Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Spectral Shrinkage of Gaussian Entropic Optimal Transport

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read For any two Gaussian measures, the entropic optimal coupling is uniquely determined by a single spectral filter applied to the aligned correlation operator; as ε→0 it selects the most diffusive optimal plan, never Monge.

desk verdict The spectral-shrinkage derivation is clean and the rates are nice, but the universal 'no assumptions' claims and the selection principle rest on the author's own unreviewed preprint [40]—real content, with the universality not yet earned. read the letter →

arxiv 2512.19457 v2 pith:N2AOV6DK submitted 2025-12-22 math.OC math.FA

Ho Yun This is my paper · ORCID
classification math.OCmath.FA MSC 49Q2260B1147B1046N30
keywords EntropicOptimalTransportGaussianMeasuresBures-WassersteinGeometryCovarianceOperatorsSpectralShrinkageFunctionalAnalysisSchurcomplementconvergencerates
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

This paper claims that entropic optimal transport (EOT) between Gaussian measures on a separable Hilbert space—without assuming invertibility or range inclusion of the covariances—has an exact, unique solution given by a universal spectral shrinkage. The optimal cross-covariance is G f_ε(G*M) M*, where G and M are any properly aligned square roots of the two covariances and f_ε(x)=2x/(√(4x²+ε²)+ε) is a single scalar function. This recasts the problem from iterative Sinkhorn computation into a one-time eigendecomposition. As ε→0 the entropic couplings converge in Wasserstein distance to the most diffusive optimal Kantorovich coupling—the centroid of the set of optimal plans—and whenever the Schur complement does not vanish the limit is provably not a Monge solution. The paper also derives closed-form entropic-bias and coupling-distance formulas, recovering the finite-dimensional rate ε log(1/ε) and giving non-parametric rates ε^{1−1/α} for polynomially decaying spectra.

What carries the argument

The key object is the properly aligned Green's pair: factorizations G G* = A and M M* = B with G*M ⪰ 0, whose product spectrum is invariant and equals the spectrum of (A^{1/2} B A^{1/2})^{1/2}. The universal spectral shrinkage function f_ε(x) = 2x/(√(4x²+ε²)+ε) acts as the scalar optimizer that decouples the entropic variational problem into independent one-dimensional problems, yielding the closed-form correlation operator R_ε = f_ε(G*M). The generalized Schur complement B/A measures the residual variance in one marginal unexplained by the other and controls uniqueness (B/A = 0) and the failure of Monge selection in the limit.

What would settle it

Take A = diag(1,1,0) and B = diag(0,1,1) on R³, build the canonical aligned pair from (6), compute G f_ε(G*M) M* for small ε, and compare the resulting cross-covariance with a Sinkhorn solution of the Gaussian EOT problem; a persistent mismatch beyond numerical tolerance refutes Theorem 3.4. A second check is whether the ε→0 limit equals the N₁₂ = 0 centroid coupling rather than a Monge plan with N₁₂*N₁₂ = B/A.

Watch

Extended reading notes

Core claim

The central discovery is Theorem 3.4: for trace-class covariances A and B and any properly aligned Green's operators G and M, the unique Gaussian EOT coupling is N(0, Σ_ε) with cross-covariance G R_ε M*, where R_ε = f_ε(G*M). Because f_ε is a monotone scalar function mapping [0,∞) to [0,1), the entropic solution is a spectral contraction of the correlation operator of the optimal Kantorovich plan. Theorem 3.9 then identifies the ε↓0 limit: the couplings converge in W₂ to the centroid of the convex set of optimal Kantorovich couplings, constructed by setting the free parameter N₁₂ = 0. This implies that in any degenerate regime where the generalized Schur complement B/A is nonzero, the entrop

Load-bearing premise

The degenerate-case conclusions—Theorem 3.9, the never-Monge claim, and the ε→0 selection principle—rest on the parameterization of all optimal Kantorovich couplings imported from the companion preprint [40] (Section 2.2, Theorem 2.1), which is not proved here; the non-parametric rates additionally assume exact polynomial eigenvalue decay s_k = c k^{−α}.

