acceptodds
Under review as a conference paper at ICLR 2027

Elysium: Simulation-Free Unbalanced Schrödinger Bridges

12 views

Abstract

Reconstructing cellular dynamics from snapshots is challenging because the destructive nature of single-cell measurements makes tracking the same cells over time impossible. Cellular dynamics are inherently stochastic and involve changes in population size through proliferation and apoptosis. Growth-based methods represent population change as a rate on the density, which does not assign exits and entries to individual cells. Unbalanced Schrödinger bridges model individual events, but neural solvers for killed diffusion bridges simulate trajectories during training. We introduce ELYSIUM, a simulation-free solver that jointly learns state transition and mass change. Using a killed Brownian reference and its time reversal, ELYSIUM constructs direct regression objectives from an optimal endpoint coupling and conditional bridge distributions, without simulating trajectories during training. We show that the minimizers of the expected training objectives recover the optimal unbalanced Schrödinger bridge. Empirically, ELYSIUM reconstructs state transitions and population changes on simulated data with known killed diffusion dynamics, outperforming all compared methods. On real single-cell datasets, it recovers cell state distributions at unobserved time points and captures population mass changes, achieving state-of-the-art results in most comparisons.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
optimal transport, Schrödinger bridge, trajectory inference, single-cell

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_XRANOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors proposed a simulation-free method to handle both entering and existing in an unbalanced SB problem under killed BM reference with Poisson process on killing event. The main idea is to have a lifted cemetery state and reduce things to balanced SB with this somewhat data augmentation approach, and simulation free could be handled by properties of the BM reference.

Critical strengths
  • There is no need to forward or backward simulate any SDEs during training.
  • Handles entering and exiting naturally.
  • Relatively well written
Critical weaknesses

I consider these not to be critical weakness for rejection but unidentified limitations worth discussing. All related to previously identified strengths.

  • I think the simulation free nature relies on BM reference with homogeneous waiting time for killing event. It might be extended to linear reference though.
  • I have a bit difficulty mapping entering and exiting to biological processes. See question for more details.
Questions for authors
  • I suspect the method can be extended to handle linear drift in the reference. The difficulty is usually not the bridge conditioning on matched particles but solving the static coupling problem. This maybe done by things like Zhang et al. 2025. What do the authors think?
  • Is a homogeneous exponential waiting time necessary? How much of the simulation-free nature relies on the existing having an exponential waiting time across all time? Is the exponential waiting time also needed so that time reversal makes physical sense?
    • A subquestion is that I think the existing event have be memoryless (o.w. time reversal will definitely be problematic, as the particles will have to remember the future). Does this imply we can only do exponential waiting time?
  • Re: exiting and entering. Unlike balanced SB, the method could have some particles from tit_{i} match with multiple particles in ti+1t_{i+1} through branching or match to none of the particles through exiting, both handled in lifted, cemetery state. (Please correct me if I misunderstood this part). I can understand it mathematically but I have some difficulty mapping it back into biology: all the cells are different cells as they are killed to see what is inside. So matching is a modeling need, not what happened physically. How should we understand the entering and existing physically/biologically?

Zhang, S. and Stumpf, M., 2026. Learning non-equilibrium diffusions with schrödinger bridges: from exactly solvable to simulation-free. Advances in Neural Information Processing Systems, 38, pp.119861-119898.

Reviewer_QdLkOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The author presented a simulation-free neural solver for an existing killed-diffusion Schrödinger bridge problem.

