acceptodds
Under review as a conference paper at ICLR 2027

Generative Modeling via Nonlinear Schrödinger Bridges and Branching Diffusions

Abstract

We propose a new generative modeling framework grounded in forward-backward partial differential equations (PDEs). These equations arise naturally in mean-field game (MFG) theory, and are closely related to nonlinear Schrödinger systems. Our main contribution is to reinterpret this connection within the setting of generative modeling and to develop an efficient numerical methodology for its practical implementation. Specifically, we derive and characterize a nonlinear extension of the classical Schrödinger bridge problem through a coupled system of PDEs, yielding a model we call SBNL (Schrödinger Bridge NonLinear). To compute its solution, we design a simulation algorithm based on branching diffusions and the probabilistic representation of semilinear PDEs. The resulting method is embarrassingly parallel, scalable to high dimensions, and applicable to a broad class of nonlinear interaction functionals. We validate the approach through numerical experiments, including recovery of the classical Schrödinger bridge, nonlinear extensions with convex penalties (Gross-Pitaevskii) as well as potential interaction terms, and high-dimensional generative tasks such as image synthesis on the MNIST and FFHQ human face dataset. Furthermore, the same framework embeds score-based generative modeling, diffusion models, and tractable gaussian Schrödinger bridges.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Generative modeling, Mean field games, Schrödinger Bridge

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gt7zOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies Schrödinger bridges with running and terminal costs that depend on the evolving marginal distribution. It expresses the optimality conditions through coupled Schrödinger potentials, related by a Cole–Hopf transformation to a Hamilton–Jacobi–Fokker–Planck system. Its main computational proposal is a two-type branching diffusion representation: offspring of the opposite type restart at the reflected time T−sT-s, allowing the coupled potentials to be evaluated by weighted Monte Carlo trees. Neural terminal potentials are trained jointly by matching the two boundary conditions with an MMD objective, and samples are generated using Monte Carlo estimates of the logarithmic gradient of a potential.

Validation includes classical bridge recovery, soft terminal constraints, a two-dimensional density-dependent penalty, obstacle avoidance, and latent transport on MNIST and FFHQ. The paper reports competitive classical-bridge results and a favorable FFHQ FID. However, the large-dimensional experiments shown in the main text use F=0F=0, and several central theoretical and objective-level claims require correction or additional justification.

Critical strengths
  1. A meaningful computational approach to coupled nonlinear potentials. The combination of two particle types, reflected offspring times, and joint MMD fitting is an interesting way to approach a coupled forward–backward problem without a spatial mesh. The paper explicitly acknowledges its connection to existing mean-field formulations in Remark 2.3. The novelty should therefore be assessed primarily as a computational construction, rather than as the discovery of law-dependent Schrödinger bridges themselves. Prior work such as Deep Generalized Schrödinger Bridge already connects bridges with mean-field structure, but that does not eliminate the potential value of the different branching-based solver proposed here.

  2. Substantial initial validation with relevant baselines and some appropriately bounded claims. Figure 3 compares against DeepGSB, LightSB, LightSBB-M, DSBM, and BranchSBM, rather than relying only on parameter sweeps. The reported mixture Wasserstein distance is competitive, and Figure 7 reports an FFHQ FID of 36.03, compared with 38.92 for DSBM and 39.54 for BranchSBM-M. Appendix K also provides a useful analytical sanity check for a time-reflected nonlinear equation. These experiments support the claim that the implementation can produce useful transports in the tested settings. They do not establish uniform superiority, which the discussion correctly refrains from claiming.

Critical weaknesses
  1. The central branching and sampling guarantees are not established as stated.

Proposition 3.1 assumes bounded terminal potentials and bounded polynomial coefficients, then claims an unbiased, square-integrable representation. Appendix L argues that its absolute-moment majorant is globally bounded because the product of two reflected potentials is constant. Constancy does not establish that this product can satisfy both endpoint conditions.

A concrete diagnostic is the spatially homogeneous case F(m)=mF(m)=m, c=T=1c=T=1, and φ(T,x)=φˉ(T,x)=1\varphi(T,x)=\bar\varphi(T,x)=1. These functions meet the proposition's stated boundedness assumptions. In the canonical decomposition of Equation (10), a0=1a_0=1 and a1=−1a_1=-1. Appendix L's absolute-coefficient majorant consequently has coefficient function 1−(1+r)=−r1-(1+r)=-r. Solving its differential equations with the two terminal values forces its positive constant product to satisfy r=err=e^r, which has no positive solution. In contrast, the original signed equations admit a positive homogeneous solution with r=e−rr=e^{-r}. This example concerns the proposition's stated function class, not a normalized probability bridge; it directly invalidates the claimed automatic existence of the majorant. The familiar need for moment and non-explosion conditions in branching diffusion representations makes it consequential to specify and prove the actual admissible regime here.

