pith. sign in

arxiv: 2507.11719 · v2 · pith:4GFWTOXTnew · submitted 2025-07-15 · 📊 stat.ME · stat.CO· stat.ML

Barycentric model aggregation in the Wasserstein space of distributions and a variational approach to consistency

Emmanouil Androulakis , Georgios I. Papayiannis , Athanasios N. Yannacopoulos This is my paper

Pith reviewed 2026-05-22 00:33 UTC · model grok-4.3

classification 📊 stat.ME stat.COstat.ML
keywords Wasserstein barycentermodel aggregationGamma-convergenceconsistencyprobability measuresvariational methodsdata-driven calibration
0
0 comments X

The pith

A Gamma-convergence argument shows data-driven Wasserstein barycenter weights converge to the true optimum for fixed candidate models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a framework for aggregating a fixed finite collection of probability models on the real line by learning weights that form their Wasserstein barycenter and match an observed target distribution. The weights are calibrated by minimizing a discrepancy functional on empirical samples from the target. A variational analysis based on Gamma-convergence then proves that these empirical weights converge to the weights that would be optimal for the true target distribution, and that the resulting barycentric estimator is consistent. This matters because it supplies a statistically reliable way to combine multiple models when the target is only observed through samples, as occurs in sensor networks or forecasting tasks.

Core claim

For a fixed finite collection of candidate probability measures on the real line, the weights minimizing the Wasserstein distance from the barycenter to the empirical target measure converge to the weights minimizing the distance to the true target. The associated barycentric estimators therefore converge to the population barycenter, with the entire argument resting on a Gamma-convergence analysis of the variational problem under mild conditions.

What carries the argument

The variational problem that selects aggregation weights to minimize the Wasserstein distance between the weighted barycenter of the candidates and the target measure.

Load-bearing premise

The target and candidate measures live on the real line, the collection of candidates is fixed and finite, and their barycenters are well-defined so that the Gamma-convergence argument applies.

What would settle it

Numerical experiments or real data in which the learned weights fail to approach the theoretically optimal weights as the number of samples from the target grows without bound would show the claimed consistency does not hold.

Figures

Figures reproduced from arXiv: 2507.11719 by Athanasios N. Yannacopoulos, Emmanouil Androulakis, Georgios I. Papayiannis.

Figure 1
Figure 1. Figure 1: Illustration per year of the observed distributional characteristics (gray line) and the [PITH_FULL_IMAGE:figures/full_fig_p014_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the calibrated upper tail ES from the pure barycenter (black dashed [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the 1-year ahead predictions for the upper tail ES by the pure barycenter [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of the calibrations for the upper tail Value at Risk by the pure barycenter [PITH_FULL_IMAGE:figures/full_fig_p025_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of the 1-year ahead predictions of the upper tail Value at Risk from the [PITH_FULL_IMAGE:figures/full_fig_p026_5.png] view at source ↗
read the original abstract

We study the problem of model aggregation within the Wasserstein space for probability measures on the real line. Given a fixed finite collection of candidate probability models, we consider the associated class of Wasserstein barycenters and develop a data-driven calibration framework in which the aggregation weights are statistically learned from empirical information associated with a target distribution. From a variational perspective based on $\Gamma$-convergence, we establish consistency of the resulting aggregation scheme, showing that empirical minimizers converge to the minimizers of the actual problem, along with the associated barycentric estimators, under mild conditions. The performance of the proposed method is evaluated through synthetic experiments and illustrated on a real dataset from a temperature monitoring network of sensors.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper develops a data-driven aggregation scheme for a fixed finite collection of candidate probability models on the real line, using Wasserstein barycenters. Aggregation weights are calibrated by minimizing an empirical variational objective based on the 2-Wasserstein distance between the barycenter and the target empirical measure. Consistency of the resulting empirical minimizers (and thus of the barycentric estimators) is proved via Γ-convergence under mild conditions. The approach is illustrated on synthetic experiments and a real temperature-monitoring sensor dataset.

