nucleation

LIT-776: Conditional Simulation Using Diffusion Schrödinger Bridges

2 min read

Properties5
StatusActive — read 2026-10-09 (NOTE-589); worth reading for the result that posterior sampling from simulation alone, when (X, Y) can be drawn but the likelihood cannot be evaluated, is a Schrödinger bridge between the joint p(x, y) and a reference p_ref(x|y)p(y) with y held fixed along the path (Proposition 1), solvable by iterative proportional fitting with learned drifts (CDSB). Its first iteration is the ordinary conditional score-based model, so later ones refine it, and the reference end can be a cheap posterior approximation rather than noise. Gains over the conditional score model are at equal small step counts and are not uniform across metrics; it filters the Lorenz-63 model with 20 steps where the score model diverges.
Tagsprobabilistic-modeling, mathematics, anthology-candidate
SupportsCLAIM-123 — In conditional generation, adding scores composes two conditions exactly when the ratio p(y₁,y₂|x)/(p(y₁|x)p(y₂|x)) does not vary with the noisy state x, which conditional independence of the conditions at that state is sufficient for but does not require, CLAIM-007 — In a bridge or chain of progressive reconstruction the path itself is observable: which constraints survive early, which are introduced late, and which relations emerge only through refinement
Cited as “paper” byNOTE-589 — Conditional Simulation Using Diffusion Schrödinger Bridges
Cited as “source” byTHEORY-515 — Sampling a posterior p(x|y) when only joint samples (X, Y) can be drawn is solved by a Schrödinger bridge on the joint space, with y held fixed along the path, between the joint distribution and any reference sharing its y-marginal; the bridge averages the per-observation bridges, its first iterative-proportional-fitting step is the conditional score-based model, and the error of later steps at the data end falls as 1/n times the reference's distance from the bridge

Graph View

Yuyang Shi, Valentin De Bortoli, George Deligiannidis and Arnaud Doucet (2022), UAI 2022, PMLR 180 — ARXIV-2202.13460

Key takeaways

  • The construction. Conditional simulation, sampling p(x|y_obs) when only (X, Y) can be simulated, cannot use a bridge to the posterior directly, since the posterior cannot be sampled. The paper poses instead the bridge between the joint p(x, y) and p_ref(x|y)p(y), with y frozen along the path, and proves it is the average over y of the per- observation bridges (Proposition 1). Iterative proportional fitting then splits into conditional forward and backward chains in x, trainable with De Bortoli et al.’s Diffusion SB losses with y as an extra input (Proposition 2, Algorithm 1).
  • Refining the score model. The first IPF iteration is the conditional score-based model (CSGM); later ones correct its bias. The data-end error of the n-th iterate is at most 2/n times the KL from the initial noising process to the bridge (Proposition 3), which argues for a y-dependent forward process.
  • Informed references. The reverse process can start from an approximation to the posterior: an upsampled low-resolution image, an EnKF Gaussian, or a pre-trained SRFlow model, which a short bridge (N = 10) then refines (FID 30.92 to 15.00 on CelebA 8× super-resolution, at a small cost in PSNR).
  • Evidence. Sharper 2D posteriors than a monotone GAN; BOD posterior moments closer to MCMC than MGAN or inverse transport on most statistics; better image metrics than CSGM at equal small N in most but not all cells (Table 2); Lorenz-63 filtering at N = 20 where CSGM diverges, with the advantage gone at N = 100.
  • Limits the authors state: the minimal step count is unknown in practice, atypical observations are poorly served by amortization, and each filtering step solves a new bridge.

Standing in the record

Filed on 2026-10-09 at the owner’s request, as one of the works in the reference list of the owner’s working manuscript (October 2026) that the record did not yet hold. See the curation entry of that day. It is a machine-learning paper an anthology topic could hold (generative-modeling, analysis-and-evaluation), so it carries the anthology-candidate flag (ADR-005); it is here because the owner asked for the manuscript’s references to be filed in this record.

Backlinks