Lifted Schr\"odinger Bridges for Gaussian Mixture Endpoints: Projection Gaps and Path-Space Obstructions

George Rapakoulias; Panagiotis Tsiotras; Siddhartha Ganguly

arxiv: 2605.24795 · v1 · pith:KP7ZH6URnew · submitted 2026-05-24 · 🧮 math.OC · cs.LG· cs.RO· cs.SY· eess.SY

Lifted Schr\"odinger Bridges for Gaussian Mixture Endpoints: Projection Gaps and Path-Space Obstructions

Siddhartha Ganguly , George Rapakoulias , Panagiotis Tsiotras This is my paper

Pith reviewed 2026-06-30 00:22 UTC · model grok-4.3

classification 🧮 math.OC cs.LGcs.ROcs.SYeess.SY

keywords Schrödinger bridgeGaussian mixturelifted constructionprojection gapSinkhorn scalingstochastic controlBrownian motionentropic optimal transport

0 comments

The pith

Augmenting Schrödinger bridges with component labels for Gaussian mixtures creates a projection gap that prevents recovery of the direct unlabeled solution.

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

The paper studies stochastic density control between Gaussian-mixture endpoints under Brownian dynamics. Since direct bridges lack closed form, it lifts each trajectory with source-target labels to decompose into explicit component bridges and a Sinkhorn entropic coupling. Projecting by forgetting labels satisfies the endpoints but incurs a nonnegative relative-entropy gap from lost label information. This gap shows the lifted optimizer is generally not the same as the direct unlabeled bridge. The work also gives the averaged drift, a kinetic bound, and a vanishing-gap condition.

Core claim

We introduce a lifted path-space construction in which each trajectory is augmented with a source--target component label. Consequently, the problem decomposes into Gaussian component-to-component Schrödinger bridges with explicit marginal, drift, and cost formulas, while the mixture-level assignment reduces to a finite-dimensional entropic coupling problem with a Sinkhorn scaling form. The projected law satisfies the original Gaussian-mixture endpoint constraints, but its relative entropy generally differs from the lifted relative entropy by a nonnegative conditional label-information gap. This gap reveals a path-space obstruction: the lifted optimizer cannot, in general, be identified with

What carries the argument

The lifted path-space construction augmenting trajectories with source-target component labels, which decomposes the mixture problem into component bridges and a finite-dimensional entropic coupling solved by Sinkhorn scaling.

If this is right

The projected marginal flow has an explicit posterior-averaged Markov drift.
The relative entropy of the projected law differs from the lifted relative entropy by the nonnegative label-information gap.
A kinetic-energy upper bound holds for the projected process.
The projection gap vanishes under a common path-potential condition.
Numerical illustrations demonstrate density and shape control under the lifted construction.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

One may need to retain labels to achieve the true minimal entropy control for mixture endpoints rather than solving the unlabeled problem directly.
Similar projection gaps may appear in other labeled extensions of optimal transport or stochastic control problems.
Numerical comparison of lifted and direct solutions for low-component mixtures could quantify the typical size of the gap.
The Sinkhorn coupling at mixture level may connect to assignment problems in other stochastic control settings with discrete choices.

Load-bearing premise

The assumption that the mixture problem decomposes into independent component-to-component Schrödinger bridges with the assignment solved by a finite-dimensional entropic coupling via Sinkhorn scaling.

What would settle it

Computing the relative entropy for both the projected lifted bridge and a direct numerical approximation of the unlabeled bridge on a two-component Gaussian mixture and checking whether they coincide or differ.

Figures

Figures reproduced from arXiv: 2605.24795 by George Rapakoulias, Panagiotis Tsiotras, Siddhartha Ganguly.

**Figure 1.** Figure 1: Roadmap of our lifted Schrödinger bridge construction. perspective. Rather than treating a Gaussian-mixture bridge only as a computational superposition of pairwise Gaussian bridges, we formulate an augmented entropy-minimization problem on a label–trajectory space, in which the source–target component assignment is itself a probabilistic object. In this formulation, the prior component coupling 𝜂 specif… view at source ↗

**Figure 2.** Figure 2: One-dimensional Gaussian-mixture Schrödinger bridge experiment. We choose the prior coupling to be 𝜂 prod 𝑖 𝑗 := 𝛼 0 𝑖 𝛼 1 𝑗 . Solving the lifted finite-dimensional problem by Sinkhorn scaling (with 100 Sinkhorn iterations) yields 𝜋 ★ = 0.35 0.30 0 0.35 , where entries below 10−5 are displayed as zero. The row and column sums satisfy Í 𝑗 𝜋 ★ 𝑖 𝑗 = 𝛼 0 𝑖 , and Í 𝑖 𝜋 ★ 𝑖 𝑗 = 𝛼 1 𝑗 , as required. The inte… view at source ↗

