Collision-based Dynamics for Multi-Marginal Optimal Transport

Hossein Gorji; Mohsen Sadr

REVIEW 4 major objections 5 minor 12 references

Collision-based Dynamics for Multi-Marginal Optimal Transport

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Random pairwise index swaps approximate multi-marginal optimal transport with linear sample cost and exponential convergence to a near-optimal answer.

desk verdict Useful linear-scaling randomized swap heuristic for multi-marginal OT, but near-optimality is measured against a truncated heuristic, not a proven guarantee. read the letter →

arxiv 2412.16385 v2 pith:XTTDC4XI submitted 2024-12-20 cs.AI cs.LGstat.CO

Mohsen Sadr , Hossein Gorji This is my paper · ORCID

classification cs.AIcs.LGstat.CO MSC 49Q2265C05

keywords multi-marginaloptimaltransportWassersteindistancecollision-baseddynamicsBoltzmannkineticsiterativeswappingalgorithmrandommini-batchinglinear-timemaplearning

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 proposes a stochastic algorithm for discrete multi-marginal optimal transport that repeatedly proposes random pairs of sample indices inside a marginal, swaps them only when the swap lowers the total transport cost, and iterates until the cost stops improving. This collision-based dynamics is a randomized version of the Iterative Swapping Algorithm (ISA): instead of checking all $O(N_p^2)$ pairs, it checks $N_p/2$ randomly chosen disjoint pairs per sweep, borrowing the collision-pair randomization of kinetic gas theory. The aim is to show that this randomization keeps the solution near-optimal while cutting per-iteration cost to $O(N_p)$ and total cost to $O(nK^2N_p)$ for $L_p$-Wasserstein costs, with memory $O(nKN_p)$ and no distance matrix. If right, the method makes multi-marginal OT practical at scales where exact linear programming, entropic-regularization solvers, and exhaustive pairwise swapping are too slow, and the paper supports this with exponential-convergence observations and experiments on synthetic densities and image datasets.

What carries the argument

The engine is a swap-acceptance kernel that compares the empirical cost before and after exchanging the labels of two particles in one marginal; the identity matrix keeps the state, the exchange matrix $J_{2n\times 2n}$ swaps it, and the swap is accepted only when the cost strictly decreases. For $L_p$-Wasserstein costs the acceptance test collapses to an $O(nK)$ local sum, $$\sum_{l\neq k}\bigl(\|$X^{{(l)}}$_i-$X^{{(k)}}$_j\|_p^p+\|$X^{{(l)}}$_j-$X^{{(k)}}$_i\|_p^p-\|$X^{{(l)}}$_i-$X^{{(k)}}$_i\|_p^p-\|$X^{{(l)}}$_j-$X^{{(k)}}$_j\|_p^p\bigr),$$ independent of $N_p$. The other ingredient is random collision-pair sampling: per sweep each marginal's indices are shuffled into an even random list, split into two halves, and only the resulting $N_p/2$ disjoint pairs are tested, replacing the exhaustive $N_p(N_p-1)/2$ scan of ISA. A Boltzmann-type operator with Heaviside collision kernel is introduced as the continuous-time heuristic that motivates the exponential-convergence expectation.

What would settle it

Construct a two-marginal instance with $N_p$ samples per side in which the only cost-reducing move is a 4-cycle: swapping sample $i$ with $j$ and sample $k$ with $\ell$ simultaneously, while each single swap raises the cost. Run the collision-based algorithm with freshly randomized pairings until the cost stops decreasing. If its terminal cost is strictly above the exhaustive-ISA or exact EMD optimum, which finds the 4-cycle, then the near-optimality claim as stated fails for that instance.