Editorial extensions

If this is right

  • For any trace-class A and B, the Gaussian EOT solution is unique for every ε > 0 and independent of the chosen properly aligned Green pair, removing the usual injectivity and range-inclusion assumptions.
  • As ε → 0 the entropic coupling converges to the most diffusive optimal Kantorovich coupling (the N₁₂ = 0 centroid), so whenever the Schur complement is nonzero the limit is never a Monge plan.
  • The entropic cost bias has the closed form tr[g_ε((A^{1/2} B A^{1/2})^{1/2})]; in finite rank it scales as ε log(1/ε) with constant equal to the rank, and under polynomial spectral decay s_k = c k^{−α} it scales as ε^{1−1/α}.
  • A single eigendecomposition of G*M yields exact EOT costs and couplings for every ε > 0—an O(d³) one-time cost followed by O(d) per ε—replacing Sinkhorn iterations and their divergence as ε ↓ 0.
  • The Wasserstein distance between the entropic coupling and the canonical OT coupling is given explicitly by W₂²(πε, π₀) = 2 tr[A + B − √Q_ε], with operator-based upper bounds expressed through spectral perturbation functions.

Reading between the lines

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

  • If the spectral-shrinkage formula is taken as the defining structure of a 'Gaussian entropic semigroup', then changing the divergence (for example to an α-divergence) or the reference coupling would correspond to replacing f_ε by another scalar function, and the same spectral decomposition would still yield exact multi-scale paths without re-solving.
  • The selection of the most diffusive plan at ε→0 is plausibly a general phenomenon in entropic OT beyond Gaussians: in any degenerate problem where the optimal plan set is a convex polytope, entropic regularization may steer the limit to the set's centroid rather than to an extreme Monge point. This is a testable extension, not established by the paper.
  • The one-step update of Proposition 3.8 offers a practical way to construct proper alignment from arbitrary Cholesky factors; if implemented in high dimensions it would sidestep explicit Schur-complement block decompositions, which matters for operator-valued data.
  • The non-parametric rates carry an implicit practical warning: for functional data with slow spectral decay (small α), entropic bias decays slowly, so ε should be chosen relative to the spectrum rather than as a universal small constant.
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.

Referee Report

3 major / 4 minor

Summary. The paper studies entropic optimal transport (EOT) between centered Gaussian measures on a separable Hilbert space with trace-class covariances A and B. It claims that for any properly aligned Green's operators G,M (GG*=A, MM*=B, G*M≥0), the unique optimal Gaussian entropic coupling has cross-covariance Cε=G fε(G*M) M*, with universal spectral shrinkage fε(x)=2x/(√(4x²+ε²)+ε). This is proved by diagonalizing G*M, applying the von Neumann trace inequality, and solving a scalar problem. The paper then derives the entropic cost bias (Cor. 3.5), recovering the finite-rank rate ε log(1/ε) and a non-parametric rate ε^{1−1/α} under exact polynomial spectral decay; examines the ε→0 limit (Thm 3.9), claiming convergence to the most diffusive optimal Kantorovich plan (the N12=0 centroid), which is never a Monge plan when the Schur complement is nonzero; and gives stability bounds and closed-form coupling distances (Thms 3.10, 3.12). Structural facts about degenerate Gaussian OT—existence of properly aligned pairs and parameterization of all optimal couplings—are imported from the author's companion preprint [40].