**Figure 3.** Figure 3: The evolution of the projected density 𝜌 𝜋 ★ 𝑡 (·) at several timeslices (left subfigure) and sample paths of the Markov diffusion under Euler-Maruyama steps (right subfigure) [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗

**Figure 4.** Figure 4: The evolution of the sample paths of the Markov diffusion under Euler-Maruyama steps (left subfigure) and the corresponding density evolution (right subfigure) at different time-slices. left subfigure in [PITH_FULL_IMAGE:figures/full_fig_p028_4.png] view at source ↗

**Figure 5.** Figure 5: The crescent source and star target shapes and their GMM approximations (left subfigure). The sample paths of the Markov diffusion under Euler-Maruyama steps (right subfigure). § 4.2.3. Example 3 (Shape matching). As a final visual illustration, we consider an image-inspired density steering problem. A crescent-shaped source silhouette and a starshaped target silhouette are first converted into point clo… view at source ↗

**Figure 6.** Figure 6: shows the corresponding analytic projected density evolution 𝑥 ↦→ 𝜌 𝜋 ★ 𝑡 (𝑥) which deforms the crescent-shaped density into the star-shaped density [PITH_FULL_IMAGE:figures/full_fig_p029_6.png] view at source ↗

**Figure 7.** Figure 7: Effect of the prior component coupling in the shape-matching example. The top row shows the prescribed prior couplings 𝜂, and the bottom row shows the corresponding optimized couplings 𝜋 ★ 𝜂 . The product prior is neutral, the diagonal-favoring prior encourages 𝑖 ↦→ 𝑖 assignments, and the rotated prior encourages the shifted assignment 𝑗 = 𝑖+3 (mod 10). All entries represent fractions of total mass assigne… view at source ↗

read the original abstract

We study stochastic density control between Gaussian-mixture endpoint distributions under Brownian prior dynamics. Since the direct Schr\"odinger bridge between Gaussian mixtures is generally not available in closed form, we introduce a lifted path-space construction in which each trajectory is augmented with a source--target component label. Consequently, the problem decomposes into Gaussian component-to-component Schr\"odinger bridges with explicit marginal, drift, and cost formulas, while the mixture-level assignment reduces to a finite-dimensional entropic coupling problem with a Sinkhorn scaling form. We then analyze the projection obtained by discarding or forgetting the label. By construction, the projected law satisfies the original Gaussian-mixture endpoint constraints, but its relative entropy generally differs from the lifted relative entropy by a nonnegative conditional label-information gap. This gap reveals a path-space obstruction: the lifted optimizer cannot, in general, be identified with the direct unlabeled Schr\"odinger bridge after projection. We also derive the posterior-averaged Markov drift associated with the projected marginal flow, prove a kinetic-energy upper bound, and identify a common path-potential condition under which the projection gap vanishes. Several numerical illustrations showing density and shape control are recorded for a self-contained exposition.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note private letter to a colleague

The lifted construction gives explicit component-wise bridges and a feasible projection, but the claimed path-space obstruction to the direct Schrödinger bridge rests on a decomposition that probably fails to produce the global optimizer when Gaussians overlap.

read the letter

The paper's core move is to augment paths with discrete source-target labels so the mixture problem splits into separate Gaussian-to-Gaussian Schrödinger bridges (with closed-form drifts and costs) plus a finite-dimensional entropic assignment solved by Sinkhorn. That decomposition is new in this setting and lets them write down the projected marginal flow, an averaged Markov drift, a kinetic-energy bound, and a clean condition under which the label-information gap disappears. Those derivations are concrete and useful for anyone who needs a workable control law even if it is not exactly optimal.

The soft spot is the optimality claim. The abstract treats the lifted measure as the optimizer whose projection then differs from the direct unlabeled bridge. But when the endpoint Gaussians overlap, the relative entropy with respect to the Brownian reference does not factor cleanly across label pairs; the optimal drift at any point depends on the total density, which couples the components. The constructed measure satisfies the marginals and is feasible, yet it is not obviously the global minimizer in the lifted space. Without that, the reported gap does not establish an obstruction relative to the true direct bridge. The stress-test note flags exactly this issue, and nothing in the abstract resolves it.

