We study the numerical approximation of Schrödinger Bridges in the setting where samples from the target distribution are unavailable and only an unnormalized target density is given. In this regime, the widely used iterative proportional fitting (IPF) procedure (and its variants) often exhibits severe numerical instabilities, stemming from ill-conditioned underlying control problems and from the fact that the optimal solution may lie far from the initialization. To address these issues, we propose a trust region variant of IPF that constrains successive updates to remain at a fixed Kullback-Leibler divergence from the current iterate. This restriction significantly improves numerical stability. Moreover, we show that the resulting method admits an interpretation as a mirror descent scheme with an adaptively chosen step size. We demonstrate the effectiveness of our approach on several challenging sampling tasks.
Figures & tables
Figure 1
Figure 2: Overview of IPF-based samplers. The columns differ in how the mirror descent step size of the forward update is chosen. The rows show the path space update, which reweights the current iterate \mathbbmPui by wiγ ( Proposition 3.1 ), its implementation by matching objectives learned from samples of \mathbbmPui with this importance weight (IW), and the variants without it.
Figure 3: Results on Many-Well potential. We evaluate robustness to increasing modes and robustness to mode distance. (a, b) Performance vs. number of modes. (c, d) Performance vs. mode distance. Our proposed method (TR-SBS, red) achieves the lowest estimation error and marginal distance for large m and δ , handling both combinatorial complexity and high energy barriers.
Figure 4: Ablation study on the step size γ . Legend is shared across all subplots. (a) Adaptive TR-SBS matches the robustness of conservative γ=0.1 . (b, c) Dynamics reveal the trade-off: Small γ=0.1 is stable (b) but slow (c). TR-SBS converges efficiently (c) while maintaining stability (b). (d) The adaptive γ starts conservative to prevent collapse and increases to accelerate convergence.
Figure 5: Results on GMM (mean ± standard deviation over three seeds). (a, b) Performance vs. dimension d∈{10,25,50,100} with K=10 . (c, d) Performance vs. number of components K∈{5,10,20,30,40} with d=10 . TR-SBS (red) attains the lowest SD and TVD in every setting. PIS and DIS use 128 instead of 200 time steps and a different training budget ( Section B.6 ).
Figure 6: Alanine dipeptide. Left: Ramachandran plots of the reference, of TR-SBS after 120,000 epochs, and of ASBS with clipped corrector targets after 70,000 epochs ( Section C.6 ); the dashed box marks C7ax. Right: C7ax population over training (lines: evaluations of 2,000 samples, averaged over 2,000 epochs; dots: 107 samples, open dots from ASBS resumed at 48,500 epochs, see Section C.6 ; squares: ASBS with clipped corrector targets, 107 samples; cross: divergence of ASBS without clipping; dashed: reference).
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Algorithm 3 Trust Region Schrödinger Bridge Sampler (TR-SBS)
Algorithm 4 PopulateBuffer: Auxiliary routine for replay buffer construction
Parameter
GMM (Scale d )
GMM (Robust K )
MW (Scale m )
MW (Robust δ )
GL (Periodic)
σprior
5.0
30.0
1.0
1.0
1.0
σmax
1/2/1/2‡
1/2/2/4‡
0.2
0.1/0.2
0.2
σmin
0.1
0.1
0.01
0.01
0.01
ε
0.1
0.3
0.3
0.3
0.3
cIW
10
10
1
1
1
I
≤150†
≤150†
50
35
80
Appendix
Table 2: Unified hyperparameters for the GMM, MW, and GL experiments (alanine dipeptide: Section B.4 ; LJ-13: Section C.7 ). For MW, values separated by a slash (e.g., 0.1/0.2 ) indicate that both values were evaluated for each specific target configuration, and the one yielding better performance was selected.
Figure 7: Additional metrics on GMM (mean ± standard deviation over three seeds). (a)-(c): K=10 , varying d . (d)-(f): d=10 , varying K . Lower is better for W2 (a, d) and sliced W2 (b, e); higher is better for the raw path ESS (c, f).
Figure 8: Samples on GMM with d=10 and K=40 (first two coordinates, seed 0). Each panel shows the 4,000 evaluation samples of the checkpoint with the lowest SD during training. Blue points show generated samples; contours show the target marginal log density. All panels use identical contours and coordinate limits and show every sample. PIS and DIS use 128 instead of 200 time steps ( Section B.6 ).