Significance. If the Γ-convergence argument holds under the stated conditions, the work supplies a variational consistency guarantee for Wasserstein-space model aggregation that is not available from direct empirical-risk arguments. The combination of a Polish-space setting (P_2(R)), explicit barycenter map, and both synthetic and real-data validation gives the result practical as well as theoretical weight; the framework could extend to other distributional aggregation tasks in statistics and environmental science.

major comments (2)
  1. [Abstract / main consistency theorem] The central consistency claim rests on Γ-convergence of the empirical functional F_n(λ) = W_2(B(λ, μ̂_n), μ̂_n) to F(λ) = W_2(B(λ, μ), μ). The abstract invokes only “mild conditions,” yet the argument requires (i) uniform integrability of second moments to guarantee μ̂_n → μ in W_2 and (ii) continuity of the barycenter map λ ↦ B(λ,·) in the W_2 topology. Without an explicit statement of these hypotheses (or a proof that they follow from the finite-collection assumption), it is impossible to verify that the Γ-limit is attained at the same point as the population problem.
  2. [Consistency section (presumably §3)] The manuscript states that the target distribution belongs to P_2(R) and that the candidate models form a fixed finite collection whose barycenters are well-defined, but does not indicate whether the proof supplies the required compactness or moment control when the target may have arbitrarily heavy tails still compatible with finite second moment. This leaves open whether the empirical barycenters converge in the topology needed for the variational limit.
minor comments (2)
  1. [Method section] Notation for the barycenter map B(λ, μ) should be introduced once and used consistently; the current text occasionally switches between “barycentric estimator” and “aggregated distribution” without cross-reference.
  2. [Numerical experiments] The real-data experiment (temperature sensor network) would benefit from a brief description of the number of sensors, time resolution, and how the empirical target measure is constructed from the raw readings.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thorough review and valuable comments on our manuscript. We address the major comments below and will incorporate clarifications in the revised version to strengthen the presentation of the consistency results.

read point-by-point responses

  1. Referee: [Abstract / main consistency theorem] The central consistency claim rests on Γ-convergence of the empirical functional F_n(λ) = W_2(B(λ, μ̂_n), μ̂_n) to F(λ) = W_2(B(λ, μ), μ). The abstract invokes only “mild conditions,” yet the argument requires (i) uniform integrability of second moments to guarantee μ̂_n → μ in W_2 and (ii) continuity of the barycenter map λ ↦ B(λ,·) in the W_2 topology. Without an explicit statement of these hypotheses (or a proof that they follow from the finite-collection assumption), it is impossible to verify that the Γ-limit is attained at the same point as the population problem.

    Authors: We appreciate this observation. In the revised manuscript, we will explicitly list the mild conditions, including the uniform integrability of second moments which follows from μ ∈ P_2(R) and ensures W_2 convergence of the empirical measures. We will also add a proof that the barycenter map is continuous on the simplex due to the finite fixed collection of models. This ensures the Γ-convergence holds and the minimizers coincide. revision: yes

  2. Referee: [Consistency section (presumably §3)] The manuscript states that the target distribution belongs to P_2(R) and that the candidate models form a fixed finite collection whose barycenters are well-defined, but does not indicate whether the proof supplies the required compactness or moment control when the target may have arbitrarily heavy tails still compatible with finite second moment. This leaves open whether the empirical barycenters converge in the topology needed for the variational limit.

