The Monge optimal transport barycenter problem

Andrew D. Lipnick; Esteban G. Tabak; Giulio Trigila; Wenjun Zhao; Xuancheng Ye; Yating Wang

REVIEW 3 major objections 6 minor 1 cited by

The Monge optimal transport barycenter problem

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper shows that the data-driven Monge optimal transport barycenter problem can be solved by a pure gradient-descent minimization whose independence penalty is the first singular value of a low-rank matrix, and that the resulting…

desk verdict A genuinely useful algorithmic reduction to a pure minimization, held back by an unproved closed-form inverse that every downstream application leans on. read the letter →

arxiv 2507.03669 v3 pith:4TSXTWEW submitted 2025-07-04 math.OC stat.ME

Andrew D. Lipnick , Esteban G. Tabak , Giulio Trigila , Yating Wang , Xuancheng Ye , Wenjun Zhao This is my paper

classification math.OCstat.ME MSC 49Q2262G07

keywords MongeoptimaltransportbarycenterindependenceviatestfunctionsfirstsingularvaluepenaltyconditionaldensityestimationBayesianinferencedata-drivenflowsfactordiscovery

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that the data-driven Monge optimal transport barycenter problem can be solved as a pure minimization over the samples $y_i$: minimize transport cost plus a penalty for dependence between $y$ and the factors $z$. The dependence penalty is cast adversarially, but the best adversarial functions can be found in closed form as the first principal components of a small matrix, so no saddle-point optimization is needed. If correct, this yields an algorithm whose per-iteration cost is linear in sample size and outcome dimension, and an explicit inverse map that makes conditional density simulation, density estimation, and model-free Bayesian inference practical.

What carries the argument

The machinery is an adversarial independence condition: $y$ and $z$ are taken to be independent when $E[g(y)f(z)]=0$ for all $f$ in $F$ and $g$ in $G$, where $F$ and $G$ are finite-dimensional inner-product spaces spanned by data-adapted kernel-like features. Representing $f=Q_z a$ and $g=Q_y b$ with orthogonal $Q$'s reduces the maximization to the first singular vectors of the $n_z \times n_y$ matrix $A_{hl}=\sum_i Q_z^h(z_i)Q_y^l(y_i)$. Because $\sigma_j=a_j'Ab_j$, Lemma 4 makes the derivative of the penalty exact at fixed $(a,b)$, so gradient descent on $y$ descends the true objective; Lemma 3 keeps $Q_y$ orthogonal under evolution, and the stationary condition at the samples is rearranged into the closed-form inverse (10), whose $y$-derivatives give conditional densities.

What would settle it

On a generative model with known $\rho(x|z)$, hold out a random subset of pairs, fit on the rest, then check two things: whether $X(T(x_i,z_i),z_i)$ recovers $x_i$ on held-out pairs, and whether simulated $x^*$ at a held-out $z^*$ passes a goodness-of-fit test against the true conditional. A percentage of reconstruction failures far above the training error would refute the closed-form inversion claim.

Watch

Extended reading notes

Core claim

The central discovery is that relaxing the pushforward/independence condition to uncorrelation over finite-dimensional inner-product spaces turns the barycenter problem into a minimization with an exact adversarial penalty: the largest singular value of $A(y)=Q_z'Q_y$, and later a weighted sum of the first $J$ squared singular values. The map $T$ is then obtained by gradient-descent flows on $y$, initialized at $x$. The paper further derives a closed-form expression for the inverse map, $X(y,z)=y+2n\sum_j \lambda_j \sigma_j f_j(z)\nabla g_j(y)$, which is smooth in $(y,z)$ and reproduces the training pairs; this inverse powers simulation of $\rho(x|z^*)$, density estimation via the change-of-variable formula, and Bayesian posterior updates.

Load-bearing premise

The load-bearing premise is that the formula used to invert the transportation map, which the algorithm derives from the optimality condition at the $n$ observed pairs, also holds at any new value of the factors $z$; if it fails there, the simulated conditional samples, density estimates, and Bayesian updates built on it are not reliable.