Significance. If the imported structural results are sound, the spectral-shrinkage formula is a significant unifying result: it removes injectivity and range-inclusion assumptions, gives a transparent geometric interpretation, and reduces computation to one spectral decomposition for all ε. The derivation of Theorem 3.4 is clean and self-contained conditional on the existence of a properly aligned pair; the bias and coupling-distance formulas are explicit, and the finite-rank ε log(1/ε) recovery is reassuring. The predicted non-parametric rates for infinite-dimensional settings are a useful caution. However, the paper's universality and selection principles are only as strong as the unproved results in [40]; the rate theorem's exact power-law assumption also requires scrutiny. With those points addressed, the paper would be a valuable contribution.

major comments (3)
  1. [§2.2, Thm 3.4, Thm 3.9] The proof of Theorem 3.4 assumes existence of a properly aligned Green pair (G,M) for arbitrary trace-class A,B; this existence is not proved here but imported from [40]. Similarly, Theorem 3.9's identification of the entropic limit with the N12=0 plan and the 'never Monge' conclusion depends entirely on [40, Theorem 2.1], which parameterizes all optimal Kantorovich couplings. Since these structural results are load-bearing for the paper's central universality and selection claims, the manuscript is not self-contained. Please either prove these results in an appendix or explicitly state the theorems under the assumption that such a properly aligned pair exists / that Theorem 2.1 holds, and flag the degree of reliance.
  2. [Cor 3.5(2), Ex 3.1] The nonparametric rate theorem assumes an exact power-law spectrum s_k=c k^{-α}. The Riemann-sum calculation in Cor 3.5(2) uses this exact identity to substitute u_k=k(ε/c)^{1/α}. For integrated Brownian motions, only the asymptotic equivalence s_k≍(πk)^{-n} is available; the exact equality fails at every finite k. Therefore the claimed 'exact asymptotic convergence rate' for IBMs does not follow from the stated proof. A comparison or Tauberian argument is needed to justify the same exponent (and constant) under asymptotic polynomial decay, or the theorem should be restricted to operators with exact power-law spectra.
  3. [Prop 3.8, Remark 3.2] The purported constructive update M0=K0+(B−K0K0*)^{1/2} (and the general Remark 3.2) is not proved: it is not shown that the cross term K0(B−K0K0*)^{1/2} vanishes for the coordinate-free K0, nor that the operator square root can be chosen to make M0 a Green's operator for B; the block verification assumes the canonical block decomposition and cites [40, Lemma 2.4]. Since this construction is offered as the route to proper alignment for arbitrary A,B, it needs a full proof or should be removed from the claims of full generality.
minor comments (4)
  1. [After Cor 3.5] Typo: 'our result recovers the our result recovers' should read 'our result recovers'.
  2. [Thm 3.10] In the display for Qε, '(M∗M)' should be '(M0∗M0)'.
  3. [References] Reference [40] is an unpublished preprint; given the heavy reliance, the reference should be marked as 'submitted' or 'preprint' and, ideally, the relevant statements reproduced.
  4. [§2.1/Thm 3.4] The spectral convention in §2.1 orders degenerate nonzero eigenvalues by magnitude starting at λ2, while Theorem 3.4 writes λ1 ≥ λ2 ≥ ... > 0 with zero excluded; please harmonize the notation.

Circularity Check

2 steps flagged · score 4.0 of 10

Core spectral-shrinkage derivation is self-contained, but the 'no-assumptions' universality and the 'never Monge' selection principle are imported from the author's own unreviewed preprint [40].

  1. self citation load bearing [Section 2.2, Definition 2.5 / Theorem 2.1 (imported from [40]); used in Theorem 3.4]
    "It has been established in [40] that for arbitrary A, B ∈ B+1(H), A ⇝ B holds if and only if dim[R(B1/2) ∩ H2] ≤ dim[N(B1/2A1/2) ∩ H1], and thus either A ⇝ B or B ⇝ A; that is, a Monge solution always exists in at least one direction."

    Theorem 3.4 is stated for arbitrary trace-class A,B and assumes existence of properly aligned Green operators G,M. That existence is not proved here; it is taken from [40], an unreviewed preprint by the same author. The paper's advertised universality ('no nullity conditions nor inclusion assumptions') therefore rests on a load-bearing self-citation. If the reachability/alignment theorem from [40] is defective, the universal spectral-shrinkage formula is unsupported rather than derived.

