--- title: Barycentric Quantile Map Overview url: https://www.emergentmind.com/topics/barycentric-quantile-map type: topic --- # Barycentric Quantile Map Overview Searching arXiv for recent papers directly relevant to barycentric quantile maps, transport-based quantiles, and barycentric projection. A **barycentric quantile map** is not a uniformly standardized object across the recent arXiv literature. The term is used most naturally for constructions that combine a common probability-indexing device with barycentric averaging, but the underlying mathematical realizations differ substantially. In one line of work, the closest object is a **conditional vector quantile map** defined as a Brenier transport from a fixed reference law to a posterior distribution, with nested images of latent balls yielding credible regions [2410.08378]. In another, the closest object is a **convex combination of optimal transport maps** that generates Wasserstein barycenters under compatibility assumptions [2201.12195]. A third line introduces an explicit multivariate quantile-like representation, the **HiMAP quantile map** \(\mathbf Q_\mu:[0,1]\to\mathbb R^d\), whose pointwise affine averages define well-posed multivariate barycenters and Fréchet regression estimators [2603.03674]. Related work on barycentric projections, barycentric algebras, and contractive barycentric maps supplies adjacent geometric, variational, and algebraic interpretations, but does not itself define the term in a probabilistic quantile sense [2606.07926], [2501.00937], [1805.08558]. ## 1. Terminological status and conceptual scope The recent literature does **not** present a single canonical definition of “barycentric quantile map.” Both "Deep Generative Quantile Bayes" and "Measure Estimation in the Barycentric Coding Model" explicitly state that they do **not** define an object with that name [2410.08378], [2201.12195]. This suggests that the phrase is best treated as an umbrella term for several nearby constructions rather than as an established primitive. Across these papers, three recurrent ideas organize the topic. First, a **quantile-like parameterization** provides a common coordinate or rank variable, such as a latent \(u\in\mathcal U\) or a scalar \(t\in[0,1]\). Second, a **transport or pushforward map** sends that parameter into the target distribution or target measure family. Third, a **barycentric operation** combines maps, displacement fields, or quantile representations across distributions. The precise meaning of “barycentric” therefore depends on context: it may refer to Wasserstein barycenters, convex combinations of Monge maps, or affine averaging in an induced function space. A useful distinction is between **transport-defined multivariate quantiles** and **barycentric averaging formulas**. The former emphasize Brenier or Monge structure, as in conditional posterior simulation [2410.08378]. The latter emphasize the algebra of averaging maps or displacement vectors, as in the barycentric coding model and HiMAP [2201.12195], [2603.03674]. A plausible implication is that the phrase “barycentric quantile map” is most precise when a single indexing variable has a distribution-invariant meaning and supports pointwise averaging across measures. ## 2. Optimal-transport quantiles and conditional Brenier maps In the most transport-theoretic usage, the closest analogue to a barycentric quantile map is the **vector quantile map** \(Q_P\), defined for a target law \(P\) on \(\mathbb R^d\) as the gradient of a convex potential \(\psi\). When \(P\) has finite second moments, \(Q_P\) is exactly the Monge map minimizing quadratic cost, \[ Q_P = \arg\min_{Q:\,Q\#F_U=P}\; \mathbb E_{U\sim F_U}\|Q(U)-U\|^2, \] and by Brenier’s theorem one has \(Q_P=\nabla\psi\) [2410.08378]. This is the paper’s main bridge between quantiles and Wasserstein geometry. For posterior simulation, the same paper uses the **conditional** version \[ Q_{\theta\mid X=x}(u)=\nabla_u\psi(u,x), \] with