Editorial extensions

If this is right

The independence constraint becomes a pure minimization, eliminating saddle-point alternation with its convergence problems.
Per-iteration cost scales linearly in the number of samples $n$ and in the outcome dimension, in contrast to quadratic-cost kernel-centered approaches.
A single run yields, for any target $z^*$, simulated samples of $\rho(x|z^*)$ and an estimate of the conditional density through the Jacobian of the inverse map.
The regular Monge optimal transport problem is recovered as the binary-$z$ case, with the map $Q(x_0^i)$ given directly.
The functions $f_j(z)$ extracted by the inversion formula serve as discovered factors describing how $x$ depends on $z$.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The closed-form inverse (10) is an extrapolation beyond the $n$ stationary points: stationarity is enforced only at the training pairs, no error bound is given, so the inversion's validity for new $z$ is the part of the claim I would test first.
The independence guarantee holds only within the chosen spaces $F$ and $G$, so dependence invisible to those spaces will survive; successive runs with richer spaces act like boosting and should be treated as part of the method, not an optional add-on.
The accuracy of the whole pipeline is likely governed by the bandwidths and centers of the kernel features; a natural testable extension is to replace manual bandwidth choices by held-out likelihood cross-validation, as the paper already does for the sea-surface-temperature forecast.
If the out-of-sample inverse holds, the same flow could provide a checkable generative model for high-stakes simulation tasks, where a permutation test on residuals could be reported alongside the conditional samples.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Bottom line: the paper's main algorithmic idea is real and worth engaging with—replacing the adversarial independence constraint with a first-singular-value penalty and then solving a pure minimization—but the closed-form inverse in Eq. (10) is an extrapolation from n training points to arbitrary (y,z), and the paper simply asserts its validity. The stress-test note lands. Everything in Sections 8 and 9 inherits that gap.

What is actually good: Lemma 4 is clean and correct, and the PCA reduction does turn a saddle-point problem into pure minimization. That is a practical improvement over the authors' earlier kernel and network flows. The factor extraction through the f_j(z) is a nice byproduct. The experiments are broad—conditional density estimation, Bayesian inference, hidden-signal recovery, SST forecasting—and the synthetic checks (z-tests, KL and MSE against known densities) are appropriate. The complexity measurements support roughly linear scaling in n, with the usual caveat that iteration counts are empirical. Heavy self-citation is not itself a problem here; the prior work is genuinely theirs and the new elements are identifiable.

The soft spots, in order of importance. First, Eq. (10) is load-bearing. The stationarity condition holds at the training pairs, and the paper says "we extend the validity of this expression" without proof or error bound. For an out-of-sample y, A is built from the training y_i and does not change, so the claimed gradient of the singular value is an ansatz, not a derivative of the object that appeared in the optimization. The termination criterion asking for small gradient is sensible, but small gradient at training points does not establish that the extended formula is a true inverse elsewhere. Second, the O(n) claim depends on the number of iterations being independent of n, which is observed in experiments rather than shown. Third, the real-data demonstrations have no baselines; they show plausible behavior, not predictive skill against alternatives. These are addressable rather than fatal.

Who this is for: researchers working on conditional optimal transport, barycenters, or factor-adjusted simulation and density estimation. It deserves a serious referee; I would send it out and ask the authors to close the inverse gap with either a theorem or a held-out validation of the inverse formula on data not used to fit it.

Referee Report

3 major / 6 minor

Summary. The paper develops a data-driven algorithm for the Monge optimal-transport barycenter problem, in which the pushforward condition is reformulated as statistical independence between y=T(x,z) and z. Independence is relaxed to vanishing correlation over finite-dimensional function spaces F and G, reducing the adversarial formulation to a pure minimization of the transport cost plus the first J squared singular values of a low-rank matrix A(y). The authors derive a closed-form expression for the inverse map, X(y,z)=y+2n Σ_j λ_j σ_j f_j(z)∇g_j(y) (Eq. 10), propose a stagewise gradient-descent flow, and apply the framework to the Monge optimal transport problem, conditional density estimation, Bayesian inference, and weather/climate forecasting.