The work is technically serious: it engages the literature on entropic OT and Schrödinger bridges, avoids circular definitions, and supplies explicit formulas rather than hand-waving. It is aimed at the stochastic control community that already works with mixture endpoints. A referee should check whether the decomposition is justified or whether an extra cross-term argument is needed; if the math holds after that clarification, the explicit constructions and the vanishing condition are worth having. I would send it to review rather than desk-reject.

Referee Report

1 major / 0 minor

Summary. The manuscript introduces a lifted path-space construction for the Schrödinger bridge problem between Gaussian mixture distributions under Brownian prior. Trajectories are augmented with source-target component labels, allowing decomposition into explicit component-to-component bridges and a Sinkhorn-solved assignment problem. The projection of the lifted optimizer satisfies the mixture marginals but differs in relative entropy from the direct bridge by a label-information gap, establishing a path-space obstruction. Additional results include the posterior-averaged drift, a kinetic-energy bound, and a vanishing condition for the gap, supported by numerical examples.

Significance. If the decomposition holds, the work offers a computationally tractable approach to Schrödinger bridges for mixtures via closed-form component solutions and finite-dimensional optimization, while highlighting an intrinsic obstruction in projecting labeled optima. The explicit drift derivation and vanishing condition provide concrete insights. The numerical illustrations demonstrate practical density and shape control. These contributions advance understanding of entropic optimal transport in mixture settings.

major comments (1)

[Abstract] Abstract (paragraph beginning 'Consequently, the problem decomposes'): The assertion that the lifted problem decomposes into independent component-to-component Schrödinger bridges assumes that relative entropy separates across label pairs with no cross terms. Because the endpoint Gaussians overlap, the Brownian reference measure does not factorize cleanly with discrete labels; the optimal drift depends on the total evolving density, which can induce path-space correlations across components not captured by independent bridges. The constructed measure is feasible for the original marginals but need not be the global minimizer in the lifted space, so the reported label-information gap does not necessarily demonstrate an obstruction relative to the true direct Schrödinger bridge. A derivation establishing separation (or quantifying cross terms) is required to support the central claim.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and the insightful comment on the decomposition claim. We address the point directly below and will strengthen the manuscript accordingly.

read point-by-point responses

Referee: [Abstract] Abstract (paragraph beginning 'Consequently, the problem decomposes'): The assertion that the lifted problem decomposes into independent component-to-component Schrödinger bridges assumes that relative entropy separates across label pairs with no cross terms. Because the endpoint Gaussians overlap, the Brownian reference measure does not factorize cleanly with discrete labels; the optimal drift depends on the total evolving density, which can induce path-space correlations across components not captured by independent bridges. The constructed measure is feasible for the original marginals but need not be the global minimizer in the lifted space, so the reported label-information gap does not necessarily demonstrate an obstruction relative to the true direct Schrödinger bridge. A derivation establishing separation (or quantifying cross terms) is required to support the central

Authors: We agree that an explicit derivation of the separation is needed to fully support the claim. In the lifted construction the reference measure on path-label space is the mixture, over source-target label pairs (i,j), of the Brownian motion started from component i and ended at component j; labels are sampled from the discrete mixture weights independently of the path under the reference. This factorization implies that the relative entropy KL(P_lifted || R_lifted) decomposes exactly as the π-weighted sum of the individual component-bridge relative entropies plus the discrete KL on the label coupling. Consequently the optimization separates into independent Gaussian-to-Gaussian Schrödinger bridges (closed-form) and a finite-dimensional Sinkhorn problem; the optimal lifted drift is label-conditioned and does not involve the unlabeled total density. The projection onto unlabeled paths then produces the stated nonnegative label-information gap by the chain rule for relative entropy. We will insert this derivation as a new subsection in Section 2 of the revised manuscript, confirming both that the constructed measure is the global minimizer in the lifted space and that the gap constitutes a genuine obstruction to the direct bridge. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained construction on standard entropic OT

full rationale

The paper introduces a lifted construction augmenting paths with discrete labels, states that this yields independent Gaussian-to-Gaussian Schrödinger bridges plus a finite-dimensional Sinkhorn coupling (abstract and section on decomposition), then defines the projection gap as the nonnegative difference in relative entropy between the lifted optimizer and its label-forgetting projection. This gap is exhibited by direct comparison of the two measures and is not obtained by fitting any parameter to data or by renaming an input quantity. No equation reduces a claimed result to a self-referential definition, a fitted input called a prediction, or a load-bearing self-citation chain. The argument relies on standard properties of relative entropy, Brownian reference measures, and entropic optimal transport, all treated as external. The central obstruction claim is therefore a comparison between explicitly constructed objects rather than an identity forced by the paper's own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The contribution rests on standard properties of relative entropy and the known closed-form Gaussian Schrödinger bridge; the new element is the label-augmented construction itself.