  2. uniqueness imported from authors [Section 3.3, before Theorem 3.9]
    "When this condition fails, the set of optimal unregularized couplings is parametrized by the choice of N12 : H2 → H1 in Theorem 2.1 satisfying R(N12) ⊂ N(B1/2 11 A1/2 11 ) and N∗ 12N12 ⪯ B/A. ... In stark contrast, we claim that the entropic regularization mechanism selects the most diffusive solution among the optimal set. This leads to a direct implication: whenever the optimal transport solution is non-unique (i.e., B/A ̸= 0), the EOT limit is never a Monge solution."

    The 'never Monge' and 'most diffusive/centroid' selection principle is not derived from the spectral formula alone. It relies on [40, Theorem 2.1] to assert that all optimal Kantorovich couplings are parameterized by N12, and that Monge solutions are exactly those saturating N12*N12 = B/A. Theorem 3.9 only proves that the paper's own Rε formula converges to the self-defined canonical Σ0 (the N12=0 choice); calling that limit the centroid of the whole optimal set is an imported uniqueness/parameterization theorem from the authors' prior work, used as if it were an external mathematical fact.

full rationale

The central optimization is self-contained and non-circular. The paper explicitly maximizes Pε(R) = 2 tr[R(M*G)] + ε log det(I − R*R) over the admissible Hilbert-Schmidt correlations; von Neumann's trace inequality fixes the singular vectors to the eigenbasis of G*M, and the scalar first-order condition gives Rε = fε(G*M) with fε(x)=2x/(√(4x²+ε²)+ε). No fitted constant, data subset, or hidden parameter shapes this prediction; it is a direct variational consequence of the Gaussian KL identity (Prop. 3.3) and the Feldman-Hájek dichotomy. The entropic-bias formulas (Cor. 3.5) and non-parametric rates are spectral sums evaluated under stated polynomial-decay assumptions, not fitted to the target quantities. The circularity concern is limited to the degenerate-case machinery imported en bloc from the author's own unreviewed preprint [40]: the existence of properly aligned Green pairs for arbitrary A,B, the reachability dichotomy A⇝B or B⇝A, and the characterization of ALL optimal Kantorovich couplings by N12. These are load-bearing for the paper's 'no assumptions' universality and for the 'EOT limit is never Monge' selection principle, but they are not part of the independent spectral-shrinkage derivation. Hence the paper has some load-bearing self-citation and one imported uniqueness theorem, but the main formula retains independent content; score 4.

Assumptions & free parameters 0 free parameters · 4 assumptions · 2 invented entities

No fitted parameters appear: the shrinkage function f_ε, the bias function g_ε, and the rate constants (e.g., c^{1/α}∫g_1(t^{−α})dt) are derived from the entropic objective and the spectral model, not tuned to data. ε is the problem's regularization input, not a free parameter. The axioms are the standard analytic toolkit plus three paper-specific premises — the imported degenerate-case geometry of [40], the polynomial spectral-decay model for the rates, and the Volterra asymptotics — none of which are proved in this text. Invented entities are limited to two internal operator-theoretic constructions, both lacking independent external evidence by their nature.