Separately, Section 4.2 and Algorithm 2 identify ∇log⁡φ^\nabla\log\widehat\varphi as an unbiased estimate of ∇log⁡φ\nabla\log\varphi, where φ^=M−1∑j=1MΦj\widehat\varphi=M^{-1}\sum_{j=1}^{M}\Phi_j. In general, E[∇φ^/φ^]≠∇φ/φ\mathbb{E}[\nabla\widehat\varphi/\widehat\varphi]\ne\nabla\varphi/\varphi, even if the potential and its gradient are separately unbiased. This is already visible in a bounded, strictly positive linear heat example: with φ(T,x)=2+sin⁡x\varphi(T,x)=2+\sin x, T−t=1T-t=1, x=0x=0, and M=1M=1, the true score is e−1/2/2≈0.303265e^{-1/2}/2\approx0.303265, whereas the expectation of the estimator is approximately 0.324611. Increasing MM can reduce bias; it does not establish finite-sample unbiasedness. Signed branching weights additionally require a justification that the empirical potential remains positive before its logarithm is taken. The explanation on page 8 attributing the absence of variance collapse to an unbiased score estimator is therefore unsupported by the stated argument. These issues concern core guarantees rather than optional future variance reduction.

  1. The intended objective and the evidence for the nonlinear contribution remain insufficiently connected.

Remark 2.4 calls the terminal penalty reverse KL, KL(μT∥mT)\mathrm{KL}(\mu_T\|m_T), but its exponential conjugate and power-law boundary condition do not follow from that probability-domain penalty. For the standard forward KL on probability measures, the conjugate is λlog⁡EμT[ev/λ]\lambda\log\mathbb{E}_{\mu_T}[e^{v/\lambda}], rather than the exponential integral displayed in Equation (6). A suitable unnormalized entropy functional has a different domain and normalization convention. Those distinctions matter because the process law is a probability distribution and the soft terminal condition is used in training.

There is also a direct inconsistency in Equation (15): the second squared-measure term uses A^(Z,∙)φθ(Z)\widehat A(Z,\bullet)\varphi_\theta(Z), as required by Equations (13) and (27), but its corresponding cross term uses A^(Z,∙)φˉη(Z)\widehat A(Z,\bullet)\bar\varphi_\eta(Z). This may be a typographical error, but the paper does not establish which expression was implemented. Algorithm 1 further sums all minibatch pairs, including the diagonal, while reusing one-tree estimates. On a diagonal, squaring an unbiased tree estimate does not give an unbiased estimate of the squared potential; it introduces a parameter-dependent second-moment contribution. Thus, the expected minibatch objective also needs to be distinguished from the intended population MMD.

Empirically, the high-dimensional mixture, MNIST, and FFHQ experiments in the main text all use F=0F=0. They validate the classical specialization, not scalability of genuinely density-dependent branching. The F(m)=κmF(m)=\kappa m experiment is two-dimensional and principally shows a penalty trajectory; the obstacle example uses F=V(x)F=V(x) and does not test dependence on the evolving density. A small set of quantitative nonlinear comparisons, together with endpoint error and tree-cost measurements, would substantially strengthen the central contribution. I am not requiring an exhaustive benchmark suite. The required distinction is between demonstrating an interesting nonlinear mechanism and establishing that it reliably solves the intended nonlinear bridge. Appendix O may contain additional relevant protocols or evidence, which could not be assessed from this copy.

Questions for authors
  1. What additional conditions ensure existence, tree termination, and finite second moments in Proposition 3.1? In particular, how does Appendix L accommodate F(m)=mF(m)=m, c=T=1c=T=1, and constant terminal potentials equal to one, for which its absolute-moment majorant requires r=err=e^r? Please provide a corrected argument or restricted theorem, and identify which experimental configurations satisfy its conditions. This could materially change my soundness assessment.

  2. Is Algorithm 2 actually intended to provide an unbiased score, or a biased but consistent ratio estimator? Please justify the claim for the implemented estimator or revise it and characterize the finite-MM bias. How are nonpositive empirical potentials caused by signed weights handled? Any clipping, tree truncation, or alternative coefficient decomposition should be stated, since it can change both the estimator and its guarantees.

  3. Which potential multiplies the black-tree estimator in the implemented cross term: φθ\varphi_\theta as required by Equations (13) and (27), or φˉη\bar\varphi_\eta as printed in Equation (15)? Are diagonal minibatch products evaluated with independent trees, removed as in a U-statistic, or retained with their bias? Please state the objective whose gradient is actually estimated and whether correcting these discrepancies changes the reported results.