    Authors: We agree that this aspect requires more explicit discussion. The proof relies on the fact that finite second moments provide sufficient compactness in P_2(R) via Prokhorov's theorem adapted to the Wasserstein metric. For distributions with heavy tails but finite variance, the empirical convergence in W_2 still holds, and the variational argument applies without additional moment assumptions. We will insert a clarifying paragraph in the consistency section. revision: yes

Circularity Check

0 steps flagged

No circularity: consistency established via standard Γ-convergence on Polish space P_2(R)

full rationale

The paper defines a variational problem for learning aggregation weights λ from empirical measures μ̂_n associated with a target μ in the Wasserstein space on the real line, then invokes Γ-convergence to show that empirical minimizers converge to population minimizers (and thus the barycentric estimators converge). This is a standard variational limit argument relying on the Polish property of P_2(R) and unspecified mild conditions for compactness and moment control; it does not reduce any claimed prediction or estimator to a fitted quantity by construction, nor does it rest on self-citations for uniqueness theorems or ansatzes. The derivation chain remains independent of the specific data realizations used for evaluation and is self-contained against external results in optimal transport.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

Based on the abstract alone, the central claim rests on standard properties of the Wasserstein metric on the real line, the existence of barycenters for finite collections of measures, and the applicability of Gamma-convergence theory; no new entities are introduced.

free parameters (1)
  • aggregation weights
    The weights are statistically learned from empirical information associated with the target distribution and therefore constitute data-dependent parameters.
axioms (2)
  • domain assumption Probability measures live in the Wasserstein space on the real line and barycenters exist for any finite convex combination of them.
    Invoked when the class of Wasserstein barycenters is defined for the candidate models.
  • standard math Gamma-convergence applies to the empirical objective and yields convergence of minimizers under mild conditions.
    Central tool used to establish consistency of the aggregation scheme.

pith-pipeline@v0.9.0 · 5665 in / 1445 out tokens · 47115 ms · 2026-05-22T00:33:23.042089+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

46 extracted references · 46 canonical work pages

  1. [1]

    Agueh, M. and G. Carlier (2011). Barycenters in the W asserstein space. SIAM Journal on Mathematical Analysis\/ 43\/ (2), 904--924

    work page 2011

  2. [2]

    Avanzi, B., Y. Li, B. Wong, and A. Xian (2024). Ensemble distributional forecasting for insurance loss reserving. Scandinavian Actuarial Journal\/ 2024\/ (9), 971--1012

    work page 2024

  3. [3]

    Bishop, A. N. and A. Doucet (2021). Network consensus in the W asserstein metric space of probability measures. SIAM Journal on Control and Optimization\/ 59\/ (5), 3261--3277

    work page 2021

  4. [4]

    Lin, and H.-G

    Chen, Y., Z. Lin, and H.-G. M \"u ller (2023). Wasserstein regression. Journal of the American Statistical Association\/ 118\/ (542), 869--882

    work page 2023

  5. [5]

    Gillingham, and W

    Christensen, P., K. Gillingham, and W. Nordhaus (2018). Uncertainty in forecasts of long-run economic growth. Proceedings of the National Academy of Sciences\/ 115\/ (21), 5409--5414

    work page 2018

  6. [6]

    Dal Maso, G. (2012). An introduction to -convergence , Volume 8. Springer Science & Business Media

    work page 2012

  7. [7]

    and H.-G

    Dubey, P. and H.-G. M \"u ller (2019). Fr \'e chet analysis of variance for random objects. Biometrika\/ 106\/ (4), 803--821

    work page 2019

  8. [8]

    Dunlop, M. M., D. Slep c ev, A. M. Stuart, and M. Thorpe (2020). Large data and zero noise limits of graph-based semi-supervised learning algorithms. Applied and Computational Harmonic Analysis\/ 49\/ (2), 655--697

    work page 2020

  9. [9]

    Fan, J. and R. Li (2001). Variable selection via nonconcave penalized likelihood and its oracle properties. Journal of the American Statistical Association\/ 96\/ (456), 1348--1360

    work page 2001

  10. [10]

    Fr \'e chet, M. (1948). Les \'e l \'e ments al \'e atoires de nature quelconque dans un espace distanci \'e . In Annales de l'institut Henri Poincar \'e , Volume 10, pp.\ 215--310

    work page 1948

  11. [11]