Watch

Extended reading notes

Core claim

The paper's central claim is that the discrete multi-marginal optimal transport problem—minimize expected cost over couplings of $K$ empirical measures, which is equivalent to optimizing a permutation of sample indices—can be approximated by a stochastic process that only ever exchanges two sample indices inside one marginal, accepting the exchange exactly when it lowers the empirical cost. The authors show this collision-based process preserves marginals exactly, decreases the cost monotonically, and, for $L_p$-Wasserstein costs, has per-sweep complexity $O(N_p)$ because the acceptance test for a single swap is $O(nK)$ independent of $N_p$; over $K$ marginals the total is $O(nK^2N_p)$ with $O(nKN_p)$ memory. They argue by analogy with the Boltzmann equation and the Wild expansion that the process converges exponentially to a stationary measure that is near-optimal rather than exactly optimal, since a stationary point can still harbour improvements that require more than one simultaneous swap, and they report convergence rates and EMD-relative errors consistent with this picture on images and synthetic densities.

Load-bearing premise

The linear-time speedup rests on the unproven assumption that randomly sampling only half as many collision pairs as there are samples, rather than checking every pair like ISA, does not bias the process away from the permutations that exhaustive swapping would reach; the paper cites existing random-batch consistency results but explicitly leaves the proof for this process to future work.

Editorial extensions

If this is right

Large-sample multi-marginal OT becomes affordable: per-iteration work is $O(N_p)$, total work $O(nK^2N_p)$, and memory $O(nKN_p)$, so no distance matrix is ever constructed.
Because the cost decreases monotonically and the observed convergence is exponential in the number of sweeps, a fixed number of sweeps (tens to hundreds) reaches practical accuracy across a wide range of $N_p$ and $K$ in the reported tests.
The algorithm can be used as a fast approximate OT oracle inside learning pipelines: it returns aligned sample pairs, which the paper uses to train neural maps between densities and between gray and color images.
Marginals are preserved to machine precision and collision pairs in a sweep are disjoint, so the scheme parallelizes over collision candidates without data races.
In the two-marginal case the same algorithm acts as a fast approximate Wasserstein solver, matching the exact linear-programming solution's reported distances with large speed-ups in the reported tests.

Reading between the lines

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

If the mini-batch selection is unbiased as assumed, the same randomization recipe applies to any local-search algorithm whose single-step improvement test is cheap, so quadratic pairwise-scan methods for matching, assignment, and rearrangement problems could be accelerated the same way.
The reported exponential convergence with a sample-independent rate suggests that an adaptive stopping rule based on the observed cost-decay exponent could be built, stopping when the predicted remaining improvement falls under a tolerance; the paper does not implement such a rule.
A useful stress test beyond the paper's benchmarks would be rugged or random cost landscapes, where pairwise monotone descent is known to get stuck; if the gap to the exact optimum grows there, adding a temperature or restart mechanism is the natural remedy.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Quick take: the collision-based swap heuristic is a genuinely useful practical algorithm for discrete MMOT, and the paper is honest about most of its gaps. The multi-marginal experiments, however, measure error against a 10-iteration ISA reference, so the central 'near-optimal' claim is not yet established. That doesn't sink the paper, but it should be reframed or substantiated.

What's new: this is the first minibatch/random-pair variant of ISA I know of. Instead of checking all Np(Np-1)/2 pairs per marginal per iteration, it shuffles and checks Np/2 pairs. The swap-cost update for Lp costs is O(1), so per-iteration cost is O(nK Np) and claimed total is O(nK^2 Np) if iterations are independent of Np. Marginal preservation and monotone cost decrease are by construction. The K=2 tests against exact EMD are real evidence: the error decays empirically and the scaling curves match the linear claims. The five-marginal map learning and image colorization are nice applications, and the tables show the method is competitive with ISA at lower wall-clock time.

