{"id":"405dbc7b-0907-4a34-809e-68420b9cc209","arxiv_id":"2509.26364","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new data-to-energy iterative proportional fitting algorithm trains Schrödinger bridges when endpoint distributions are known only through unnormalised densities.","lead":"This paper extends iterative proportional fitting for Schrödinger bridges to settings where one or both endpoint distributions are given only as unnormalised densities, not as samples. The resulting data-to-energy algorithm makes it possible to train stochastic transports without data from the target, with applications to posterior sampling in latent spaces and image-to-image translation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Forward IPF step trains on x0 from an off-policy buffer; without proof that this marginal covers p0, the alternating IPF fixed point may not satisfy the SB constraints, so the first-method claim is not yet established.","rationale":"The reader's weakest assumption focuses on the forward step not constraining the time-0 marginal to p0. That specific framing is not quite right: in IPF the forward half-bridge is not meant to preserve p0; the alternating backward step enforces it, and the fixed point of the conditional recursion (p0⊗forward conditional) satisfies both marginals. However, the reader's underlying instinct — that the off-policy buffer makes the forward loss insufficiently anchored — is valid. The forward loss must be evaluated on the correct x0 set; if p_train^0 is not p0, the per-x0 proportionality can fail on p0's support, and the alternating scheme may fix the wrong conditionals. The paper's own §§3.1–3.2 and Table 3 show the practical sensitivity to p_train^0 without a theoretical guarantee. The energy-to-energy claim is even less supported because no quantitative metrics are reported. Therefore the central claim should be accepted only conditionally, pending a sharper argument for the buffer's x0 coverage (or a change to sample x0 directly from p0 in data-to-energy runs).","tokens_in":18312,"tokens_out":20543,"duration_ms":168503,"concrete_test":"In the 2D Gauss↔GMM data-to-energy experiment (Table 4), instrument Algorithm 2 to record the empirical distribution of x0 used in the forward-loss batches at each IPF iteration. Measure its Wasserstein distance to the true p0 (e.g., EMD between 10k buffer-x0 samples and 10k p0 samples). Run two variants: (A) the standard replay-buffer scheme; (B) a variant that samples x0 directly from p0 for the loss, keeping the buffer only for exploration. If the buffer x0 marginal deviates substantially from p0 and variant B achieves significantly lower W2/path KL, or if the loss (8) can be driven to zero while the learned forward conditional still violates the p1 constraint on some p0-mass region, the concern lands. If the buffer marginal matches p0 and the two variants agree, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the central claim (first general data-to-energy/energy-to-energy SB method), Algorithm 2 must implement a correct IPF forward half-bridge. The forward loss (8) minimizes, for each x0, the variance of log[→p_θ(τ|x0)/(←p_φ(τ|x1)p1(x1))]. This enforces the required proportionality only on the support of p_train(τ|x0) and only for x0 drawn from p_train^0. In the implementation, p_train^0 is the x0 marginal of backward rollouts from replay-buffer x1 samples (Algorithm 2 lines 11–13; §3.2). Nothing guarantees that this marginal equals p0 or even covers p0's support; it depends on the current ←p_φ and buffer state. The loss is then averaged over this possibly biased p_train^0, so the θ fixed point may satisfy the proportionality on the buffer's x0 support but not on the regions of p0 that the next backward half-bridge (which conditions on x0∼p0) will query. The paper acknowledges the loss only 'approximately enforces' the constraint and that the choice of p_train^0 is 'very important', but provides no analysis of when the buffer scheme converges to the correct conditional. Without that, the alternating recursion need not converge to the Schrödinger bridge. The same issue applies a fortiori to energy-to-energy (§5.2), where no ground-truth x0 samples exist.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an extension of iterative proportional fitting (IPF) for Schrödinger bridge problems in which one or both endpoint marginals are available only as unnormalised densities (data-to-energy and energy-to-energy settings). The forward half-bridge is trained with a conditional log-variance (VarGrad) loss, Eq. (7)–(8), using off-policy replay buffers, reverse-trajectory reuse, and Langevin updates; the backward half-bridge is trained by maximum likelihood as in data-to-data IPF. The authors also train the diffusion coefficients of the forward/backward processes, claiming this improves discrete-time data-to-data IPF, and they apply the method to latent-space posterior sampling of GAN/VAE priors for image translation. The main positive results are on 2D synthetic benchmarks and on FID/reward metrics for CIFAR-10 and MNIST latent-space experiments. The central claim