the defining pushforward relation \[ Q_{\theta\mid X=x}\#F_U=\pi(\theta\mid X=x). \] The base distribution \(F_U\) is supported on the unit Euclidean ball \(\mathcal U=S^d(1)\), with \(U=r\phi\), \(r\sim\mathrm{Unif}([0,1])\), \(\phi\) uniform on the sphere \(\mathcal S^{d-1}(1)\), and \(r\perp\phi\) [2410.08378]. In this framework, a latent “rank” \(u\) is mapped to a posterior draw \(\theta\), and the image of the reference law under the map is the posterior itself. The same paper also gives the inverse conditional rank map \[ R_{U\mid X=x}(\theta)=\nabla_\theta\psi^*(\theta,x), \] where \(\psi^*\) is the convex conjugate in the \(\theta\)-variable. This duality is central because it identifies quantile structure with OT geometry rather than with a Wasserstein-barycenter construction. Accordingly, the paper’s own characterization is that its geometry is **OT/Brenier/MK-depth geometry rather than an explicit Wasserstein-barycenter construction** [2410.08378]. A related but distinct plan-to-map construction appears in "Barycentric Projections of Optimal Transport Plans on Riemannian Manifolds." There the intrinsic barycentric projection assigns to each source point \(x\) the conditional Fréchet mean of its conditional destination law, \[ B_\pi(x)\in\argmin_{z\in M}\frac12\int_M d(z,y)^2\,d\pi(y\mid x), \] and is shown to be the best deterministic representative of a coupling under squared geodesic loss [2606.07926]. This is not a quantile map in the classical sense, but it is a canonical barycentric surrogate of a transport object. In Euclidean space it reduces to conditional expectation, while in Monge cases it recovers the actual transport map [2606.07926]. This suggests a broader interpretation in which barycentric quantile maps can also arise as deterministic representatives extracted from probabilistic transport couplings. ## 3. Barycentric transport maps in Wasserstein barycenter geometry "Measure Estimation in the Barycentric Coding Model" develops the barycentric side of the topic. It studies an unknown measure \(\mu_0\) assumed to lie in the set of Wasserstein-2 barycenters generated by known reference measures \(\{\mu_i\}_{i=1}^p\), with barycenter \[ \nu_\lambda = \arg\min_{\nu\in \mathcal P_{2,ac}(\mathbb R^d)} \frac12\sum_{i=1}^p \lambda_i W_2^2(\nu,\mu_i), \qquad \lambda\in\Delta^p, \] where the \(\lambda_i\) are barycentric coordinates [2201.12195]. The paper does not introduce quantile functions, but it identifies the nearest object to a barycentric quantile map as a **barycentric combination of optimal transport maps**. Let \(T_i\) denote the optimal map from \(\mu_0\) to \(\mu_i\). The gradient of the variance functional is \[ \nabla G_\lambda(\nu)= -\sum_{i=1}^p \lambda_i (T_i-\mathrm{Id}), \] so the barycenter condition is expressed as a balancing relation among displacement maps in the Wasserstein tangent space [2201.12195]. The associated Gram matrix \[ A_{ij}=\langle T_i-\mathrm{Id},\,T_j-\mathrm{Id}\rangle_{\mu_0} \] reduces coordinate recovery to the convex quadratic program \[ \min_{\lambda\in\Delta^p}\lambda^\top A\lambda. \] The strongest map formula appears under the paper’s **compatibility** assumption, where for all \(i,j,k\), \[ T_i^k = T_j^k\circ T_i^j. \] In that regime the barycenter admits the exact pushforward representation \[ \nu_\lambda = \left[\sum_{i=1}^p \lambda_i T_0^i\right]\#\mu_0. \] Equivalently, \[ T_\lambda=\sum_{i=1}^p \lambda_i T_0^i, \qquad T_\lambda-\mathrm{Id}=\sum_{i=1}^p \lambda_i (T_0^i-\mathrm{Id}). \] This is the paper’s closest explicit realization of a **barycentric map** [2201.12195]. The paper notes that **continuous measures on \(\mathbb R\)** are among the compatible classes, but it does **not** state the 1D formula \[ Q_{\nu_\lambda}(t)=\sum_i\lambda_i Q_{\mu_i}(t). \] That quantile representation is described only as an inference by standard OT theory, not as a theorem of the paper [2201.12195]. This distinction is important: the paper proves a barycentric