assumptions (4)
  • domain assumption Existence of properly aligned Green's pairs and the reachability dichotomy A⇝B or B⇝A for any trace-class A,B (Section 2.2, Definition 2.4, Theorem 2.1, quoted from [40])
    The EOT spectral theorem and the ε→0 selection principle presuppose this structure; proofs are in the author's unreviewed preprint [40] and are not reproduced here.
  • domain assumption N₁₂-parameterization of all optimal Kantorovich couplings (Theorem 2.1 from [40]): optimal Gaussian couplings are exactly those with cross-covariance depending on N₁₂: H₂→H₁ with R(N₁₂)⊂N(B₁₁^{1/2}A₁₁^{1/2}), N₁₂*N₁₂ ⪯ B/A; uniqueness iff B/A=0
    Load-bearing for the centroid claim (N₁₂=0 is not Monge when B/A≠0) and for Theorem 3.9.
  • domain assumption Polynomial spectral decay model s_k = c k^{−α} (α>1) for the non-parametric rate theorems (Corollary 3.5.2, Corollary 3.11) and Volterra singular-value asymptotics σ_k(G_n) ≍ (πk)^{−n} in Example 3.1
    Exact rates are derived only under this stylized assumption; real integrated Brownian motion operators satisfy it asymptotically.
  • standard math Standard operator theory: spectral theorem/SOT functional calculus, von Neumann trace inequality, Douglas factorization, Feldman–Hájek dichotomy, Fredholm determinants (Sections 2.1 and 3.1)
    Background results; usage is standard and internally consistent.
invented entities (2)
  • Properly aligned Green's operator pair (G,M)
    purpose: Coordinate-free, degeneracy-robust square-root factorization of A and B making G*M non-negative definite and the entropic objective spectrally diagonal
    Internal construction (imported from [40]); its validity is judged by internal consistency and by matching known finite-dim closed forms, not by an external falsifiable handle.
  • Green's correlation operator R = G†C(M*)†
    purpose: Encodes the cross-covariance of a Gaussian coupling relative to any Green's pair; constrained to F(G,M) for finite entropy
    Bookkeeping device from the Douglas factorization; no independent evidence — the factorization C = G R M* is directly verifiable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spectral Shrinkage of Gaussian Entropic Optimal Transport." pith.science (2026). https://pith.science/paper/N2AOV6DK

@misc{pith2026251219457,
  author       = {Pith},
  title        = {Pith review of: Spectral Shrinkage of Gaussian Entropic Optimal Transport},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N2AOV6DK}},
  note         = {Machine review of arXiv:2512.19457}
}
abstract

We present a functional calculus treatment of Entropic Optimal Transport (EOT) between Gaussian measures on separable Hilbert spaces, providing a unified framework that handles infinite-dimensional degeneracy. By leveraging the notion of proper alignment and the Schur complement, we reveal that the Gaussian EOT solution operates as a precise \textit{spectral shrinkage}: the optimal coupling is uniquely determined by contracting the spectrum of the correlation operator via a universal scalar function. This geometric insight facilitates an algorithmic shift from iterative fixed-point schemes (e.g., Sinkhorn) to direct algebraic computation, enabling efficient multi-scale analysis, where a single spectral decomposition allows for the exact evaluation of entropic costs across arbitrary regularization parameters $\varepsilon > 0$ at negligible additional cost. Furthermore, we investigate the asymptotic behavior as $\varepsilon \downarrow 0$ in settings where the unregularized Optimal Transport problem admits non-unique solutions. We establish a selection principle that the regularized limit converges to the most diffusive optimal coupling --characterized as the centroid of the convex set of optimal Kantorovich plans. This demonstrates that in degenerate regimes, the entropic limit systematically rejects deterministic Monge solutions (extremal points) in favor of the optimal solution with minimal Hilbert-Schmidt correlation, effectively filtering out spurious correlations in the null space. Finally, we derive stability bounds and convergence rates, recovering established parametric rates ($\varepsilon \log(1/\varepsilon)$) in finite dimensions while identifying distinct non-parametric rates dependent on spectral decay in infinite-dimensional settings.

Figures

Figures reproduced from arXiv: 2512.19457 by the authors.