Significance. The conceptual core of the paper is attractive and partially sound: the reduction of the adversarial independence formulation to a minimization over singular values is elegant, and Lemma 4 on derivatives of singular values is correct. If the out-of-sample inversion formula were rigorously justified, the paper would offer an efficient per-iteration O(n) method for barycenter problems with continuous covariates, with a useful byproduct of extracted factors. The numerical experiments are extensive and suggestive. However, the closed-form inverse is the linchpin of the conditional simulation, density estimation, and Bayesian inference claims; until that step is proved or bounded, those claims remain conditional and are not established by the numerical examples alone.

major comments (3)

[§5, Eq. (10)] Equation (10) is an extrapolation, not a derivation. The stationarity condition ∇_{y_i}L=0 used to write x_i = y_i + 2n Σ_j λ_j σ_j ∇_{y_i}σ_j holds only at the n training pairs (y_i,z_i). For a new pair (y,z), the quantity σ_j is not defined as a function of the new y in the way the formula requires, since A(y) is built from the training points; the replacement ∇_y σ_j|_{y,z} = f_j(z)∇g_j(y) is an ansatz. No proof or error bound is given that T(X(y,z),z)=y, nor that X(y,z) from Eq. (10) is the inverse of the map produced by the optimization. Because Section 8 and all of Section 9 (conditional density estimation, Bayesian inference, climate reconstruction) evaluate X at off-sample z and at arbitrary y, these applications inherit this gap. The authors should either prove a consistency result under stated assumptions, provide a quantitative error bound, or substantially weaken the claims and validate the inversion against held-out data.
[§6.6, complexity] The advertised linear-complexity result counts only the work per gradient-descent iteration. The statement that 'for a fixed problem, the number of iterations should not depend on n' is an empirical observation, not a proved scaling law. Without control on the number of iterations or on the growth of m_y and m_z, the total cost of the algorithm is not shown to be linear in n. The comparison with O(n^2) methods in the Introduction and Section 6.6 should be restated as a per-iteration comparison, or the iteration-count scaling should be supported by a theoretical argument or by systematic experiments across a range of n.
[§8, density estimation] The conditional density formula ρ(X(y,z)|z) = |∇_y X(y,z)|^{-1} μ(y) is written without the absolute determinant of the Jacobian, which is required when d_x>1. More importantly, the density estimator depends on ∂X/∂y from Eq. (11), which is derived from the unproved inversion formula Eq. (10). Consequently, the density-estimation and Bayesian results in Sections 8 and 9 do not constitute independent evidence for the validity of Eq. (10); they are downstream applications of the same extrapolation.

minor comments (6)

[§6.4] In the definition of the kernel-based feature spaces, the second line reads 'G_j(z) = K_y(y, y_j^c)'; this should be 'G_j(y) = K_y(y, y_j^c)'.
[§6.5] There is a typo: 'prrvious run' should be 'previous run'.
[§9.5] The subsection title 'Lotka-Volterra pray-predator models' should read 'prey-predator models'.
[References] Several references are duplicated: [4] and [49] are the same paper, [25] and [46] are the same paper, and [21] and [48] are the same paper.
[Algorithm 1, line 14] The stage-termination condition 'Done with stage = ∥y−y0∥2 > 1/10 ∥y0 − y0∥2' is ambiguous because the right-hand side appears to be zero; it should presumably involve the variance of the y0 samples or a similar reference quantity.
[§8] The determinant notation should be made explicit: in the multivariate case, the change-of-variable formula requires |det ∇_y X(y,z)|, not the scalar |∇_y X(y,z)|.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the core derivation is self-contained, and the closed-form inverse (10) is an explicitly labeled extrapolation rather than a circular prediction.

full rationale

The paper's derivation does not assume its target results. The independence condition is introduced through standard test-function characterizations, and the reduction of the adversarial maximization to the first singular values of A(y) is exact algebra (Lemma 4), not an imported conclusion. The minimization objective is optimized without reference to the conditional densities later estimated. The inversion formula (10) is derived from the stationarity condition at the training pairs and therefore reproduces x_i at (y_i,z_i) by construction; the paper states "we extend the validity of this expression" to arbitrary (y,z), which is an unproved extrapolation rather than a circular step, since it does not presuppose the simulation or density-estimation outputs it is used to produce. Numerical sections validate the method against known synthetic distributions and external data, so the central derivation is self-contained. Self-citations are frequent in the literature review but are not the load-bearing justification for any central claim; cited prior work is not used to forbid alternatives or to define away the problem.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several tuned hyperparameters (λ_j, bandwidths, number of centers, J_max, thresholds) and on four substantive assumptions: smoothness and invertibility of conditional distributions, map-supported optimal couplings (cited), richness of feature spaces, and out-of-sample validity of the closed-form inverse. The last two are acknowledged in the text but not proven. No new physical entities are introduced.