transport-map characterization, not a quantile-function formalism. ## 4. HiMAP and explicit multivariate barycentric quantile maps "HiMAP: Hilbert Mass-Aligned Parameterization for Multivariate Barycenters and Frećhet Regression" provides the most direct realization of a barycentric quantile map. For a multivariate probability measure \(\mu\in\mathcal P_\infty(\mathbb R^d)\), the paper defines the **HiMAP quantile map** \[ \mathbf Q_\mu:[0,1]\to\mathbb R^d, \] designed so that the scalar parameter \(t\in[0,1]\) has a common, distribution-invariant “mass level” meaning across different measures [2603.03674]. The construction recursively partitions the support through **equiprobable conditional-median splits** arranged according to **Hilbert curve ordering rules**. Starting from a bounded axis-aligned box \(\mathcal M=\prod_{j=1}^d [L_{0,j},U_{0,j}]\), the split at depth \(l\) is the conditional median \[ q_l(\mu;B_{l-1}) := \inf\left\{x\in\mathbb R: \mu\big(\{\mathbf u\in B_{l-1}:u_{j_l}\le x\}\big)\ge \frac12 \mu(B_{l-1}) \right\}, \] so that each child cell has half the mass of its parent [2603.03674]. This makes the recursion **mass aligned** rather than geometrically uniform. For fixed depth \(L\), the finite-depth representative is \[ \mathbf H_{\mu,L}(t):= \big(q_{k_1(L)}(t),q_{k_2(L)}(t),\dots,q_{k_d(L)}(t)\big)^\top. \] Under the assumptions that \(\mu\) admits a density \(f\) on \(\mathcal M\) with \(00, \] with \[ \nu_\oplus = (\mathbf Q_{\nu_\oplus})_\# \mathrm{Unif}[0,1]. \] Within the cited literature, this is the clearest exact meaning of **barycentric quantile map** [2603.03674]. The same closure property makes HiMAP especially suitable for Fréchet regression with affine weights that sum to one but may be negative. The paper gives the population regression identity \[ \mathbf Q_{m_\oplus(\mathbf x)}(t) = \mathbb E\!\left[ w(\mathbf X,\mathbf x)\,\mathbf Q_Y(t) \right], \] and the empirical estimator \[ \widehat{\mathbf Q}_{m(\mathbf x)}(t) = \sum_{i=1}^m \widehat w_i(\mathbf x)\,\widehat{\mathbf Q}_{Y_i}(t) \] [2603.03674]. This generalizes the 1D algebra of quantile averaging to multivariate responses in an induced Hilbert geometry rather than in exact \(W_2\) geometry. ## 5. Bayesian posterior geometry and nested credible regions In "Deep Generative Quantile Bayes," the conditional vector quantile map is used as a posterior generator rather than as a barycenter operator. Training data consist of simulated triples \[ \{(\theta_i,X_i,U_i)\}_{i=1}^N, \] where \(\theta_i\sim\pi(\theta)\), \(X_i\sim L(\cdot\mid\theta_i)\), and \(U_i\sim F_U\) i.i.d. [2410.08378]. The learned map has domain \(u\in\mathcal U=S^d(1)\), conditioning variable \(x\in\mathcal X\), codomain \(\theta\in\Theta\subset\mathbb R^d\), and pushforward \[ U\sim F_U\mapsto \hat Q^N_{\theta\mid X=x}(U)\sim \hat\pi^N(\theta\mid X=x). \] A core modeling step is the assumption of a learned sufficient summary statistic \(f(X)\) satisfying \[ X\perp\theta\mid f(X), \] together with the affine-in-summary potential parameterization \[ \psi(u,x)=\varphi(u)+b(u)^\top f(x). \] The neural architecture uses **ICNNs** for \(\varphi\) and \(b\), with “3 hidden layers of width 512, and CELU activation,” and combines a **DeepSet** \(h_1\) with an **LSTM** \(h_2\) to define \[ f(X)=[h_1(X),h_2(X)]\in\mathbb R^q \] [2410.08378]. The paper’s central optimization problem is the empirical dual OT loss \[ \mathcal L_1(\varphi,b,f\mid \mathbf X,\mathbf\theta,\mathbf U) = \sum_{i=1}^N \left( \varphi(U_i)+ \max_{j\in\{1,\dots,N\}} \{U_j^\top \theta_i-\varphi(U_j)-b(U_j)^\top f(X_i)\} \right). \] The paper’s most geometric contribution is its use of **Monge–Kantorovich depth** for credible sets. For an unconditional distribution \(P\), the MK-depth region of content \(\tau\in(0,1)\) is \[ Q_P(S^d(\tau)), \] where \(S^d(\tau)\) is the ball of radius \(\tau\), yielding nested regions \[ Q_P(S^d(\tau))\subset Q_P(S^d(\tau')),\qquad \tau\le\tau'. \] In the posterior setting