Figure 1
Figure 1. (a) The spectral shrinkage function fε(x) for varying regularization strengths. As ε → 0, the function approaches the indicator function f0(x) = I(x > 0), forcing the correlation of all positive eigenmodes to 1, whereas larger ε dampens the correlations, effectively blurring the transport plan. (b) The logarithmic divergence of the scaled entropic bias gε(x)/ε = g1(x/ε) ≈ log(x/ε) (see Lemma A.2) near the origin hig… view at source ↗
Figure 2
Figure 2. Behavior of the perturbation function. (a) [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. The ratio W2 2 (πε, π)/ε as a function of the regularization parameter ε (on a reverse logarithmic scale) for dimensions d ∈ {10, 50, 200, 500}. For each dimension, we perform 5 independent trials with randomly generated full rank covariance matrices A, B ∈ R d×d . The limiting rate appears to scale linearly with dimension, concentrating around ≈ 0.36d for large d, which is strictly lower than the theoretical baseli… view at source ↗

Discussion (0). Continue with ORCID 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

  1. Entropic optimal transport need not select a zero-temperature limit

    math.OC 2026-07 accept novelty 8.0 of 10

    Entropic optimal-transport minimizers need not converge as ε→0 even for compact, atomless, bounded-Lipschitz costs; the cluster set can be an interval of optimal plans.

Reference graph

Works this paper leans on

48 extracted references · 3 linked inside Pith · cited by 1 Pith paper

  1. [40]

    Villani, C. et al. (2008). Optimal transport: old and new 338. Springer

    2008

  2. [1]

    Baker, C. R. (1970). Mutual information for Gaussian processes. SIAM Journal on Applied Math- ematics 19 451–458

    1970

  3. [2]

    and Lim, Y

    Bhatia, R., Jain, T. and Lim, Y. (2019). On the Bures–Wasserstein distance between positive definite matrices. Expositiones Mathematicae 37 165–191

    2019

  4. [3]

    Bogachev, V. I. (1998). Gaussian measures 62. American Mathematical Soc

    1998

  5. [4]

    and Krause, A

    Bunne, C., Hsieh, Y.-P., Cuturi, M. and Krause, A. (2023). The schr¨ odinger bridge between gaussian measures has a closed form. In International Conference on Artificial Intelligence and Statistics 5802–5833. PMLR

    2023

  6. [5]

    and Schmitzer, B

    Carlier, G., Duval, V., Peyr´e, G. and Schmitzer, B. (2017). Convergence of entropic schemes for optimal transport and gradient flows. SIAM Journal on Mathematical Analysis 49 1385–1418

    2017

  7. [6]

    and Tamanini, L

    Carlier, G., Pegon, P. and Tamanini, L. (2023). Convergence rate of general entropic optimal transport costs. Calculus of Variations and Partial Differential Equations 62 116

    2023

  8. [7]

    , Georgiou, T

    Chen, Y. , Georgiou, T. T. and Pavon, M. (2021). Optimal transport in systems and control. Annual Review of Control, Robotics, and Autonomous Systems 4 89–113

    2021

Show all 48 references
  1. [8]

    and Peyr´e, G

    Chizat, L., Roussillon, P., L´eger, F., Vialard, F.-X. and Peyr´e, G. (2020). Faster Wasser- stein distance estimation with the Sinkhorn divergence. Advances in neural information processing systems 33 2257–2269

    2020

  2. [9]

    Conway, J. B. (2019). A course in functional analysis 96. Springer

    2019

  3. [10]

    Csiszar, I. (1975). I-Divergence Geometry of Probability Distributions and Minimization Problems. The Annals of Probability 3 146 – 158

    1975

  4. [11]

    Cuturi, M. (2013). Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems 26

    2013

  5. [12]

    Da Prato, G. (2006). An introduction to infinite-dimensional analysis. Springer Science & Business Media

    2006

  6. [13]

    Douglas, R. G. (1966). On majorization, factorization, and range inclusion of operators on Hilbert space. Proceedings of the American Mathematical Society 17 413–415

    1966

  7. [14]

    Dryden, I. L. and Mardia, K. V. (2016). Statistical shape analysis: with applications in R . John Wiley & Sons