    Friederichs, P. and T. L. Thorarinsdottir (2012). Forecast verification for extreme value distributions with an application to probabilistic peak wind prediction. Environmetrics\/ 23\/ (7), 579--594

    work page 2012

  12. [12]

    a is \"a nen, K. Takahashi, E. Ter \

    Fronzek, S., Y. Honda, A. Ito, J. P. Nunes, N. Pirttioja, J. R \"a is \"a nen, K. Takahashi, E. Ter \"a m \"a , M. Yoshikawa, and T. R. Carter (2022). Estimating impact likelihoods from probabilistic projections of climate and socio-economic change using impact response surfaces. Climate Risk Management\/ 38 , 100466

    work page 2022

  13. [13]

    Gilat, D. and T. P. Hill (1992). One-sided refinements of the strong law of large numbers and the G livenko- C antelli theorem. The Annals of Probability\/ , 1213--1221

    work page 1992

  14. [14]

    Gneiting, T. and R. Ranjan (2013). Combining predictive distributions. Electronic Journal of Statistics\/ 7 , 1747--1782

    work page 2013

  15. [15]

    Hansen, B. E. (2007). Least squares model averaging. Econometrica\/ 75\/ (4), 1175--1189

    work page 2007

  16. [16]

    Hartman, B. M. and C. Groendyke (2013). Model selection and averaging in financial risk management. North American Actuarial Journal\/ 17\/ (3), 216--228

    work page 2013

  17. [17]

    Hoerl, A. E. and R. W. Kennard (1970). Ridge regression: Biased estimation for nonorthogonal problems. Technometrics\/ 12\/ (1), 55--67

    work page 1970

  18. [18]

    Hong, L. and R. Martin (2017). A flexible B ayesian nonparametric model for predicting future insurance claims. North American Actuarial Journal\/ 21\/ (2), 228--241

    work page 2017

  19. [19]

    Hsiang, S., R. Kopp, A. Jina, J. Rising, M. Delgado, S. Mohan, D. J. Rasmussen, R. Muir-Wood, P. Wilson, M. Oppenheimer, et al. (2017). Estimating economic damage from climate change in the U nited S tates. Science\/ 356\/ (6345), 1362--1369

    work page 2017

  20. [20]

    Kim, Y.-H. and B. Pass (2017). Wasserstein barycenters over R iemannian manifolds. Advances in Mathematics\/ 307 , 640--683

    work page 2017

  21. [21]

    Koundouri, P., G. I. Papayiannis, A. Vassilopoulos, and A. N. Yannacopoulos (2024). Probabilistic scenario-based assessment of national food security risks with application to E gypt and E thiopia. Journal of the Royal Statistical Society Series A: Statistics in Society\/ , qnae046

    work page 2024

  22. [22]

    Kravvaritis, D. C. and A. N. Yannacopoulos (2020). Variational methods in nonlinear analysis: with applications in optimization and partial differential equations . Walter De Gruyter Gmbh & Co Kg

    work page 2020

  23. [23]

    Spokoiny, and A

    Kroshnin, A., V. Spokoiny, and A. Suvorikova (2021). Statistical inference for B ures-- W asserstein barycenters. The Annals of Applied Probability\/ 31\/ (3), 1264--1298

    work page 2021

  24. [24]

    and J.-M

    Le Gouic, T. and J.-M. Loubes (2017). Existence and consistency of W asserstein barycenters. Probability Theory and Related Fields\/ 168 , 901--917

    work page 2017

  25. [25]

    Lu, X. and L. Su (2015). Jackknife model averaging for quantile regressions. Journal of Econometrics\/ 188\/ (1), 40--58

    work page 2015

  26. [26]

    Miljkovic, T. and B. Gr \"u n (2021). Using model averaging to determine suitable risk measure estimates. North American Actuarial Journal\/ 25\/ (4), 562--579

    work page 2021

  27. [27]

    Moral-Benito, E. (2015). Model averaging in economics: An overview. Journal of Economic Surveys\/ 29\/ (1), 46--75

    work page 2015

  28. [28]

