m : \theta_i \ge 0,\ \sum_{i=1}^m \theta_i = 1\}.$$2. If $$\Delta_m = \{\theta \in \mathbb{R}^m : \theta_i \ge 0,\ \sum_{i=1}^m \theta_i = 1\}.$$3 and the simplex intersects the nullspace of $$\Delta_m = \{\theta \in \mathbb{R}^m : \theta_i \ge 0,\ \sum_{i=1}^m \theta_i = 1\}.$$4, then the minimizer is unique. If $$\Delta_m = \{\theta \in \mathbb{R}^m : \theta_i \ge 0,\ \sum_{i=1}^m \theta_i = 1\}.$$5 and multiple eigenvectors associated with the zero eigenvalue intersect the simplex, there may be infinitely many minimizers. If the nullspace does not intersect the simplex, the minimum is strictly positive; in that case $$\Delta_m = \{\theta \in \mathbb{R}^m : \theta_i \ge 0,\ \sum_{i=1}^m \theta_i = 1\}.$$6 is not exactly a barycenter of the references, and the quadratic program returns an approximate coordinate vector (Werenski et al., 2022).

The sampling model in the empirical theory assumes i.i.d. samples from the target and each reference:

• $$\Delta_m = \{\theta \in \mathbb{R}^m : \theta_i \ge 0,\ \sum_{i=1}^m \theta_i = 1\}.$$7 from $$\Delta_m = \{\theta \in \mathbb{R}^m : \theta_i \ge 0,\ \sum_{i=1}^m \theta_i = 1\}.$$8,
• $$\Delta_m = \{\theta \in \mathbb{R}^m : \theta_i \ge 0,\ \sum_{i=1}^m \theta_i = 1\}.$$9 from $$W_2^2(\mu,\nu)=\inf_{\pi\in\Pi(\mu,\nu)} \int \|x-y\|^2\, d\pi(x,y),$$0 for $$W_2^2(\mu,\nu)=\inf_{\pi\in\Pi(\mu,\nu)} \int \|x-y\|^2\, d\pi(x,y),$$1.

The empirical procedure has three stages. First, one estimates the optimal maps from $$W_2^2(\mu,\nu)=\inf_{\pi\in\Pi(\mu,\nu)} \int \|x-y\|^2\, d\pi(x,y),$$2 to each $$W_2^2(\mu,\nu)=\inf_{\pi\in\Pi(\mu,\nu)} \int \|x-y\|^2\, d\pi(x,y),$$3 using entropic regularization, specifically the entropic map estimator of Pooladian–Niles-Weed. This yields empirical displacement fields $$W_2^2(\mu,\nu)=\inf_{\pi\in\Pi(\mu,\nu)} \int \|x-y\|^2\, d\pi(x,y),$$4. Second, one forms empirical Gram entries

$$W_2^2(\mu,\nu)=\inf_{\pi\in\Pi(\mu,\nu)} \int \|x-y\|^2\, d\pi(x,y),$$5

A point-cloud variant replaces $$W_2^2(\mu,\nu)=\inf_{\pi\in\Pi(\mu,\nu)} \int \|x-y\|^2\, d\pi(x,y),$$6 by a Sinkhorn plan and computes Gram entries through a trace formula on the observed support. Third, one solves

$$W_2^2(\mu,\nu)=\inf_{\pi\in\Pi(\mu,\nu)} \int \|x-y\|^2\, d\pi(x,y),$$7

for example by projected gradient descent or an off-the-shelf quadratic-program solver (Werenski et al., 2022).

The statistical theory establishes convergence rates for both Gram estimation and coordinate recovery. The paper gives an entrywise bound of the form

$$W_2^2(\mu,\nu)=\inf_{\pi\in\Pi(\mu,\nu)} \int \|x-y\|^2\, d\pi(x,y),$$8

with $$W_2^2(\mu,\nu)=\inf_{\pi\in\Pi(\mu,\nu)} \int \|x-y\|^2\, d\pi(x,y),$$9 and $$W_2^2(\mu,\nu)=\min_{T_{\#}\mu=\nu}\int \|T(x)-x\|^2\, d\mu(x).$$0, under supplementary smoothness assumptions on the optimal maps. Using these Gram-entry rates, Corollary 1 yields

$$W_2^2(\mu,\nu)=\min_{T_{\#}\mu=\nu}\int \|T(x)-x\|^2\, d\mu(x).$$1

under a uniqueness condition requiring a simple zero eigenvalue and exact barycentricity $$W_2^2(\mu,\nu)=\min_{T_{\#}\mu=\nu}\int \|T(x)-x\|^2\, d\mu(x).$$2. The constants may depend on the number of references, the eigengap $$W_2^2(\mu,\nu)=\min_{T_{\#}\mu=\nu}\int \|T(x)-x\|^2\, d\mu(x).$$3, the support diameter, density bounds, and the regularity of the optimal maps (Werenski et al., 2022).