Figure 9: Visualization at small noise ( σmax=0.1 ). Rows (a)–(d) show the four methods; columns show the marginal histograms of dimensions 0–2 and the energy distribution. Due to insufficient noise, all methods (a-d) fail to reach the wells (underfitting).
Figure 10: Visualization at large noise ( σmax=0.2 ). Rows and columns as in Figure 9 . Aggressive noise facilitates exploration but causes instability. (a) ASBS suffers from mode collapse, and (b) ASBS-IW misbalances the modes. (c) TR-ASBS prevents total collapse but lacks fidelity. (d) Only TR-SBS successfully captures the global structure and energy distribution.
Figure 11: Ablation study on the maximum noise scale σmax . Legend is shared across subplots (a-c). (a) Large σmax=0.2 causes mode collapse in ASBS (blue) as barrier distance δ increases, whereas TR-SBS (red) remains robust. (b) Training dynamics confirm that ASBS diverges in EMD while TR-SBS remains stable. (c) A larger σmax leads to faster convergence. (d) Histograms confirm that ASBS collapses to a single mode under large noise, while TR-SBS maintains the global structure.
NFC
NBC
Marginal EMD ↓
∣ΔlogZ∣↓
10
100
0.294
200.694
100
10
0.215
2.455
100
100
0.131
0.006
100
300
0.191
546.959
300
100
0.212
99.551
Appendix
Table 3: Sensitivity of TR-SBS to NFC and NBC on MW ( d=8 , m=5 , δ=4 ). The setting 100/100 is the main result averaged over three seeds; the other settings use a single seed.
Method
ESS ↑
W2↓
ASBS
0.005
1.78
TR-SBS
0.445
0.15
Appendix
Table 4: Results on the discretized Ginzburg-Landau model.
Method
Epoch
KL ϕ
KL ψ
KL γ1
KL γ2
KL γ3
W2
C7ax (%)
αL (%)
ASBS
31k
0.2784
0.0243
0.0032
0.0158
0.0263
0.6659
1.5352
1.2509
TR-SBS
33k
0.1862
0.0306
0.0193
0.0210
0.0300
0.3205
0.3342
0.3646
ASBS
48.5k
0.0256
0.0132
0.0036
0.0159
0.0262
0.4922
0.0788
0.0837
TR-SBS
48.5k
0.0570
0.0093
0.0074
0.0162
0.0264
0.2700
0.3500
0.2259
ASBS, resumed
54k
0.0457
0.0276
0.0032
0.0157
0.0262
0.5744
0.0338
0.0746
ASBS (diverged)
58.5k
0.2967
0.3786
0.8315
0.0333
0.0423
1.7838
0.1927
0.3708
Appendix
Table 5: Alanine dipeptide: metrics ( 107 samples) of TR-SBS and ASBS with the same prior at similar stages of training and at the end of training. Only some checkpoints of each run were saved and evaluated with 107 samples, so paired rows show the closest such checkpoints, whose epochs differ by up to 6,500 . ASBS diverges between 54,000 and 54,500 epochs; the row at 54,000 epochs is from the run resumed at 48,500 epochs, and the row at 58,500 epochs is after the divergence. The rows “ASBS, clipped” are from the run resumed at 48,500 epochs with clipped corrector targets ( Section C.6 ). Reference populations: C7ax 0.169% , αL0.049% .
Figure 12: Alanine dipeptide: populations of C7ax (left) and αL (right) over training for TR-SBS and ASBS with the same prior. Lines: evaluations of 2,000 samples, averaged over 2,000 epochs; dots: 107 samples, open dots from ASBS resumed at 48,500 epochs; squares: ASBS with clipped corrector targets, resumed at 48,500 epochs, 107 samples, up to 70,000 epochs; dashed: reference; cross: divergence of ASBS without clipping, after which its populations are not shown (see Table 5 ).
Method
W2↓
E(⋅)W2↓
ASBS
1.571±0.002
0.94±0.07
TR-SBS
1.585±0.001
0.82±0.11
ASBS, reported in [ 30 ]
1.59±0.03
1.99±1.01
Reference vs. reference samples
1.32±0.01
0.39±0.04
Appendix
Table 6: LJ-13: mean ± standard deviation over three evaluation trials of 2,000 samples.