Soft spots: the paper explicitly leaves the consistency between the random collision process and ISA to future work (Sec. 3), and notes the stationary solution only guarantees no improvement via binary swaps (Sec. 5). So the theoretical core is a hope, not a result. The exponential convergence argument imports Wild expansion for the Boltzmann equation with constant alpha, but alpha in Eq. (15) is state-dependent; the discrete swap process is not shown to satisfy that kinetic equation. The total-complexity claim depends on iteration count being independent of Np, which is only empirical. The multi-marginal relative errors are computed against ISA with 10 iterations, so they measure closeness to a truncated heuristic. Also, no code is provided; for an algorithmic paper, that's a real gap.

I think the stress-test's minibatch-bias worry is plausible but not demonstrated; the K=2 EMD comparisons suggest the bias is small in practice on those problems, but the multi-marginal case remains open. The paper's own caveats are fair, and it doesn't hide the theory gap. Still, 'approximates the solution' in the abstract overstates what is proven.

Who it's for: someone looking for a scalable heuristic for MMOT with large Np, especially in high-dimensional or memory-limited settings. A serious referee should engage with it; the algorithm is simple enough to test and the practical value is clear. I'd ask the authors to release code, replace the ISA-10-iteration reference with a much tighter one (or an exact LP on small instances), and either prove consistency or explicitly market the method as a heuristic. My recommendation: send to peer review, with major revision expectations.

Referee Report

4 major / 5 minor

Summary. The paper proposes a collision-based dynamics for discrete multi-marginal optimal transport (MMOT), in which each iteration randomly selects Np/2 disjoint pairs of samples per marginal and swaps them if the swap lowers the empirical transport cost. This is presented as a randomized, mini-batch version of the Iterative Swapping Algorithm (ISA), inspired by Bird's DSMC method for the Boltzmann equation. The authors claim per-iteration complexity O(Np), total complexity O(nK^2 Np) for Lp costs, memory O(nKNp), and exponential convergence to a near-optimal stationary solution. The method is tested on toy problems, a five-marginal map-learning task, and image datasets (JAFFE, butterfly, CelebA), with comparisons against EMD, Sinkhorn, and ISA.

Significance. If the algorithmic claims were fully established, this would be a practically valuable method for large-scale MMOT: it is simple, parallelizable, preserves marginals by construction, and the reported wall-clock and memory scaling are substantially better than Sinkhorn or EMD in the tested regimes. The K=2 validation against exact EMD (Fig. 3 and Fig. 6) gives real evidence that the method can approximate the optimal transport solution in the two-marginal case. However, the central near-optimality claim for the multi-marginal setting is not proven: the paper explicitly defers the consistency proof between the collision-based process and ISA, and the multi-marginal comparisons use a truncated ISA run as the reference solution. The current manuscript is best read as a promising empirical heuristic with rigorous per-iteration complexity analysis, not as a proven approximation scheme for MMOT.

major comments (4)