free parameters (6)

Penalization coefficients λ_j = Adaptive, updated by rule in Section 6.3 (e.g., λ_l = γ - min(c_uv, 0))
Balance transport cost against independence penalty; fitted to drive σ_j below σ*.
Bandwidth scale γ_z = Cross-validated over [0.2,20] in SST example; set by scale in temperature example
Controls the resolution of covariate dependence captured by kernel features in F and G; a user-chosen free parameter.
Number of kernel centers m = m = min(ceil(sqrt(n)), m_max)
Sets the richness of feature spaces F and G; the user provides m_max.
J_max = user-set
Truncation index for the singular-value penalty in (9); affects which dependence modes are penalized.
Target correlation σ* = 0.2 / sqrt(n)
Termination threshold for remaining dependence; chosen heuristically from the sampling standard deviation of correlations.
Gradient and stage tolerances α and δ = 0.0025 and 0.1
Termination criteria in Section 6.2; chosen by hand.

assumptions (5)

domain assumption ρ(x|z) is absolutely continuous on X and vanishes on small subsets for all z.
Section 2; needed for existence of Monge solutions and invertibility of optimal maps (Lemma 2).
standard math The optimal couplings in the extended Wasserstein barycenter are supported on maps.
Lemma 1 relies on this, citing [6, Theorem 5.1] and [8, Corollary 3.3.3].
ad hoc to paper The chosen feature spaces F and G are rich enough to approximate delta functions, so vanishing correlation over F and G is equivalent to independence.
Section 3; the equivalence is only approximate with finite spaces and finite samples, and the error is not quantified.
ad hoc to paper The inversion formula (10) extends from training samples to all (y,z).
Section 5; 'we extend the validity of this expression' is an assumption without proof.
ad hoc to paper The number of gradient iterations for convergence is independent of n.
Section 6.6; needed for the claimed linear total complexity; supported only by numerical observation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Monge optimal transport barycenter problem." pith.science (2026). https://pith.science/paper/4TSXTWEW

@misc{pith2026250703669,
  author       = {Pith},
  title        = {Pith review of: The Monge optimal transport barycenter problem},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4TSXTWEW}},
  note         = {Machine review of arXiv:2507.03669}
}

abstract

A novel methodology is developed for the solution of the data-driven Monge optimal transport barycenter problem, where the pushforward condition is formulated in terms of the statistical independence between two sets of random variables: the factors $z$ and a transformed outcome $y$. Relaxing independence to the uncorrelation between all functions of $z$ and $y$ within suitable finite-dimensional spaces leads to an adversarial formulation, for which the adversarial strategy can be found in closed form through the first principal components of a small-dimensional matrix. The resulting pure minimization problem can be solved very efficiently through gradient descent driven flows in phase space. The methodology extends beyond scenarios where only discrete factors affect the outcome, to multivariate sets of both discrete and continuous factors, for which the corresponding barycenter problems have infinitely many marginals. Corollaries include a new framework for the solution of the Monge optimal transport problem, a procedure for the data-based simulation and estimation of conditional probability densities, and a nonparametric methodology for Bayesian inference.

Figures

Figures reproduced from arXiv: 2507.03669 by the authors.

**Figure 1.** displays the data and barycenter for dx = dz = 1 and figure 2 shows the running times for various values of n, dx, and dz. Each data point displayed is the median across 10 trials of the mean time spent in each portion of the algorithm. When the dependence on n is being considered, dx and dz are both kept at 1. When either dx or dz are being varied, the other is kept at 1 and n is kept at 2000. The first two rows us… view at source ↗

**Figure 2.** Median time dependency on number and dimension of data. The top and bottom rows display the [PITH_FULL_IMAGE:figures/full_fig_p022_2.png] view at source ↗

**Figure 3.** distribution of p-values of samples generated with map inversion compared to their expectation and [PITH_FULL_IMAGE:figures/full_fig_p025_3.png] view at source ↗

Figures from the paper (14 more)

**Figure 4.** Figure 4: One-dimensional Gaussian. The leftmost column displays the samples [PITH_FULL_IMAGE:figures/full_fig_p026_4.png]

