Source
Shi, De Bortoli, Deligiannidis & Doucet (2022), “Conditional Simulation Using Diffusion Schrödinger Bridges”, UAI 2022, LIT-776, read in NOTE-589. Sections 1–8 closely; Appendix C (the proofs of Propositions 1–3) followed for structure only.
What was actually shown
The setting. One can sample X ∼ p_data and Y ∼ g(·|X), and g need not be evaluable. Densities are positive and the relevant KL divergences finite. The bridge for one observation cannot be built, because its posterior end cannot be sampled.
Proposition 1. The averaged problem, the expectation over Y of KL(π_Y | p_Y) under both marginal constraints, is solved by the ordinary Schrödinger bridge on X × Y between p(x, y) and p_ref(x|y)p(y), with the y-coordinate held constant. So one network conditioned on y serves every observation.
Proposition 2. The iterative-proportional-fitting iterates factor as p(y) times conditional forward and backward chains in x, so the diffusion-Schrödinger-bridge losses carry over with y as an input.
Proposition 3. For n ≥ 1, E[KL(π^{c,n}_{Y,0} | p(·|Y))] ≤ (2/n) E[KL(π^{c,*}_Y | p_Y)]. The data-end error falls as 1/n, scaled by how far the initial noising process is from the bridge.
The first step. With one IPF iteration the algorithm is the conditional score-based model; later iterations correct it (Section 8, Figure 2). The reference end is free, so it can be an informed guess: an upsampled image, an EnKF Gaussian, or SRFlow’s output.
Empirically, at equal small step counts. CelebA 4× super-resolution with noise, N = 20: FID 92.02 (score model), 57.22 (bridge), 44.44 (score model with informed reference), 28.41 (bridge with informed reference). Not uniform: on CelebA inpainting at N = 20 the bridge’s FID (19.85) is worse than the score model’s (17.62). In Lorenz-63 filtering with N = 20 both score-model variants diverge, while the bridge beats the EnKF (0.178 against 0.354 at M = 2000).
What could have come out otherwise: the averaged problem need not have reduced to one bridge on the extended space, and the iterates need not have factored. Both are proved.
What this does not say
- That a long noising horizon is not what makes conditional diffusion correct. That is NOTE-589’s reading of Proposition 1. A source that compared horizons at fixed endpoint error would be needed.
- That the learned networks achieve the bound. Proposition 3 bounds exact IPF iterates, not trained, discretized ones.
- That the method serves every observation well. Training never uses y_obs, so atypical observations are poorly served (the authors’ limitation).
- That it is faster than diffusion in wall-clock terms. Gains are at equal small N. Training costs roughly twice the score model’s, and there is no comparison with distillation or fast solvers.
- That adding conditions composes. Nothing here composes several conditions; CLAIM-123’s conditional-independence point is untouched.
Connections
- The denoising losses rest on the same Gaussian-channel calculus as the I–MMSE identity (LIT-058, THEORY-617). Neither paper cites the other.
- Classifier-free guidance (LIT-839, THEORY-611) also conditions by feeding y to the network; this paper changes the endpoints and the path instead of the sampler’s field.
- CLAIM-123 cites this paper as a path-space view of conditional transport.