[Sec. 3, item 2 and Sec. 4] The mini-batch randomization is load-bearing for the O(Np) per-iteration claim, but the paper explicitly states that the theoretical consistency between the collision-based process and ISA is left to separate studies. The unbiasedness of Bird's DSMC randomization applies to collision rates in kinetic Monte Carlo, not to the fixed point of a greedy combinatorial local search. Because the swap acceptance rule depends on the current configuration, random pair selection could in principle bias the stationary permutation away from what exhaustive ISA would find. Please either provide a proof of unbiasedness for this greedy setting or add experiments on small problems (e.g., Np ≤ 50, K ∈ {2,3,4}) where the exact MMOT optimum is computable, measuring the gap between the fixed points of random mini-batching and exhaustive scanning.
[Sec. 5, 'Monotone convergence' and Appendix A] The exponential convergence claim in Eq. (17) is imported from the Wild expansion for the Boltzmann equation under a constant collision frequency alpha-hat, but the collision frequency alpha in Eq. (15) is state-dependent, and the paper does not show that the discrete-time swap process satisfies the continuous-time kinetic equation (14). The sentence "we expect" and the reference to Wagner (1992) do not constitute a proof. Consequently, the total complexity O(nK^2 Np) also rests on the unproven assumption that the number of iterations to convergence is independent of Np; Fig. 5 shows this empirically, but the paper should present this as an empirical observation rather than a proven complexity bound.
[Sec. 6, Figs. 2 and 5, Tables 1-8] The multi-marginal experiments, which constitute the paper's main novelty, use "ISA with 10 iterations" as the reference solution. The reported relative errors therefore measure closeness to a truncated heuristic, not to the MMOT optimum. The K=2 comparisons against EMD (Fig. 3 and Fig. 6) are strong evidence for the two-marginal case, but they do not cover the multi-marginal claim. For small K (e.g., K=3 or 4) and small Np, please compare the collision-based method against the exact MMOT solution from linear programming, or provide a nontrivial lower bound on the optimal cost, to substantiate the near-optimality claim.
[Sec. 5, 'The stationary solution may not be optimal'] The admission that the stationary solution is only a local optimum with respect to binary swaps is in tension with the abstract's claim of "approximates the solution" of the MMOT problem. The paper provides no quantitative optimality gap for the multi-marginal cases; the error bars in Tables 1-8 are relative to ISA(10 iterations), and Fig. 2 shows errors of order 10^-2 to 10^-1 for both methods. Please clarify what "near-optimal" means and provide evidence, or a clear caveat, that the stationary permutation is close to the global MMOT optimum for the tested cases.

minor comments (5)

[Appendix A title] The title "Wild expansion for the Bolztmann equation" contains a typo: "Bolztmann" should be "Boltzmann".
[Algorithm 2 output line] The final line of Algorithm 2 reads "Ouput: X"; this should be "Output: X".
[Eqs. (14)-(15) and Appendix A] The notation alpha(x,y) in Eq. (15) suggests a function of position, but the integral also depends on the time-dependent density rho_t; the paper later assumes a constant alpha-hat in the Wild expansion. Please clarify the time/state dependence and the relationship between alpha(x,y) and alpha-hat.
[Sec. 5, complexity paragraph] In the sentence on memory consumption, "O(nKN p)" should be written as "O(nKNp)" for consistency with the rest of the paper.
[Fig. 5, left panel] The exponential fit annotation "exp(-0.34 NIter)" spans multiple Np values; please clarify whether this is a single fit to all curves or a representative curve, and add the corresponding fit details.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the proposed algorithm is a self-contained randomized minibatch variant of ISA, and the cited Boltzmann/Wild-expansion material is explicitly heuristic rather than load-bearing.

full rationale

The derivation chain is self-contained: Section 2 reduces the discrete MMOT problem to a permutation search (Eqs. (8)-(9)) using standard facts about bistochastic matrices and permutation matrices; Section 3 defines ISA and then the collision-based process as a randomized minibatch version of ISA; Sections 4-5 specify Algorithm 1 and its per-iteration cost. The linear per-iteration complexity O(nK^2 Np) follows by counting local cost checks (each O(nK), repeated over Np/2 pairs and K marginals), not by fitting or by importing a conclusion from the authors' prior work. The exponential-convergence claim in Section 5 is explicitly conditioned ("If the proposed collision-based dynamics behaves similar to the Boltzmann kinetics, we expect..."), and the paper openly defers the consistency proof ("we leave the theoretical consistency between collision-based process and ISA to separate studies"). Appendix A's Wild expansion assumes a constant alpha in Eq. (18), whereas Eq. (15) defines a state-dependent alpha; this is an unproven analogy that weakens the theoretical justification, but it is not a circular reduction because the claim is not derived from the quantity it purports to predict. The multi-marginal optimality evidence in Figs. 2 and 5 and Tables 1-8 uses ISA with 10 iterations as the reference solution, which limits how strongly the results support near-optimality but does not make the algorithm's output equivalent to its inputs. The self-citations (Sadr et al. 2024a,b) are background references, not load-bearing; no fitted parameter is renamed as a prediction, and the exponential overlays in Figs. 1 and 5 are descriptive fits rather than predictions derived from the theory. Therefore no circular step is present.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The contributions are supported by standard OT theory (discrete permutation reduction, Gangbo-Swiech maps) and by an unproven analogy to kinetic theory. The main uncharged load is the mini-batch randomization and the Boltzmann-type kinetic description, both acknowledged by the authors as requiring future work.