is that this is the first general method for data-to-energy and energy-to-energy Schrödinger bridges.","tokens_in":18778,"tokens_out":8092,"duration_ms":74297,"significance":"If the algorithm were proven to converge to the Schrödinger bridge, it would fill a genuine gap: existing IPF and Schrödinger-bridge matching methods require samples from both marginals, while many applications (Bayesian posteriors, reward-weighted priors) provide only unnormalised densities. The off-policy formulation is a plausible and potentially useful way to adapt diffusion-sampler losses to the bridge setting, and the authors have provided code. The empirical evaluation uses external metrics (W2 against oracle samples, FID against real images), which reduces circularity. However, the central convergence claim is not established: the forward IPF step is only heuristically justified, and the reported experiments do not demonstrate that the fixed point satisfies the p0 marginal constraint that defines the Schrödinger bridge. The secondary claim about learning diffusion coefficients is also only weakly supported by the tables. The paper is therefore more of a promising empirical proposal than a validated general method at this stage.","major_comments":[{"comment":"The forward IPF step is trained by minimising Eq. (8) over x0 drawn from p_train^0. In practice p_train^0 is a mixture of on-policy samples x0~p0 and off-policy samples obtained by reverse rollouts from a replay buffer. The paper states only that this loss 'approximately enforce[s] this proportionality' and that the choice of p_train^0 is 'very important'; no argument is given that the fixed point satisfies the p0 marginal constraint. This is load-bearing because the backward step in Algorithm 2 (line 4) conditions on x0~p0 and will query the forward conditional at those x0. For the energy-to-energy variant (§D.3), there is no on-policy p0 component and buffers are initialised with Gaussian noise, so coverage of p0 is even less certain. Please provide a support/convergence analysis, or modify the algorithm so that p_train^0 provably covers p0.","section":"§3.1, Eq. (8) and Algorithm 2"},{"comment":"The variance in Eq. (7) is defined over a full-support distribution p_train(τ|x0), but the implementation uses N=2 non-i.i.d. trajectories: one reverse roll-out from the buffer-conditioned backward process and one (or more) forward on-policy roll-outs. The empirical variance of a non-exchangeable batch is not an unbiased estimator of the variance of either marginal trajectory distribution, and at a global optimum it need not force p_θ(τ|x0) ∝ p_φ(τ|x1)p1(x1) for draws from the two different proposal mechanisms. This undermines the theoretical link between Eq. (8) and the half-bridge constraint. Please either use i.i.d. batches or provide a separate consistency argument for the non-i.i.d. estimator.","section":"§3.1, Eq. (7)–(8)"},{"comment":"The claim that learning the diffusion coefficient improves existing data-to-data IPF is not supported by the reported numbers. In Table 4, data-to-data learnt-variance has higher Path KL than fixed-variance at K=5, 10, 20 and 40, and in Table 1 it is worse or mixed on the Gauss↔GMM and Gauss↔Two Moons tasks. The statement that the results 'clearly show the benefits of training the variances' is therefore an overclaim. Please re-analyse the results or temper the contribution to specific settings (e.g., small numbers of steps and W2 metric).","section":"§5.1, Tables 1 and 4"},{"comment":"The image-translation experiments contain no baselines: there is no comparison to the outsourced diffusion sampling method of Venkatraman et al. (2025) that the paper generalises, nor to a non-Schrödinger stochastic transport baseline. Consequently the claim that modelling a Schrödinger bridge is beneficial 'as opposed to a diffusion sampler' is unsupported. Additionally, the model's FID is lower than rejection-sampling FID for several classes, which is surprising and suggests either a bias in the classifier reward, mode collapse, or an artefact of the FID computation; this should be explained before the image results can be interpreted.","section":"§5.3, Table 2 and Fig. 4"}],"minor_comments":[{"comment":"The Euler-Maruyama increment in Eq. (17) writes the noise term as σΔt ξ_k; it should be σ√Δt ξ_k under the convention used in Eq. (5a). This affects the intermediate algebra of the path-KL derivation, though the final limit is plausible.","section":"Appendix C, Eq. (17)"},{"comment":"Typos: '2-dimentional', 'leant variance', 'undelying'. Appendix headings contain 'outsoursed'.","section":"§5.1"},{"comment":"The comparison between data-to-data and data-to-energy IPF in Fig. 3 is visual only; reporting quantitative metrics with error bars for the energy-to-energy setting would strengthen the claim of 'comparable' performance.","section":"§5.2"},{"comment":"The notation for the forward and backward IPF steps is inconsistent: Eq. (3a)/(3b) are labelled with arrows but are later referred to as 'backward' and 'forward' respectively. Please align the terminology.","section":"§3.3 and Algorithm 2"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the off-policy x0 distribution lands: the forward IPF step in Algorithm 2 does not have a proof of enforcing the half-bridge constraint, and the energy-to-energy variant is even less protected. The paper would be suitable for a journal after a substantial revision that either adds a rigorous analysis of the training distribution, or significantly reframes the contributions as an empirical heuristic. The learned-variance claim should also be corrected in light of Tables 1 and 4."