The Schrödinger bridge problem is concerned with finding a stochastic dynamical system bridging two marginal distributions that minimises a certain transportation cost. This problem, which represents a generalisation of optimal transport to the stochastic case, has received attention due to its connections to diffusion models and flow matching, as well as its applications in the natural sciences. However, all existing algorithms allow to infer such dynamics only for cases where samples from both distributions are available. In this paper, we propose the first general method for modelling Schrödinger bridges when one (or both) distributions are given by their unnormalised densities, with no access to data samples. Our algorithm relies on a generalisation of the iterative proportional fitting (IPF) procedure to the data-free case, inspired by recent developments in off-policy reinforcement learning for training of diffusion samplers. We demonstrate the efficacy of the proposed data-to-energy IPF on synthetic problems, finding that it can successfully learn transports between multimodal distributions. As a secondary consequence of our reinforcement learning formulation, which assumes a fixed time discretisation scheme for the dynamics, we find that existing data-to-data Schrödinger bridge algorithms can be substantially improved by learning the diffusion coefficient of the dynamics. Finally, we apply the newly developed algorithm to the problem of sampling posterior distributions in latent spaces of generative models, thus creating a data-free image-to-image translation method. Code: https://github.com/mmacosha/d2e-stochastic-dynamics
The Schrödinger Bridge Problem connects an initial distribution to a terminal one along a minimum-energy stochastic process. Its mean-field extension, the Mean-Field Schrödinger Bridge, governs interacting populations whose dynamics and costs depend on the collective distribution. When these interactions are nonlocal, their direct evaluation scales quadratically with the population size, making large ensembles intractable within FBSDE-based solvers. We replace these terms with neural surrogates in state and time, trained on empirical interaction values along sampled trajectories and embedded in a four-stage alternating scheme that updates the forward and backward potentials and the surrogates in turn, while preserving forward--backward consistency and the prescribed endpoint marginals. We derive Grönwall-type stability bounds quantifying how surrogate errors propagate to the generated trajectories under a small-gain condition. On crowd-navigation and high-dimensional opinion-dynamics benchmarks, the surrogates reproduce the trajectories obtained with exact evaluation at reduced training cost. The advantage is most significant when the interaction is a nonlinear functional of the measure, such as the normalized bounded-confidence drift, for which random-batch subsampling is biased and unstable whereas the learned surrogate remains accurate.
Daisuke Inoue, Dante Kalise, Mathieu Laurière
Department of Mathematics, Imperial College London, London SW7 2AZ, United Kingdom. · Shanghai Frontiers Science Center of Artificial Intelligence and Deep Learning, NYU-ECNU Institute of Mathematical Sciences, NYU Shanghai, Shanghai 200126, People’s Republic of China.
Over the past few years, diffusion-based Schrödinger bridge models have been proposed to approximate optimal transport dynamics between two prescribed boundary distributions, with successful applications to generative modeling. More precisely, these methods aim to estimate a path measure whose initial and terminal marginals match the two boundary distributions, while minimizing the Kullback-Leibler divergence with respect to a reference Markov process. In this work, we consider the generalized Schrödinger bridge problem, in which the reference process is a twisted Brownian motion, that is, a Feynman-Kac transform of a Brownian motion induced by a time-dependent differentiable potential. Building on the Iterative Markovian Fitting (IMF) paradigm, and in particular on its special case Diffusion Schrödinger Bridge Matching (DSBM), which corresponds to the zero potential case, we introduce Twisted Schrödinger Bridge Matching (TSBM), a diffusion-based method designed to handle both continuous- and discrete-time potentials. Unlike previous approaches, TSBM provides a rigorous extension of the IMF scheme to the generalized Schrödinger bridge problem. This derivation leads to a new bridge-matching loss that depends explicitly on the gradient of the potential and recovers the DSBM objective when the potential vanishes, yielding improved performance. We further introduce trajectory-based variance-reduction techniques that substantially stabilize optimization and may be useful beyond the present setting. Finally, we empirically demonstrate the benefits of TSBM for trajectory inference across increasingly high-dimensional settings, including crowd navigation and single-cell data. Code available at https://github.com/maxencenoble/twisted-sb-matching.
Maxence Noble, Marie Scheid, Yazid Janati +2
CMAP, Ecole polytechnique · Institute of Foundation Models · MBZUAI +1