These results establish statistical consistency of BCM under empirical sampling. They also make clear that performance is sensitive to smoothness and dimensionality, since both map estimation and Monte Carlo averaging appear in the final bound.

4. Canonical applications

The measure-estimation paper demonstrates BCM in three application areas: Gaussian covariance estimation, image processing, and natural language processing (Werenski et al., 2022).

Application area	BCM formulation	Reported outcome
Gaussian measures	Recover barycentric coordinates, then solve for the Bures–Wasserstein barycenter covariance	BCM outperforms empirical covariance and is competitive with maximum-likelihood optimization while being orders of magnitude faster and numerically more stable
Image processing	Use corrupted images to estimate coordinates, then reconstruct the clean barycenter from clean references	BCM reconstructions markedly outperform linear convex-combination projection baselines and are competitive in $$W_2^2(\mu,\nu)=\min_{T_{\#}\mu=\nu}\int \|T(x)-x\|^2\, d\mu(x).$$4 reconstruction error while being over an order of magnitude faster
Natural language processing	Represent documents as empirical measures over word embeddings and classify via recovered barycentric coordinates	BCM-based predictors outperform $$W_2^2(\mu,\nu)=\min_{T_{\#}\mu=\nu}\int \|T(x)-x\|^2\, d\mu(x).$$5 nearest-neighbor and minimum-average-distance baselines, especially with small training sets

In the Gaussian case, the references are $$W_2^2(\mu,\nu)=\min_{T_{\#}\mu=\nu}\int \|T(x)-x\|^2\, d\mu(x).$$6 and the barycenter is again Gaussian. For zero-mean Gaussians, if $$W_2^2(\mu,\nu)=\min_{T_{\#}\mu=\nu}\int \|T(x)-x\|^2\, d\mu(x).$$7 and $$W_2^2(\mu,\nu)=\min_{T_{\#}\mu=\nu}\int \|T(x)-x\|^2\, d\mu(x).$$8, the optimal map is linear,

$$W_2^2(\mu,\nu)=\min_{T_{\#}\mu=\nu}\int \|T(x)-x\|^2\, d\mu(x).$$9

and the Gram entries simplify to

$\theta$0

The corresponding barycenter covariance satisfies the Bures–Wasserstein fixed-point equation

$\theta$1

The empirical study reported for $\theta$2 and $\theta$3 finds that the BCM-based “gradient norm” estimator recovers $\theta$4 quickly and robustly (Werenski et al., 2022).

In image processing, the paper treats MNIST digits as probability measures on a $\theta$5 grid. Given a corrupted image and corresponding corrupted references, BCM estimates coordinates using entropic optimal transport on point clouds and reconstructs a clean barycenter from clean references. The reported blur is attributed to entropic regularization and can be tuned through $\theta$6 (Werenski et al., 2022).

In document classification, documents are empirical measures over word2vec embeddings, with support at word vectors and masses given by bag-of-words counts. The paper considers two BCM-based decision rules, “Minimum Barycenter Loss” and “Maximum Coordinate,” and reports improvements over $\theta$7 nearest-neighbor and minimum-average-distance baselines on BBCSport and 20NEWS, especially under limited labeled data (Werenski et al., 2022).

5. Relations to neighboring frameworks

BCM is closely related to, but distinct from, several other barycentric constructions.

First, it differs from ordinary mixture models. A linear mixture $\theta$8 is a convex combination in the space of measures and does not impose geometric alignment. BCM instead uses Wasserstein barycenters, which interpolate through displacement fields. The paper explicitly contrasts these views and argues that barycenters often preserve geometric features more faithfully than mixtures, for example in the Gaussian setting (Werenski et al., 2022).

Second, BCM has a manifold analogue in barycentric subspace analysis. On a Riemannian manifold, Pennec defines exponential, Fréchet, and Karcher barycentric subspaces as loci of weighted means of reference points, with the defining first-order condition

$\theta$9

In Euclidean space these constructions reduce to affine spans, while on spheres or hyperbolic spaces they generate geometry-adapted subspaces and nested flags used in Barycentric Subspaces Analysis (Pennec, 2016). The same source explicitly introduces a “Barycentric Coding Model” on manifolds: a code is a projective or simplex-constrained weight vector over references, reconstruction solves the weighted mean equation, and encoding either solves $$\nu_\theta=\arg\min_\nu \frac12\sum_{i=1}^m \theta_i W_2^2(\nu,\nu_i).$$0 or a simplex-constrained variance minimization problem (Pennec, 2016). This suggests a broader interpretation of BCM as coordinate recovery with respect to barycentric primitives in non-Euclidean spaces.