    2016

  8. [15]

    and Nutz, M

    Eckstein, S. and Nutz, M. (2022). Quantitative stability of regularized optimal transport and convergence of sinkhorn’s algorithm. SIAM Journal on Mathematical Analysis 54 5922–5948

    2022

  9. [16]

    , Georgakis, N

    Freulon, P. , Georgakis, N. and Panaretos, V. (2025). Entropic optimal transport beyond product reference couplings: the Gaussian case on Euclidean space.arXiv preprint arXiv:2507.01709

    2025 arXiv

  10. [17]

    Gelbrich, M. (1990). On a formula for the L2 Wasserstein metric between measures on Euclidean and Hilbert spaces. Mathematische Nachrichten 147 185–203

    1990

  11. [18]

    and Peyr´e, G

    Genevay, A., Chizat, L., Bach, F., Cuturi, M. and Peyr´e, G. (2019). Sample complexity of sinkhorn divergences. In The 22nd international conference on artificial intelligence and statistics 1574–1583. PMLR

    2019

  12. [19]

    and Nutz, M

    Ghosal, P. and Nutz, M. (2025). On the convergence rate of Sinkhorn’s algorithm. Mathematics of Operations Research. 20 Spectral Shrinkage of Gaussian Entropic Optimal Transport

    2025

  13. [20]

    , Nutz, M

    Ghosal, P. , Nutz, M. and Bernton, E. (2022). Stability of entropic optimal transport and Schr¨ odinger bridges.Journal of Functional Analysis 283 109622

    2022

  14. [21]

    and Krupnik, N

    Gohberg, I., Goldberg, S. and Krupnik, N. (2012). Traces and determinants of linear operators

    2012

  15. [22]

    and Krein, M

    Gohberg, I. and Krein, M. G. (1978). Introduction to the theory of linear nonselfadjoint operators

    1978

  16. [23]

    American Mathematical Soc

  17. [24]

    and Greenewald, K

    Goldfeld, Z. and Greenewald, K. (2020). Gaussian-smoothed optimal transport: Metric struc- ture and statistical efficiency. In International Conference on Artificial Intelligence and Statistics 3327–3337. PMLR

    2020

  18. [25]

    Hairer, M. (2009). An introduction to stochastic PDEs. arXiv preprint arXiv:0907.4178

    2009 arXiv

  19. [26]

    Hall, B. C. (2013). Quantum theory for mathematicians 267. Springer Science & Business Media

    2013

  20. [27]

    , Peyr´e, G

    Janati, H., Muzellec, B. , Peyr´e, G. and Cuturi, M. (2020). Entropic optimal transport be- tween unbalanced gaussian measures has a closed form. Advances in neural information processing systems 33 10468–10479

    2020

  21. [28]

    L´eonard, C. (2012). From the Schr¨ odinger problem to the Monge–Kantorovich problem. Journal of Functional Analysis 262 1879–1920

    2012

  22. [29]

    and Minh, H

    Mallasto, A., Gerolin, A. and Minh, H. Q. (2022). Entropy-regularized 2-Wasserstein distance between Gaussian measures. Information Geometry 5 289–323

    2022

  23. [30]

    , Panaretos, V

    Masarotto, V. , Panaretos, V. M. and Zemel, Y. (2019). Procrustes metrics on covariance operators and optimal transportation of Gaussian processes. Sankhya A 81 172–213

    2019

  24. [31]

    Minh, H. Q. (2023). Entropic regularization of Wasserstein distance between infinite-dimensional Gaussian measures and Gaussian processes. Journal of Theoretical Probability 36 201–296

    2023

  25. [32]

    and Wiesel, J

    Nutz, M. and Wiesel, J. (2022). Entropic optimal transport: Convergence of potentials.Probability Theory and Related Fields 184 401–424

    2022

  26. [33]