free parameters (2)

Empirical exponential decay rates (alpha) in convergence plots = 0.28, 0.58, 0.56, 0.34 per iteration
Rates in Fig. 1 and Fig. 5 are fitted to the observed cost curves to illustrate exponential convergence; they are not derived from the collision dynamics or predicted independently.
Reference iteration count for pseudo-ground truth (ISA 10 iterations) = 10 iterations of ISA
Used in Fig. 2, Fig. 5, and Tables 2-8 as the reference to define relative error; the resulting optimality error is relative to an approximate solution, not the exact OT solution.

assumptions (5)

standard math Existence and uniqueness of optimal maps for the Gangbo-Swiech L2-MMOT cost (Eq. 4)
Invoked in Sec. 1 through Gangbo and Swiech (1998) to justify optimizing over maps; not re-derived.
standard math For discrete measures, the MMOT optimum is attained at a permutation matrix (Birkhoff-von Neumann)
Used in Sec. 2 Eqs. (8)-(9) to reduce the problem to a search over permutations.
domain assumption Empirical measures converge weakly to the true marginals as Np grows
Assumed in Sec. 2 to connect discrete permutation solutions to the continuum OT solution; standard but only valid in the large-sample limit.
ad hoc to paper Random mini-batch pair selection in the OT swap process is unbiased and preserves the near-optimality of the stationary permutation
Assumed in Sec. 3 and 4; the paper cites Liu and Wang (2024) for random batch methods but explicitly leaves the consistency proof for the swap dynamics to future work.
ad hoc to paper The discrete-time collision process can be modeled by a continuous-time Boltzmann-type kinetic equation (Eqs. 14-16) with the given collision kernel
Assumed in Sec. 3 and 5 to import Wild-expansion exponential convergence; the paper notes the two descriptions operate on different time scales and forgoes proof.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Collision-based Dynamics for Multi-Marginal Optimal Transport." pith.science (2026). https://pith.science/paper/XTTDC4XI

@misc{pith2026241216385,
  author       = {Pith},
  title        = {Pith review of: Collision-based Dynamics for Multi-Marginal Optimal Transport},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XTTDC4XI}},
  note         = {Machine review of arXiv:2412.16385}
}

read the original abstract

Inspired by the Boltzmann kinetics, we propose a collision-based dynamics with a Monte Carlo solution algorithm that approximates the solution of the multi-marginal optimal transport problem via randomized pairwise swapping of sample indices. The computational complexity and memory usage of the proposed method scale linearly with the number of samples, making it highly attractive for high-dimensional settings. In several examples, we demonstrate the efficiency of the proposed method compared to the state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2412.16385 by the authors.

**Figure 2.** Evolution of Wasserstein distance estimate as a function of number of collision candidates [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** Optimal transport between 10 randomly selected pairs of pictures from JAFFE dataset, [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Transport map MY →Z based on paired samples of a five-marginal optimal transport problem, i.e. from the normally distributed one-marginal Y = [X(1)] (top) to a four-marginal output Z = [X(2), X(3), X(4), X(5)] (middle and bottom) consisting of the Swiss roll, ring, fun…

**Figure 5.** Figure 5: Evolution of the cost function during collision-based MMOT Algorithm 1 per number of [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]

**Figure 6.** Figure 6: Evolution of the error in the Wasserstein distance [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]

**Figure 7.** Figure 7: Distribution of relative error er := |d 2 (Xcoll) − d 2 (XEMD)|/d2 (XEMD) (left), speed-up, i.e. ratio of EMD execution time to collision method (middle), and ratio of memory consumption, i.e. EMD to collision method, (right) for 100 randomly selected pairs of picture …