  4. Is the intended penalty reverse KL, forward KL on probability measures, or an unnormalized entropy functional on finite measures? Please reconcile its definition, the conjugate in Equation (6), the boundary conditions in Equations (7)–(8), and mass normalization. Does the corrected formulation change the soft-constraint experiment or the interpretation of its reported terminal KL?

  5. What quantitative results establish successful optimization with genuinely density-dependent F≠0F\ne0, beyond the two-dimensional penalty plots? A representative comparison against an appropriate nonlinear solver, reporting the nonlinear objective, endpoint mismatch, and matched computational cost, would be particularly informative. Please also identify any supporting results in Appendix O and report how tree counts, runtime, and stability depend on MM, cc, κ\kappa, and dimension. This requests evidence for the central claim, rather than an exhaustive set of future applications.

2 comments
AuthorsOpenReview ↗

Corrected branching conditions, consistent drift estimator, clarified objective

We thank the reviewer for the careful reading and the explicit counterexamples.

Q1. The reviewer is right: in App. L, the constancy of U(t)Uˉ(T−t)=rU(t)\bar U(T-t)=r does not ensure that the majorant meets both endpoint conditions. For the second-moment majorant these force r=Ψ(r):=A2B2exp⁡(cT(Q(r)−1))r=\Psi(r):=A^2B^2\exp(cT(Q(r)-1)), with A=∥φT∥∞A=\Vert\varphi_T\Vert_\infty, B=∥φˉT∥∞B=\Vert\bar\varphi_T\Vert_\infty, Q(r)=∑k∥ak∥∞2rk/pkQ(r)=\sum_k\Vert a_k\Vert_\infty^2r^k/p_k, and a bounded solution exists iff this equation has a root. The corrected Prop. 3.1 assumes (i) cTKˉ<1cT\bar K<1, Kˉ=∑kkpk\bar K=\sum_k kp_k, and (ii) Ψ(r)≤r\Psi(r)\le r for some r>0r>0. Under (i), the expected tree size is exactly 1+c(T−t)(1+2Kˉ)/(1−cTKˉ)1+c(T-t)(1+2\bar K)/(1-cT\bar K), so trees are a.s. finite; under (ii), sup⁡t,xE[Φ2]<∞\sup_{t,x}\mathbb E[\Phi^2]<\infty and E[Φ]\mathbb E[\Phi] is the unique mild solution in the dominated class. Your example: for F=κmF=\kappa m and φT=φˉT≡1\varphi_T=\bar\varphi_T\equiv1, the signed system has the solution e−cκm(T−t)e^{-c\kappa m(T-t)}, m=W0(cκT)/(cκT)m=W_0(c\kappa T)/(c\kappa T), while r=ecκTrr=e^{c\kappa Tr} has a root iff e cκT≤1e\,c\kappa T\le1. As ∣Φ∣\vert\Phi\vert equals the majorant functional for constant data, E∣Φ∣=∞\mathbb E\vert\Phi\vert=\infty when e cκT>1e\,c\kappa T>1, so (ii) cannot be dropped although the PDE solution exists. Experiments: for F∈{0,V}F\in\lbrace0,V\rbrace, K=0K=0 and (i)-(ii) hold for every TT (all non-Gross-Pitaevskii runs). For F=κmF=\kappa m (T=2T=2, p0=p1=12p_0=p_1=\frac12), (i) holds as cT≤0.6cT\le0.6, and (ii) reads 2e cTκ2A2B2ecT≤12e\,cT\kappa^2A^2B^2e^{cT}\le1. At a solution, AB≥sup⁡μ0≥(2πe)−1AB\ge\sup\mu_0\ge(2\pi e)^{-1} for whitened 2D data, so (ii) fails for (c,κ)=(0.3,10)(c,\kappa)=(0.3,10) and (0.3,100)(0.3,100), which we now present as qualitative only; for κ=1\kappa=1 it reduces to an explicit bound on ABAB.

Q2. It is a consistent ratio estimator, not an unbiased one. Under smooth data and a strict version of (ii), φ^\widehat\varphi and ∇φ^\nabla\widehat\varphi are unbiased, and s^M=∇φ^/φ^\widehat s_M=\nabla\widehat\varphi/\widehat\varphi is consistent and asymptotically normal. For positive weights with ∣∇Φ∣≤LΦ\vert\nabla\Phi\vert\le L\Phi, E[s^M]=s+b1/M+O(M−2)\mathbb E[\widehat s_M]=s+b_1/M+O(M^{-2}), b1=(s Var Φ−Cov(∇Φ,Φ))/φ2b_1=(s\,\mathrm{Var}\,\Phi-\mathrm{Cov}(\nabla\Phi,\Phi))/\varphi^2. Your heat example is such a case, with b1=s(1−e−2)/8>0b_1=s(1-e^{-2})/8>0. With signed weights we only claim consistency and the CLT. The exact potential is positive (Feynman-Kac), but the empirical average may not be; Algorithm 2 then sets the drift to zero. No clipping, tree truncation or alternative decomposition is used. The sentence on p. 8 is removed.