**Figure 5.** Figure 5: One-dimensional Gaussian mixture. The left panel displays results when the test function space [PITH_FULL_IMAGE:figures/full_fig_p026_5.png]

**Figure 6.** Figure 6: Two-dimensional Gaussian mixture. 9.4 An example of Bayesian inference This section illustrates model-free Bayesian inference using the OTBP (A different use of push-forward maps for Bayesian inference [38] pushes forward the prior to the posterior measure.) To demonst…

**Figure 7.** Figure 7: Original data, barycenter and simulated versus true posterior density [PITH_FULL_IMAGE:figures/full_fig_p029_7.png]

**Figure 8.** Figure 8: Time-discretized Ornstein–Uhlenbeck process. Testing time series for [PITH_FULL_IMAGE:figures/full_fig_p030_8.png]

**Figure 9.** Figure 9: Lotka-Volterra model: Testing time series for [PITH_FULL_IMAGE:figures/full_fig_p031_9.png]

**Figure 10.** Figure 10: Three examples with different relations Yz(w) between w and y: one-to-one on the first row, twoto-one on the second and z-dependent on the third. The analysis of the barycenter underlying the points {yi} may at first seem similar to residual analysis, whereby the dif…

**Figure 11.** Figure 11: Two-dimensional y dependence on a one-dimensional w (denoted with a colorbar) 9.7 Hidden patterns in ground-level atmospheric temperature We consider next the hourly ground-level temperature in Ithaca, NY from 2007 to 2023. The data, available from National Oceanic an…

**Figure 12.** Figure 12: Median temperature and 90% confidence interval as a function of day of year, time of day and both, [PITH_FULL_IMAGE:figures/full_fig_p034_12.png]

**Figure 13.** Figure 13: Median diurnal cycle and seasonal cycle of the temperature in Ithaca, NY, displayed for four days [PITH_FULL_IMAGE:figures/full_fig_p034_13.png]

**Figure 14.** Figure 14: The original time series x(t), the y1(t) resulting from removing the effects of periodic time factors (time of day and day of year), and the y2(t) resulting from further removing from y1 the synoptic weather and multi-year signals. On the right panels, a zoom-in versi…

**Figure 15.** Figure 15: Reconstruction of the conditional median temperature ¯x [PITH_FULL_IMAGE:figures/full_fig_p036_15.png]

**Figure 16.** Figure 16: Upper panels: EOF1,2,3 . The red box in EOF1 (left most panel) indicates the region of El Ni˜no events. Lower panel: C 1,2,3 l as a function of time tl . years, 1870-2014) and out-of-sample (10 years, 2015-2024) sets. We apply the procedure to each component Ck indepe…

**Figure 17.** Figure 17: Forecast 6 months ahead of the global Sea Surface Temperature: (a) visualization of the signal’s [PITH_FULL_IMAGE:figures/full_fig_p038_17.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Learning sufficient low-dimensional structures through conditional optimal transport
math.ST 2026-07 conditional novelty 7.0 of 10

Sufficiency forces the conditional optimal-transport map and its velocity to factor through the reduced covariate, and the resulting flow-matching estimator (SDR-COT) recovers the central subspace in the linear case.

Reference graph

Works this paper leans on

50 extracted references · 43 canonical work pages · cited by 1 Pith paper

[1]

Goldstein, Allen A On steepest descent.Journal of the Society for Industrial and Applied Mathematics, Series A: Control, 3, 1, 147–151, 1965

work page 1965
[2]

Armijo, Larry Minimization of functions having Lipschitz continuous first partial derivatives.Pacific Journal of mathematics, 16, 1, 1–3, 1966

work page 1966
[3]

E. G. Tabak, G. Trigila. Explanation of variability and removal of confounding factors from data through optimal transport.Communications on Pure and Applied Mathematics, 71(1):163–199, 2018

work page 2018
[5]

Optimal transport for applied mathematicians.Birk¨ auser, NY, 55(58-63):94, 2015

Filippo Santambrogio. Optimal transport for applied mathematicians.Birk¨ auser, NY, 55(58-63):94, 2015

2015
[6]