this becomes \[ C_\tau(\theta\mid X=x):=Q_{\theta\mid X=x}(S^d(\tau)), \qquad \hat C^N_\tau(\theta\mid X=x):=\hat Q^N_{\theta\mid X=x}(S^d(\tau)). \] Thus posterior credible sets are direct images of concentric latent balls under the conditional Brenier map [2410.08378]. The paper proves uniform consistency of the estimated potential and conjugate, consistency of the vector quantile map, \[ \sup_{u\in K_U,\;x\in K_X} \|\hat Q^N_{\theta\mid X=x}(u)-Q_{\theta\mid X=x}(u)\| \to 0, \] consistency of the recovered posterior, \[ \sup_{x\in K_X} W_2\big(\hat\pi^N(\theta\mid X=x),\pi(\theta\mid X=x)\big)\to 0, \] and Hausdorff convergence of credible sets, \[ \sup_{x\in K_X} d_H\big(\hat C^N_\tau(\theta\mid X=x),C_\tau(\theta\mid X=x)\big)\to 0 \] [2410.08378]. The Gaussian conjugate experiment also displays **support shrinkage**, meaning that posterior contours contract as sample size \(n\) increases. In this setting, the “barycentric” appearance comes from centered latent balls being transported into posterior contours, but the paper is explicit that this is not a Wasserstein-barycenter construction [2410.08378]. ## 6. Related abstractions, limitations, and common misconceptions Several nearby theories clarify what a barycentric quantile map is **not**. "Partitions of Unity and Barycentric Algebras" studies convex combinations on polytopes through barycentric algebras and the **tautological map** \[ (Tf)(a)=\sum_{i=1}^n f_i(a)v_i, \] from partitions of unity to points in a convex polytope [2501.00937]. This is a canonical barycentric coordinate map, but the paper explicitly does **not** define a probabilistic quantile map. Its relevance is algebraic: it shows how weight vectors or simplex-valued functions can be converted into points by barycentric combination. "Convergence theorems for barycentric maps" studies a different abstraction: a **contractive barycentric map** \[ \beta:\mathcal P^p(M)\to M \] satisfying \[ \beta(\delta_x)=x,\qquad d(\beta(\mu),\beta(\nu))\le d_p^W(\mu,\nu), \] and defines nonlinear conditional expectation by \[ E^\beta_{\mathcal B}(\varphi)(\omega)=\beta(\varphi_*\mathbf P_\omega) \] [1805.08558]. This again supplies a barycentric summary of measures, together with martingale, ergodic, continuity, and large deviation theory, but it does not define quantile maps or transport-based rank parameterizations. Several misconceptions therefore need to be separated. | Misconception | Correction | |---|---| | A barycentric quantile map is a standardized OT term | The cited papers do not present a single standardized definition | | It always means a Wasserstein barycenter formula | In [2410.08378] the relevant object is a Brenier quantile transport, not a barycenter | | It is always an exact multivariate analogue of 1D quantiles | HiMAP provides such an analogue only in an induced HiMAP geometry, not in exact \(W_2\) geometry [2603.03674] | | Any barycentric projection is a quantile map | Barycentric projection is a deterministic surrogate of a coupling, not a classical quantile function [2606.07926] | The most accurate synthesis is therefore plural. In OT-based Bayesian computation, the nearest object is the **conditional vector quantile map** \(Q_{\theta\mid X=x}\) [2410.08378]. In Wasserstein barycenter geometry, the nearest object is the **convex combination of optimal transport maps** \(T_\lambda=\sum_i\lambda_iT_0^i\) under compatibility [2201.12195]. In multivariate quantile-like barycenter calculus, the most explicit object is the **HiMAP quantile map** \(\mathbf Q_\mu\), whose pointwise affine averages define barycenters and Fréchet regression estimators [2603.03674]. Taken together, these works indicate that “barycentric quantile map” is best understood as a family of constructions linking common mass coordinates, transport maps, and barycentric averaging, with the exact meaning determined by the ambient geometry and the target problem. Source: https://www.emergentmind.com/topics/barycentric-quantile-map