**Figure 8.** Figure 8: After training a network to learn the map [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 10 canonical work pages

[1]

7, the proposed collision-based approach outperforms Sinkhorn in the distribution of relative error in the Wasserstein distance, speed-up and memory consumption

As shown in Fig. 7, the proposed collision-based approach outperforms Sinkhorn in the distribution of relative error in the Wasserstein distance, speed-up and memory consumption. 19 10 4 10 1 102 er [ ] 0 5 10 15 20Distribution of er Coll. ISA Sinkhorn, = 0.5 Sinkhorn, = 0.1 101 102 Speed-Up [-] 0 5 10 15Distribution of Speed-Up Collision ISA Sinkhorn, = ...

work page 2000
[1942]

Adam: A method for stochastic optimization

Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980,

arXiv
[1951]

(32) By induction, we have 13 F P (1) = ρt0 , (33) F Pr(n) = P [F Ps(m), FPt(n−m)]

denoted by Pr(n) where Pr(n) = Ps(m)Pt(n − m) . (32) By induction, we have 13 F P (1) = ρt0 , (33) F Pr(n) = P [F Ps(m), FPt(n−m)] . (34) Furthermore, consider a numerical function of Pr(n) as gr(n) by the relations g(1) = 1 (35) gr(n) = 1 n − 1 gs(m)gt(n − m) (36) where P r gr = 1 for all n. This leads to ρt(x, y) = e− ˆαt ∞X n=1 (1 − e− ˆαt)n−1 X r gr(n...

work page 1951
[1955]

On random batch methods (RBM) for interacting particle systems driven by Levy processes

Jian-Guo Liu and Yuliang Wang. On random batch methods (RBM) for interacting particle systems driven by Levy processes. arXiv preprint arXiv:2412.06291,

work page arXiv
[1990]

Hossein Gorji

Mohsen Sadr, Peyman Mohajerin Esfehani, and M. Hossein Gorji. Optimal transportation by orthogonal coupling dynamics. preprint on arXiv:2410.08060, 2024a. Mohsen Sadr, Nicolas G Hadjiconstantinou, and M Hossein Gorji. Wasserstein-penalized entropy closure: A use case for stochastic particle methods. Journal of Computational Physics, 511:113066, 2024b. Tom...

work page arXiv
[1994]

Displacement interpolation using Lagrangian mass transport

Nicolas Bonneel, Michiel Van De Panne, Sylvain Paris, and Wolfgang Heidrich. Displacement interpolation using Lagrangian mass transport. In Proceedings of the 2011 SIGGRAPH Asia conference, pages 1–12,

work page 2011
[2000]

Fine-grained repre- sentation learning and recognition by exploiting hierarchical semantic embedding

Tianshui Chen, Wenxi Wu, Yuefang Gao, Le Dong, Xiaonan Luo, and Liang Lin. Fine-grained repre- sentation learning and recognition by exploiting hierarchical semantic embedding. In Proceedings of the 26th ACM international conference on Multimedia, pages 2023–2031,

work page 2023
[2015]

Fast and robust earth mover’s distances

Ofir Pele and Michael Werman. Fast and robust earth mover’s distances. In 2009 IEEE 12th international conference on computer vision, pages 460–467. IEEE,

work page 2009

Show all 12 references

[2017]

Let us revisit the Botlzmann equation

12 A Wild expansion for the Bolztmann equation In this section, we review the Wild expansion for the Boltzmann equation which is used as the basis to justify the exponential behavior of the collisional OT. Let us revisit the Botlzmann equation. The particle-particle interactio...