Barycenter in the Wasserstein space.SIAM J

M Agueh, G Carlier. Barycenter in the Wasserstein space.SIAM J. MATH. ANAL., 43(2):094–924, 2011. 41

work page 2011
[7]

E. G. Tabak, G. Trigila, W. Zhao. Distributional barycenter problem through data-driven flows.Pattern Recognition, 130:108795, 2022

work page 2022
[8]

B. Pass. Optimal transportation with infinitely many marginals.J. Functional Analysis, 264:947–963, 2013

work page 2013
[9]

Traversing the schr¨ odinger bridge strait: Robert fortet’s marvelous proof redux.Journal of Optimization Theory and Applications, 181(1):23–60, 2019

Montacer Essid, Michele Pavon. Traversing the schr¨ odinger bridge strait: Robert fortet’s marvelous proof redux.Journal of Optimization Theory and Applications, 181(1):23–60, 2019

work page 2019

Show all 50 references

[10]

A relationship between arbitrary positive matrices and doubly stochastic matrices.The annals of mathematical statistics, 35(2):876–879, 1964

Richard Sinkhorn. A relationship between arbitrary positive matrices and doubly stochastic matrices.The annals of mathematical statistics, 35(2):876–879, 1964

1964
[11]

Sinkhorn distances: Lightspeed computation of optimal transport

Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. InAdvances in neural information processing systems, pages 2292–2300, 2013

2013
[12]

Auricchio, F

G. Auricchio, F. Bassetti, S. Gualandi, M. Veneroni. Computing wasserstein barycenters via linear pro- gramming. InInternational Conference on Integration of Constraint Programming, Artificial Intelligence, and Operations Research, 355–363. Springer, 2019

2019
[13]

An efficient linear programming method for optimal transportation

Adam M Oberman, Yuanlong Ruan. An efficient linear programming method for optimal transportation. arXiv preprint arXiv:1509.03668, 2015

2015 arXiv
[14]

Sample-based optimal transport and barycenter problems.Communications on Pure and Applied Mathematics, 72(8):1581–1630, 2019

Max Kuang, Esteban G Tabak. Sample-based optimal transport and barycenter problems.Communications on Pure and Applied Mathematics, 72(8):1581–1630, 2019

2019
[15]

Continuous regularized wasserstein barycenters.Advances in Neural Information Processing Systems, 33:17755–17765, 2020

Lingxiao Li, Aude Genevay, Mikhail Yurochkin, Justin M Solomon. Continuous regularized wasserstein barycenters.Advances in Neural Information Processing Systems, 33:17755–17765, 2020

2020
[16]

Taghvaei, B

A. Taghvaei, B. Hosseini. An optimal transport formulation of bayes’ law for nonlinear filtering algorithms. In2022 IEEE 61st Conference on Decision and Control (CDC), 6608–6613. IEEE, 2022

2022
[17]

E. G. Tabak, E. Vanden-Eijnden. Density estimation by dual ascent of the log-likelihood.Comm. Math. Sci., 8, 2010

2010
[18]

E. G. Tabak, C. V. Turner. A family of non-parametric density estimation algorithms.CPAM, LXVI, 2013

2013
[19]

Trigila, E

G. Trigila, E. G. Tabak. Data-driven optimal transport.Comm. Pure and Applied Math., 69:613–648, 2016

2016
[20]

On finding local nash equilibria (and only local nash equilibria) in zero-sum games.arXiv preprint arXiv:1901.00838, 2019

Eric V Mazumdar, Michael I Jordan, S Shankar Sastry. On finding local nash equilibria (and only local nash equilibria) in zero-sum games.arXiv preprint arXiv:1901.00838, 2019. 42

1901 arXiv
[21]

An implicit gradient-descent procedure for minimax problems.Mathematical Methods of Operations Research, 97(1):57–89, 2023

Montacer Essid, Esteban G Tabak, Giulio Trigila. An implicit gradient-descent procedure for minimax problems.Mathematical Methods of Operations Research, 97(1):57–89, 2023

2023
[22]

Stochastic wasserstein barycenters

Sebastian Claici, Edward Chien, Justin Solomon. Stochastic wasserstein barycenters. InInternational Conference on Machine Learning, pages 999–1008. PMLR, 2018

2018
[23]