Third, BCM has been extended from classical Wasserstein geometry to Gromov–Wasserstein geometry. In that setting, the unknown object is a distance matrix or weighted network, the references may have different support sizes, and the analysis problem is again cast as estimating barycentric coordinates (Martín et al., 14 Jul 2025). Two methods are proposed there. One uses the fixed-point structure of GW barycenter computation and solves

$$\nu_\theta=\arg\min_\nu \frac12\sum_{i=1}^m \theta_i W_2^2(\nu,\nu_i).$$1

which becomes a convex quadratic program. The other uses a blow-up alignment technique to obtain an explicit gradient in an aligned weighted Frobenius geometry, leading to

$$\nu_\theta=\arg\min_\nu \frac12\sum_{i=1}^m \theta_i W_2^2(\nu,\nu_i).$$2

The parallel with the Wasserstein BCM is direct: in both cases, synthesis is barycenter computation and analysis is coordinate recovery by minimizing a quadratic form built from geometry-dependent inner products (Martín et al., 14 Jul 2025).

6. Limitations, misconceptions, and terminological ambiguity

A central limitation of BCM in the Wasserstein setting is computational. The method requires estimating one transport map per reference, constructing a Gram matrix, and then solving a simplex-constrained quadratic program. The paper emphasizes that entropic regularization makes this tractable and practical at the scale of its experiments, but also notes sensitivity to map-estimation errors in high dimensions (Werenski et al., 2022). Robustness is mediated by the eigengap of the Gram matrix and by the sampling error in the empirical maps, so poor identifiability or small eigengaps can materially degrade estimation (Werenski et al., 2022).

Another limitation concerns exactness. If the target is not exactly in the barycenter family, the quadratic program no longer achieves zero. In that case BCM returns the minimizer of a tangent-space criterion, which the paper interprets as a principled approximation. For compatible families this approximation coincides with exact projection, but not in full generality (Werenski et al., 2022).

A frequent misconception arises from the acronym itself. In neuroscience and neuromorphic engineering, “BCM” commonly denotes the Bienenstock–Cooper–Munro plasticity rule rather than any barycentric model. That usage is unrelated to Wasserstein barycenters or coordinate coding and concerns synaptic modification under spike-timing-dependent plasticity (Azghadi et al., 2012). Disambiguation is therefore necessary in cross-disciplinary contexts.

There is also terminological dispersion within the barycentric literature. The hierarchical-clustering paper on mixed-type data uses “barycentric coding” to denote a pseudo-disjunctive fuzzy recoding scheme for continuous variables compatible with correspondence analysis, but it does not define a named BCM in the sense of Wasserstein barycenter analysis (Moschidis et al., 2022). Likewise, recent work on distributed computing uses “Barycentric Coding Model” to refer to barycentric rational interpolation codes for flexible-threshold coded computation in mobile edge computing (Qiu et al., 11 Sep 2025). These usages share the general idea of representing objects via barycentric structure, but they do not instantiate the same model class.

7. Broader significance and prospective directions

Within optimal transport, BCM provides a precise bridge between barycenter synthesis and inverse estimation. Its main conceptual contribution is to reinterpret the analysis problem as a convex quadratic program on the simplex whose coefficients are inner products of geometry-induced displacement fields. This yields a formulation that is simultaneously geometric, statistical, and computational (Werenski et al., 2022).

The later GW extension indicates that this synthesis–analysis paradigm is not limited to shared ambient domains or classical Wasserstein transport. In GW space, barycentric coding applies to finite metric spaces and networks, where aligned supports are absent and geometry must be mediated through couplings or blow-up constructions (Martín et al., 14 Jul 2025). The manifold literature suggests an additional direction: BCM-like encoding can be formulated directly in terms of weighted means and affine spans on curved spaces, with projective or simplex-constrained coordinates and nested subspace structure (Pennec, 2016).

Several extensions are explicitly identified in the measure-estimation work. These include other $$\nu_\theta=\arg\min_\nu \frac12\sum_{i=1}^m \theta_i W_2^2(\nu,\nu_i).$$3 metrics, manifolds endowed with Wasserstein geometry, regularized barycenters such as entropic barycenters, and structured coordinates obtained by adding linear or penalty terms to the quadratic program (Werenski et al., 2022). This suggests that BCM is best viewed not as a single algorithmic recipe but as a family of analysis models whose common core is barycentric coordinate recovery under non-Euclidean geometry.

In that sense, BCM occupies a distinctive position in contemporary geometric inference. It replaces linear reconstruction by barycentric synthesis, replaces Euclidean dictionary atoms by reference measures or geometric templates, and replaces coefficient estimation by a simplex-constrained quadratic problem determined by the ambient transport or manifold structure. The resulting framework retains explicit coordinates, identifiability conditions, and statistical rates, while extending representation theory into spaces where interpolation is governed by geometry rather than vector addition (Werenski et al., 2022).