    Pal, S. (2019). On the difference between entropic cost and the optimal transport cost. arXiv preprint arXiv:1905.12206

    2019 arXiv

  27. [34]

    M.and Zemel, Y

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

    2020

  28. [35]

    Paulsen, V. I. and Raghupathi, M. (2016). An introduction to the theory of reproducing kernel Hilbert spaces 152. Cambridge university press

    2016

  29. [36]

    Peyr´e, G., Cuturi, M. et al. (2019). Computational optimal transport: With applications to data science. Foundations and Trends® in Machine Learning 11 355–607

    2019

  30. [37]

    , Aston, J

    Pigoli, D. , Aston, J. A. , Dryden, I. L. and Secchi, P. (2014). Distances and inference for covariance operators. Biometrika 101 409–422

    2014

  31. [38]

    Takatsu, A. (2011). Wasserstein geometry of Gaussian measures. Osaka Journal of Mathematics 48 1005 – 1026

    2011

  32. [39]

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

    2021

  33. [41]

    and Zemel, Y

    Yun, H. and Zemel, Y. (2025). Gaussian Optimal Transport Beyond Brenier’s Theorem. arXiv preprint. 21 Ho Yun Appendix A: Lemmas Lemma A.1. Let G, M ∈ B2(H). If {Tn}n is a sequence of bounded operators such that Tn → T in the SOT, then the product converges in the trace norm: G...

    2025

  34. [42]

    For the other limit, let u = 1/(2x)

    For the limits, first observe that: lim x↓0 g1(x) 2x = lim x↓0 (1 − f1(x)) − log(1 − f1(x)2) 2x = 1 − lim x↓0 −f1(x)2 2x = 1 − 0 = 1. For the other limit, let u = 1/(2x). As x → ∞, u → 0. We can rewrite f1(x) as: f1(x) = 1/up 1/u2 + 1 + 1 = 1√ 1 + u2 + u = p 1 + u2 − u. 22 Spe...

  35. [43]

    For any ε >0, δε : [0, ∞) → [0, 2 − √

  36. [44]

    is monotonely decreasing in x with limx↓0 δε(x) = 2 − √

  37. [45]

    For fixed x >0, δε(x) ↓ 0 as ε ↓ 0

  38. [46]

    The limiting behavior of ψ is given by lim u↓0 ψ(u) = 1 4 , lim u↑∞ ψ(u) = 0

    For any ε, x >0, we have the scaling property: xδε(x) ε = ψ ε 2x , where ψ : (0, ∞) → 0, 1 4 , ψ (u) := 1 2u 2 − r 2 1 + p 1 + u2 − u ! , u > 0, is a monotonely decreasing function. The limiting behavior of ψ is given by lim u↓0 ψ(u) = 1 4 , lim u↑∞ ψ(u) = 0. Proof. 1. This is...

  39. [47]

    We rewrite the scaled function solely in terms of u: xδε(x) ε = 1 2u 2 − r 2 1 + p 1 + u2 − u ! =: ψ(u)

    Let u := ε 2x . We rewrite the scaled function solely in terms of u: xδε(x) ε = 1 2u 2 − r 2 1 + p 1 + u2 − u ! =: ψ(u). For the continuity of ψ(u), observe that r 2 1 + p 1 + u2 − u = p 4 − 2u + O(u2) = 2 − u 2 + O(u2), thus lim u↓0 ψ(u) = lim u↓0 u 2 + O(u2) 2u = 1 4 . Then,...

  40. [48]

    once again, it remains to show that ˜ψ(y) = ψ(u) = (y2 − 1) y2(2 − y2) 2 − √ 2y = √ 2(y2 − 1) y2( √ 2 + y) 23 Ho Yun is monotonely increasing in y. By taking the logarithmic derivative with respect to y: d dy log ˜ψ(y) = 2y y2 − 1 − 2 y − 1√ 2 + y = 2 y(y2 − 1) − 1√ 2 + y the ...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.