Vector quantile regression: An optimal transport approach.Ann

Guillaume Carlier, Victor Chernozhukov, Alfred Galichon. Vector quantile regression: An optimal transport approach.Ann. Statist, 44(3):1165–1192, 2016

2016
[24]

E. G. Tabak, G. Trigila, W. Zhao. Conditional density estimation and simulation through optimal trans- port.Machine Learning, 1–24, 2020

2020
[25]

E. G. Tabak, G. Trigila, W. Zhao. Data driven conditional optimal transport.Mach. Learning, 1–21, 2021

2021
[26]

E. G. Tabak, G. Trigila, W. Zhao. The conditional barycenter problem, its data-driven formulation and its solution through normalizing flows.Communications in Mathematical Sciences, 22(6):1635–1656, 2024

2024
[27]

Gretton, A

A. Gretton, A. Smola, O. Bousquet, R. Herbrich, A. Belitski, M. Augath, Y. Murayama, J. Pauls, B. Sch¨ olkopf, N. Logothetis. Kernel constrained covariance for dependence measurement. In volume R5 of Proceedings of Machine Learning Research, pages 112–119. PMLR, 06–08 Jan 2005

2005
[28]

Parallel streaming wasserstein barycenters.Advances in Neural Information Processing Systems, 30, 2017

Matthew Staib, Sebastian Claici, Justin M Solomon, Stefanie Jegelka. Parallel streaming wasserstein barycenters.Advances in Neural Information Processing Systems, 30, 2017

2017
[29]

Gromov-wasserstein averaging of kernel and distance ma- trices

Gabriel Peyr´ e, Marco Cuturi, Justin Solomon. Gromov-wasserstein averaging of kernel and distance ma- trices. InInternational conference on machine learning, pages 2664–2672. PMLR, 2016

2016
[30]

E. G. Tabak, G. Trigila. Conditional expectation estimation through attributable components.Information and Inference: A Journal of the IMA, 128(00), 2018

2018
[31]

De l’Imprimerie Royale, 1781

Gaspard Monge.M´ emoire sur la th´ eorie des d´ eblais et des remblais. De l’Imprimerie Royale, 1781
[32]

Polar factorization and monotone rearrangement of vector-valued functions.Communica- tions in Pure and Applied Mathematics, 44:371–417, 1991

Yann Brenier. Polar factorization and monotone rearrangement of vector-valued functions.Communica- tions in Pure and Applied Mathematics, 44:371–417, 1991

1991
[33]

Springer Science & Business Media, 2012

Jean Jacod, Philip Protter.Probability essentials. Springer Science & Business Media, 2012

2012
[34]

Gretton, O

A. Gretton, O. Bousquet, A. Smola, B. Sch¨ olkopf. Measuring statistical dependence with hilbert-schmidt norms. InAlgorithmic Learning Theory, pages 63–77. Springer Berlin Heidelberg, 2005. 43

2005
[35]

Experiments with a new boosting algorithm

Yoav Freund, Robert E Schapire, et al. Experiments with a new boosting algorithm. Inicml, volume 96, pages 148–156. Citeseer, 1996

1996
[36]

A randomized algorithm for principal component analysis

Vladimir Rokhlin, Arthur Szlam, Mark Tygert. A randomized algorithm for principal component analysis. SIAM Journal on Matrix Analysis and Applications, 31(3):1100–1124, 2010

2010
[37]

Incremental pca for on-line visual learning and recognition

Matej Artac, Matjaz Jogan, Ales Leonardis. Incremental pca for on-line visual learning and recognition. In2002 International Conference on Pattern Recognition, volume 3, pages 781–784. IEEE, 2002

2002
[38]

El Moselhy, Youssef M

Tarek A. El Moselhy, Youssef M. Marzouk. Bayesian inference with optimal maps.Journal of Computa- tional Physics, 231(23):7815–7850, 2012

2012
[39]

F. J. Anscombe, J. W. Tukey. The examination and analysis of residuals.Technometrics, 5:141–160, 1963

1963
[40]

Fitting models to data: Residual analysis, a primer.Uncertainty quantification and model calibration, 133, 2017

Julia Martin, David Daffos Ruiz De Adana, Agustin G Asuero. Fitting models to data: Residual analysis, a primer.Uncertainty quantification and model calibration, 133, 2017

2017
[41]