axioms (2)

standard math Standard properties of relative entropy and conditional mutual information
Invoked to quantify the nonnegative label-information gap between lifted and projected relative entropies
domain assumption Existence of explicit marginal, drift, and cost formulas for Gaussian-to-Gaussian Schrödinger bridges under Brownian dynamics
Required for the decomposition into component-to-component problems

invented entities (1)

Lifted trajectories augmented with source-target component labels no independent evidence
purpose: To decompose the Gaussian-mixture problem into tractable component bridges plus a finite entropic assignment
New construction introduced because the direct mixture bridge lacks closed form

pith-pipeline@v0.9.1-grok · 5768 in / 1431 out tokens · 46546 ms · 2026-06-30T00:22:17.440975+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages

[1]

Arapostathis, V

[ABG11] A. Arapostathis, V. S. Borkar, and M. K. Ghosh,Ergodic Control of Diffusion Processes, Encyclopedia of Mathematics and its Applications, Cambridge University Press, 2011, doi: https://doi.org/10.1017/CBO9781139003605. [AC24] D. O. Adu and Y. Chen,Schrödinger bridge over averaged systems, 2024, URL:https://arxiv. org/abs/2412.03294. [AEG26] D. O. A...

work page doi:10.1017/cbo9781139003605 2011
[2]

Springer-Verlag, Berlin, 2001, doi:https://doi.org/10.1007/978-3-662-13043-8. [MAJTC25] Y.Mei,M.Al-Jarrah,A.Taghvaei,andY.Chen,Flowmatchingforstochasticlinearcontrolsystems, Proceedings of the 7th Annual Learning for Dynamics & Control Conference (Necmiye Ozay, Laura Balzano, Dimitra Panagou, and Alessandro Abate, eds.), Proceedings of Machine Learning Re...

work page doi:10.1007/978-3-662-13043-8 2001
[3]

and YOR, M

Springer-Verlag, Berlin, 1999, doi:https://doi.org/10.1007/978-3-662-06400-9. [San15] F. Santambrogio,Optimal Transport for Applied Mathematicians: Calculus of Variations, PDEs, and Modeling, Progress in Nonlinear Differential Equations and their Applications, vol. 87, Birkhäuser/Springer, Cham, 2015, doi:https://doi.org/10.1007/978-3-319-20828-2. [San23]...

work page doi:10.1007/978-3-662-06400-9 1999

[1] [1]

Arapostathis, V

[ABG11] A. Arapostathis, V. S. Borkar, and M. K. Ghosh,Ergodic Control of Diffusion Processes, Encyclopedia of Mathematics and its Applications, Cambridge University Press, 2011, doi: https://doi.org/10.1017/CBO9781139003605. [AC24] D. O. Adu and Y. Chen,Schrödinger bridge over averaged systems, 2024, URL:https://arxiv. org/abs/2412.03294. [AEG26] D. O. A...

work page doi:10.1017/cbo9781139003605 2011

[2] [2]

Springer-Verlag, Berlin, 2001, doi:https://doi.org/10.1007/978-3-662-13043-8. [MAJTC25] Y.Mei,M.Al-Jarrah,A.Taghvaei,andY.Chen,Flowmatchingforstochasticlinearcontrolsystems, Proceedings of the 7th Annual Learning for Dynamics & Control Conference (Necmiye Ozay, Laura Balzano, Dimitra Panagou, and Alessandro Abate, eds.), Proceedings of Machine Learning Re...

work page doi:10.1007/978-3-662-13043-8 2001

[3] [3]

and YOR, M

Springer-Verlag, Berlin, 1999, doi:https://doi.org/10.1007/978-3-662-06400-9. [San15] F. Santambrogio,Optimal Transport for Applied Mathematicians: Calculus of Variations, PDEs, and Modeling, Progress in Nonlinear Differential Equations and their Applications, vol. 87, Birkhäuser/Springer, Cham, 2015, doi:https://doi.org/10.1007/978-3-319-20828-2. [San23]...

work page doi:10.1007/978-3-662-06400-9 1999