1951
[2018]

Projected Langevin dynamics and a gradient flow for entropic optimal transport

Giovanni Conforti, Daniel Lacker, and Soumik Pal. Projected Langevin dynamics and a gradient flow for entropic optimal transport. arXiv preprint arXiv:2309.08598,

arXiv
[2019]

Celebrating cercignani’s conjecture for the boltzmann equation

Laurent Desvillettes, Clément Mouhot, and Cédric Villani. Celebrating cercignani’s conjecture for the boltzmann equation. arXiv preprint arXiv:1009.4006,

arXiv
[2023]

We draw Np samples from the two marginals, and solve the optimal transport problem using the proposed collision-based algorithm 1, EMD and Sinkhron

for details. We draw Np samples from the two marginals, and solve the optimal transport problem using the proposed collision-based algorithm 1, EMD and Sinkhron. As shown in Fig. 6, the proposed algorithm outperforms the benchmark at a reasonable error, both in terms of execut...

2000

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

7, the proposed collision-based approach outperforms Sinkhorn in the distribution of relative error in the Wasserstein distance, speed-up and memory consumption

As shown in Fig. 7, the proposed collision-based approach outperforms Sinkhorn in the distribution of relative error in the Wasserstein distance, speed-up and memory consumption. 19 10 4 10 1 102 er [ ] 0 5 10 15 20Distribution of er Coll. ISA Sinkhorn, = 0.5 Sinkhorn, = 0.1 101 102 Speed-Up [-] 0 5 10 15Distribution of Speed-Up Collision ISA Sinkhorn, = ...

work page 2000

[2] [1942]

Adam: A method for stochastic optimization

Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980,

arXiv

[3] [1951]

(32) By induction, we have 13 F P (1) = ρt0 , (33) F Pr(n) = P [F Ps(m), FPt(n−m)]

denoted by Pr(n) where Pr(n) = Ps(m)Pt(n − m) . (32) By induction, we have 13 F P (1) = ρt0 , (33) F Pr(n) = P [F Ps(m), FPt(n−m)] . (34) Furthermore, consider a numerical function of Pr(n) as gr(n) by the relations g(1) = 1 (35) gr(n) = 1 n − 1 gs(m)gt(n − m) (36) where P r gr = 1 for all n. This leads to ρt(x, y) = e− ˆαt ∞X n=1 (1 − e− ˆαt)n−1 X r gr(n...

work page 1951

[4] [1955]

On random batch methods (RBM) for interacting particle systems driven by Levy processes

Jian-Guo Liu and Yuliang Wang. On random batch methods (RBM) for interacting particle systems driven by Levy processes. arXiv preprint arXiv:2412.06291,

work page arXiv

[5] [1990]

Hossein Gorji

Mohsen Sadr, Peyman Mohajerin Esfehani, and M. Hossein Gorji. Optimal transportation by orthogonal coupling dynamics. preprint on arXiv:2410.08060, 2024a. Mohsen Sadr, Nicolas G Hadjiconstantinou, and M Hossein Gorji. Wasserstein-penalized entropy closure: A use case for stochastic particle methods. Journal of Computational Physics, 511:113066, 2024b. Tom...

work page arXiv

[6] [1994]

Displacement interpolation using Lagrangian mass transport

Nicolas Bonneel, Michiel Van De Panne, Sylvain Paris, and Wolfgang Heidrich. Displacement interpolation using Lagrangian mass transport. In Proceedings of the 2011 SIGGRAPH Asia conference, pages 1–12,

work page 2011

[7] [2000]

Fine-grained repre- sentation learning and recognition by exploiting hierarchical semantic embedding

Tianshui Chen, Wenxi Wu, Yuefang Gao, Le Dong, Xiaonan Luo, and Liang Lin. Fine-grained repre- sentation learning and recognition by exploiting hierarchical semantic embedding. In Proceedings of the 26th ACM international conference on Multimedia, pages 2023–2031,

work page 2023

[8] [2015]

Fast and robust earth mover’s distances

Ofir Pele and Michael Werman. Fast and robust earth mover’s distances. In 2009 IEEE 12th international conference on computer vision, pages 460–467. IEEE,

work page 2009