Hannachi, I

A. Hannachi, I. T. Jolliffe, D. B. Stephenson. Empirical orthogonal functions and related techniques in atmospheric science: A review.International Journal of Climatology, 27(9):1119–1152, May 2007

2007
[42]

Takahashi, A

K. Takahashi, A. Montecinos, K. Goubanova, B. Dewitte. Enso regimes: Reinterpreting the canonical and modoki el ni˜ no: Reinterpreting enso modes.Geophysical Research Letters, 38(10), May 2011

2011
[43]

Z. Wang, E. G. Tabak. Optimal transport with a density-dependent cost function.arXiv preprint arXiv:2511.02929, 2026

2026
[44]

Baptista, et al

R. Baptista, et al. Conditional simulation via entropic optimal transport: Toward non-parametric estima- tion of conditional Brenier maps.arXiv preprint arXiv:2411.07154, 2024

2024 arXiv
[45]

Hosseini, A

B. Hosseini, A. W. Hsu, and A. Taghvaei. Conditional optimal transport on function spaces.SIAM/ASA Journal on Uncertainty Quantification, 13(1):304–338, 2025

2025
[46]

E. G. Tabak, G. Trigila, and W. Zhao. Data driven conditional optimal transport.Machine Learning, 110(11):3135–3155, 2021

2021
[47]

Alfonso, et al

J. Alfonso, et al. A generative flow model for conditional sampling via optimal transport. InNeurIPS 2023 Workshop on Optimal Transport and Machine Learning, 2023

2023
[48]

Essid, E

M. Essid, E. G. Tabak, and G. Trigila. An implicit gradient-descent procedure for minimax problems. Mathematical Methods of Operations Research, 97(1):57–89, 2023. 44

2023
[49]

Hongkang Yang, E. G. Tabak. Conditional density estimation, latent variable discovery, and optimal transport.Communications on Pure and Applied Mathematics, 2020

2020
[50]

Triangular monotonic generative models can perform causal dis- covery.Causal Representation Learning Workshop at NeurIPS 2023.2023

Xi, Quanhan, Sebastian Gonzalez, and Benjamin Bloem-Reddy. Triangular monotonic generative models can perform causal dis- covery.Causal Representation Learning Workshop at NeurIPS 2023.2023

2023
[51]

Counterfactual cocycles: A framework for robust and coherent counterfactual transports

Dance, Hugh, and Benjamin Bloem-Reddy. Counterfactual cocycles: A framework for robust and coherent counterfactual transports. arXiv preprint arXiv:2405.13844 (2024). 45

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Goldstein, Allen A On steepest descent.Journal of the Society for Industrial and Applied Mathematics, Series A: Control, 3, 1, 147–151, 1965

work page 1965

[2] [2]

Armijo, Larry Minimization of functions having Lipschitz continuous first partial derivatives.Pacific Journal of mathematics, 16, 1, 1–3, 1966

work page 1966

[3] [3]

E. G. Tabak, G. Trigila. Explanation of variability and removal of confounding factors from data through optimal transport.Communications on Pure and Applied Mathematics, 71(1):163–199, 2018

work page 2018

[4] [5]

Optimal transport for applied mathematicians.Birk¨ auser, NY, 55(58-63):94, 2015

Filippo Santambrogio. Optimal transport for applied mathematicians.Birk¨ auser, NY, 55(58-63):94, 2015

2015

[5] [6]

Barycenter in the Wasserstein space.SIAM J

M Agueh, G Carlier. Barycenter in the Wasserstein space.SIAM J. MATH. ANAL., 43(2):094–924, 2011. 41

work page 2011

[6] [7]

E. G. Tabak, G. Trigila, W. Zhao. Distributional barycenter problem through data-driven flows.Pattern Recognition, 130:108795, 2022

work page 2022

[7] [8]

B. Pass. Optimal transportation with infinitely many marginals.J. Functional Analysis, 264:947–963, 2013

work page 2013

[8] [9]

Traversing the schr¨ odinger bridge strait: Robert fortet’s marvelous proof redux.Journal of Optimization Theory and Applications, 181(1):23–60, 2019

Montacer Essid, Michele Pavon. Traversing the schr¨ odinger bridge strait: Robert fortet’s marvelous proof redux.Journal of Optimization Theory and Applications, 181(1):23–60, 2019

work page 2019