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the good news: this paper does something new. Existing SB algorithms need samples from both endpoints; the authors replace one endpoint with an unnormalized density and run IPF with a conditional VarGrad loss in the forward step and a replay-buffer/off-policy scheme to obtain x0. The synthetic experiments are plausible, the image-to-image translation in latent space is a nice application, and code is provided.\n\nThe main soft spot is the theory, or rather the lack of it. Equation (8) minimizes, for each x0, the variance of the log ratio over trajectories. That enforces proportionality of p_θ(τ|x0) to p_1(x1)p_φ(τ|x1), but the normalizing constant is free to vary with x0. The time-1 marginal of the implied forward process will be p1 only if, at the fixed point, the backward step has made the two processes exact reversals and the training distribution over x0 covers the relevant support. That fixed-point consistency might hold—the alternating structure could work it out—but the paper doesn't give the argument, and the off-policy x0 distribution (from buffer plus reverse rollouts) has no guarantee of covering p0. The authors say the loss 'approximately enforces' the constraint, which is honest, but for a paper claiming the first general method in this setting, this needs either a proof or a much clearer statement of what is being approximated.\n\nSecond, the secondary claim about learning the diffusion coefficient is not supported by the numbers. In Table 1, learned variance improves W2 on two of three tasks but makes path KL worse in all three. Table 4 is similar. The text says the tables 'clearly show' benefits—they don't.\n\nThe image experiments are suggestive but under-benchmarked: no comparison to a diffusion sampler (e.g., the outsourced diffusion sampling they generalize) or to bridge-sampling alternatives, and FID below the rejection-sampling ground truth is suspicious enough to need discussion.\n\nBottom line: the core idea is novel and the paper is worth a serious referee, but revisions need to address the fixed-point question and tone down the variance claim. I'd bring it to a reading group to kick around the IPF issue.","headline":"A genuinely new data-to-energy Schrödinger bridge algorithm that is likely the first of its kind, but the forward IPF step is heuristic and the paper's own tables contradict the secondary learned-variance claim.","tokens_in":19144,"tokens_out":10026,"would_cite":true,"duration_ms":80785,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes the first general algorithm for Schrödinger bridge problems in which one or both endpoints are known only as unnormalised densities, extending iterative proportional fitting with a variance-minimising loss and demonstrat","keywords":["Schrödinger bridge","iterative proportional fitting","unnormalised density sampling","diffusion samplers","off-policy reinforcement learning","stochastic optimal transport","learned diffusion coefficient","latent-space image translation"],"falsifier":"For a two-dimensional Gauss-to-GMM pair with a known analytic or high-accuracy sample-based reference bridge, run data-to-energy IPF to convergence, then evaluate the time-0 marginal of the forward process against p0 and compare the joint (x0,x1) law against the entropic-optimal-transport coupling. Significant deviation in the time-0 marginal or the coupling would show that the conditional variance loss alone does not identify the bridge without an additional marginal constraint.","tokens_in":18219,"feed_emoji":"🌉","tokens_out":4365,"duration_ms":37320,"temperature":0.7,"pith_summary":"This paper proposes the first general algorithm for solving Schrödinger bridge problems when one or both endpoint distributions are known only through an unnormalised density—a function you can evaluate but cannot sample from. Existing bridge algorithms require samples from both endpoints; this one replaces the sample-based half-bridge updates with a variance-minimisation loss borrowed from diffusion samplers, so the iterative proportional fitting recursion can run data-free. The authors show on synthetic problems that the resulting bridges match sample-based ones, and apply it to posterior sampling in latent spaces of generative models, yielding image-to-image translation where semantic content is preserved. A secondary claim is that learning the diffusion coefficient as well as the drift improves existing data-to-data bridge algorithms at coarse time discretisations.","feed_headline":"Bridges two distributions when only energies are