Critical strengths
  1. Analytic training targets avoid trajectory simulation during training.
  2. Thorough proofs.
Critical weaknesses
  1. The biological meaning of the population masses is unclear. Sampled cell counts are treated as population sizes. Differences could also come from sampling or capture efficiency. How is this justified for each dataset?
  2. Each interval allows only gain or loss. If endpoints μ0\mu_0 and μ1\mu_1 are equal, does that imply no events? Even when biological turnover may occur? This limits the claimed advantage over net-growth models.
Questions for authors
  1. Compared with Chen et al.’s killed SB and UDSB, is the main advance simulation-free training and analytic conditional targets? Please clarify which parts of the formulation are new, rather than inherited from those methods.
  2. How are κ,ν\kappa, \nu and the dataset-specific clocks selected? What justified these choices for the real datasets? Given the endpoint masses and reference clock, equation 30 specifies the target total-mass curve and aggregate event-time distribution before neural training. The sensitivity analysis on Simulation Gene is helpful, but how robust are the real-data results to alternative parameter and clock choices?
  3. For Q1 experiment, why do the baselines produce extra clusters absent from the target distribution? For WFR-FM and USB, how much of this difference comes from their different objectives? For UDSB-F, are the reference process and endpoint constraints identical to ELYSIUM’s, and does the performance gap remain with more training and finer numerical time steps?
Reviewer_sSEyOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper introduces simulation-free training algorithms for unbalanced Schrödinger bridges, where the population size of temporal snapshots from a given distributional dynamic (e.g., cellular dynamics) often changes. The work follows the standard pipeline of conditional flow matching where the flow (or dynamics) follows the fixed reference interpolated path for a given so-called “optimal” coupling which is inferred by extra optimal transport stages between snapshots. This paper mostly deals with how to incorporate killing/growth rate into those settings for unbalanced SB settings.

Critical strengths
  • Deriving a closed-form diffusion bridge for a given killed reference path is interesting and technically well-motivated.
  • The proposed methods seem to outperform the baseline they compared.
Critical weaknesses
  • The paper is extremely hard to follow. If I had to pick one, for example, at the Definition 1, the author defines KL-divergence that is similar to the KL-divergence used for typical Schrodinger Bridge formulation of the killed reference path measure R\mathbb{R}. But, this measure R\mathbb{R} is not denoted before, so I have to guess that the author referred to the “reference process” as whether R\mathbb{R} or not. Now, let’s say R\mathbb{R} is the “reference process” that the author defined in equation (1). For a given τ\tau, the particles propagate as a Wiener process or absorbed to a symmetry state. Then τ\tau is given by τ∼Exp(κ)\tau \sim Exp(κ). What does τ∼Exp(κ)\tau \sim Exp(κ) mean? Does it mean, the random (stopping time) τ\tau is sampled from some exponential distribution with intensity rate κ\kappa? Moreover, in the following paragraph, there are many similar ambiguities, making the explanation difficult to follow.

  • The ICLR guideline shows that “Appendix: “Authors may use as many pages of appendices (after the bibliography) as they wish, but reviewers are not required to read the appendix.”. However, the authors have placed essential technical details in the appendix that make it even harder to read. Which feels the current version of writing far below than the bar of the ICLR venues. I suggest re-writing the paper in-hand and resubmitting next venues.

  • The paper relies on a precomputed OT coupling, which makes the proposed method seem like an extension of [1], with a killing/birth rate included in the optimization and dynamics. Most standard SB formulations in this literature (e.g., [2-3] for two marginals, and [4-6] for multiple marginals) require forward-backward iterations to solve the SB problem without a predefined coupling, and the optimal coupling is obtained only after solving the problem, through simulated trajectories. However, the proposed method requires the optimal coupling by design, which might be computationally demanding to obtain for higher-dimensional problems. I wonder whether the proposed method can be extended to a forward-backward setting where the optimal coupling is unavailable.


[1] Tong et al., Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport
[2] Shi et al., Diffusion Schrödinger Bridge Matching
[3] Peluchetti et al., Diffusion Bridge Mixture Transports, Schrödinger Bridge Problems and Generative Modeling
[4] Theodoropoulos et al., Momentum Multi-Marginal Schrödinger Bridge Matching
[5] Gravier et al., Multi-marginal temporal Schrödinger Bridge Matching from unpaired data
[6] Park et al., Multi-marginal Schrödinger Bridge Matching
Questions for authors

See Weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.