    G \"u neralp, B

    Muis, S., B. G \"u neralp, B. Jongman, J. C. Aerts, and P. J. Ward (2015). Flood risk and adaptation strategies under climate change and urban expansion: A probabilistic analysis using global data. Science of the Total Environment\/ 538 , 445--457

    work page 2015

  29. [29]

    Panaretos, V. M. and Y. Zemel (2020). An invitation to statistics in W asserstein space . Springer Nature

    work page 2020

  30. [30]

    Galanis, and A

    Papayiannis, G., G. Galanis, and A. Yannacopoulos (2018). Model aggregation using optimal transport and applications in wind speed forecasting. Environmetrics\/ 29\/ (8), e2531

    work page 2018

  31. [31]

    Papayiannis, G. I. and A. N. Yannacopoulos (2018). A learning algorithm for source aggregation. Mathematical Methods in the Applied Sciences\/ 41\/ (3), 1033--1039

    work page 2018

  32. [32]

    Li, and Y

    Peng, J., Y. Li, and Y. Yang (2024). On optimality of M allows model averaging. Journal of the American Statistical Association\/ , 1--12

    work page 2024

  33. [33]

    and H.-G

    Petersen, A. and H.-G. M \"u ller (2019a). Fréchet regression for random objects with E uclidean predictors. The Annals of Statistics\/ 47\/ (2), 691--719

    work page

  34. [34]

    and H.-G

    Petersen, A. and H.-G. M \"u ller (2019b). Wasserstein covariance for multiple random densities. Biometrika\/ 106\/ (2), 339--351

    work page

  35. [35]

    Ranjan, R. and T. Gneiting (2010). Combining probability forecasts. Journal of the Royal Statistical Society Series B: Statistical Methodology\/ 72\/ (1), 71--91

    work page 2010

  36. [36]

    Santambrogio, F. (2015). Optimal transport for applied mathematicians. Birk \"a user, NY\/ 55\/ (58-63), 94

    work page 2015

  37. [37]

    Scheuerer, M. and D. M \"o ller (2015). Probabilistic wind speed forecasting on a grid based on ensemble model output statistics. The Annals of Applied Statistics\/ , 1328--1349

    work page 2015

  38. [38]

    Steel, M. F. (2020). Model averaging and its use in economics. Journal of Economic Literature\/ 58\/ (3), 644--719

    work page 2020

  39. [39]

    Theil, A

    Thorpe, M., F. Theil, A. M. Johansen, and N. Cade (2015). Convergence of the K -means minimization problem using -convergence. SIAM Journal on Applied Mathematics\/ 75\/ (6), 2444--2474

    work page 2015

  40. [40]

    Tucker, D. C., Y. Wu, and H.-G. M \"u ller (2023). Variable selection for global F r \'e chet regression. Journal of the American Statistical Association\/ 118\/ (542), 1023--1037

    work page 2023

  41. [41]

    Villani, C. (2021). Topics in optimal transportation , Volume 58. American Mathematical Soc

    work page 2021

  42. [42]

    Villani, C. et al. (2009). Optimal transport: old and new , Volume 338. Springer

    work page 2009

  43. [43]

    Wan, A. T., X. Zhang, and G. Zou (2010). Least squares model averaging by M allows criterion. Journal of Econometrics\/ 156\/ (2), 277--283

    work page 2010

  44. [44]

    Zemel, Y. and V. M. Panaretos (2019). Fr \'e chet means and P rocrustes analysis in W asserstein space. Bernoulli\/ 25\/ (2), 932--976

    work page 2019

  45. [45]

    and H.-G

    Zhou, Y. and H.-G. M \"u ller (2024). Wasserstein regression with empirical measures and density estimation for sparse data. Biometrics\/ 80\/ (4), ujae127

    work page 2024

  46. [46]

    Zou, H. and T. Hastie (2005). Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society Series B: Statistical Methodology\/ 67\/ (2), 301--320

    work page 2005