known","feed_subtitle":"Extends Schrödinger-bridge transport to unnormalised densities, enabling sampling and image translation without target samples.","key_machinery":"The central mechanism is the iterative proportional fitting (IPF) recursion, which alternates half-bridge projections onto the two endpoint marginals. In the data-to-energy setting the forward projection is carried out by a source-conditional log-variance (VarGrad-type) loss: over a batch of trajectories sharing an initial point x0, it minimises the variance of log p_theta(tau|x0) - log p_phi(tau|x1) - log p1(x1), a quantity whose normalising constant cancels. Off-policy training distributions—a replay buffer of terminal samples, Langevin refreshes, and reuse of reverse trajectories—supply the x0 locations and trajectory batches that make the loss tractable in high dimensions.","core_discovery":"The paper claims that the Schrödinger bridge between a sample-available distribution p0 and an energy-defined distribution p1 can be computed by running IPF where the forward half-bridge is solved with a source-conditional log-variance loss rather than maximum likelihood. The loss minimises the variance, over trajectories sharing the same x0, of log p_theta(tau|x0) - log p_phi(tau|x1) - log p1(x1); because the normalising constant of p1 cancels, no samples from p1 are needed. The paper further claims that learning the diffusion coefficients of the discretised processes—not only the drifts—systematically improves existing data-to-data bridge algorithms, and that applying the resulting data-to","pith_inferences":["The authors left implicit that the same variance loss could handle constraints expressed as arbitrary reward-shaped weights, not just probabilities, since only the log-weight appears in the objective.","The off-policy buffer acts as an implicit importance-sampling distribution; a natural extension would be to reweight buffer samples by importance weights or anneal the off-policy ratio dynamically, which the ablation suggests trades off mode coverage and transport cost.","The paper does not directly report the time-0 marginal of the learned forward process; a sharper fixed-point test would compare it to p0, since this is the constraint the variance loss only approximately enforces."],"forward_implications":["Bridge solvers can now target Boltzmann distributions, posteriors, and other energy models without a sample oracle, the main regime in scientific computing.","Learning the diffusion coefficient as well as the drift yields better bridges under coarse discretisation, so existing sample-based bridge pipelines can be improved by the same change.","Posterior sampling in latent spaces becomes a stochastic transport problem; class-conditional translation preserves unconstrained semantic structure such as background and digit style.","Energy-to-energy bridges, with no samples from either endpoint, are feasible when both sides are equipped with replay buffers.","Data-to-energy IPF reduces to diffusion sampler training when the forward process is already a bridge, connecting the new method to the diffusion-sampling literature."],"fun_headline_variants":["Schrödinger bridges without target samples","Bridging distributions from energies alone","Data-free Schrödinger bridge via variance loss","Learning diffusion coefficients improves Schrödinger bridges","Sample-free image translation with Schrödinger bridges"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The forward IPF step is trained by minimising the variance of a trajectory log-ratio conditional on x0, which only approximately enforces p_theta(tau|x0) proportional to p1(x1)p_phi(tau|x1) and does not by itself constrain the time-0 marginal of the learned process to equal p0; if the off-policy training distribution over x0 misses regions of p0, the fixed point of the recursion need not be the Schrödinger bridge.","fun_headline_variants_meta":{"raw":{"variants":["Schrödinger bridges without target samples","Bridging distributions from energies alone","Data-free Schrödinger bridge via variance loss","Learning diffusion coefficients improves Schrödinger bridges","Sample-free image translation with Schrödinger bridges"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000383,"raw_usage":{"total_tokens":1890,"prompt_tokens":796,"completion_tokens":1094,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":1031}},"tokens_in":540,"tokens_out":1094,"duration_ms":7947,"temperature":1.0,"reasoning_tokens":1031,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T13:32:45.812685+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a two-dimensional Gauss-to-GMM pair with a known analytic or high-accuracy sample-based reference bridge, run data-to-energy IPF to convergence, then evaluate the time-0 marginal of the forward process against p0 and compare the joint (x0,x1) law against the entropic-optimal-transport coupling. Significant deviation in the time-0 marginal or the coupling would show that the conditional variance loss alone does not identify the bridge without an additional marginal constraint.","supporting_citations":[],"review_version":1}