Q3. The cross term should contain A^(Z,∙)φθ(Z)\widehat A(Z,\bullet)\varphi_\theta(Z), as in Eqs. (13) and (27); φˉη\bar\varphi_\eta in Eq. (15) is a typo, and the code uses φθ\varphi_\theta (WT = B_Z * phi_Z / mu_Z in sbnl_toy.py). The two factors of each quadratic term use independent reference points and trees, so no squared one-tree estimate appears; Algorithm 1 now writes the loss as a U-statistic over b≠b′b\neq b', unbiased for the MMD of Prop. 4.1, as is its gradient. Our runs, however, normalise the weights by ∑b∣Wb∣\sum_b\vert W_b\vert. With positive weights (F∈{0,V}F\in\lbrace0,V\rbrace), the zeros of this objective define the same bridge, so those results are unchanged. With the signed weights of the Gross-Pitaevskii runs, it is strictly positive, so these runs minimise a surrogate; we are re-running them with the unnormalised loss.

Q4. The penalty is the forward KL: with f(u)=−log⁡u−1f(u)=-\log u-1, Df(μT,m)=∫m(log⁡mμT−1)=KL(m∥μT)−1D_f(\mu_T,m)=\int m(\log\frac{m}{\mu_T}-1)=\mathrm{KL}(m\Vert\mu_T)-1 on probability densities, so "reverse" in Remark 2.4 is wrong. Eq. (6) is the conjugate over nonnegative densities; over probability densities it is λlog⁡EμT[ev/λ]+λ\lambda\log\mathbb E_{\mu_T}[e^{v/\lambda}]+\lambda. Both agree at the optimum, and the mass constraint only adds a multiplier, absorbed by the gauge (φ,φˉ)↦(sφ,φˉ/s)(\varphi,\bar\varphi)\mapsto(s\varphi,\bar\varphi/s). Eqs. (7)-(8) and the soft-constraint experiment are unchanged; Fig. 4b reports KL(Law(XT)∥μT)\mathrm{KL}(\mathrm{Law}(X_T)\Vert\mu_T).

Q5. The paper does show the effect of non-classical running costs in dimension 2, on obstacle (F=VF=V, Fig. 5, App. O.3) and congestion (F=κmF=\kappa m, Fig. 4c) problems of the type studied by DeepGSB (Liu et al., 2022) and GSBM (Liu et al., 2024); for F=κmF=\kappa m, larger cc lowers the concentration at intermediate times. In high dimension, we chose to compare with SOTA SB methods (DSBM, LightSB, LightSBB-M, DeepGSB, BranchSBM) on their benchmarks, which are classical bridges, hence F=0F=0 for MNIST and FFHQ. The 50D run (App. O.4) does use F=κmF=\kappa m, c=0.1c=0.1; the label F=0F=0 in Sec. 5 is a typo. Whitened densities being small in d=50d=50, the coupling is weak there, and this run mainly shows that training with branching trees is stable. The expected tree size (Q1) does not depend on dd or κ\kappa, and a tree costs O(d)O(d). A quantitative 2D comparison against a grid solver (Guéant, 2012), DeepGSB and the explicit case e cκT≤1e\,c\kappa T\le1 of Q1 would further strengthen this evidence; Sec. 6 now mentions it.

Thank you very much for the detailed response and clarifications. I have noted the chairs’ reminder regarding AI use. While the corrections are helpful, some nonlinear experiments remain outside the revised theoretical guarantees, and corrected results for the signed-weight objective are still pending. These limitations leave my main concerns only partially resolved. I therefore maintain my original rating of weak rejection.

Reviewer_uMZ8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies Schrödinger bridges with distribution-dependent running costs and soft terminal costs, aiming to construct a generative model, SBNL, through mean-field optimal control, coupled semilinear PDEs, and time-reflected branching diffusions with two particle types. The method parameterizes terminal potentials with neural networks, estimates intermediate potentials using branching Monte Carlo, matches boundary distributions through MMD, and drives sampling with estimated log-potential gradients. Validation includes two-dimensional distribution transport, soft terminal constraints, nonlinear costs and potential-field obstacles, as well as high-dimensional mixtures and latent-space transport on MNIST and FFHQ. The central value proposition is a unified treatment of nonlinear bridges and a computable, unbiased stochastic representation.

Critical strengths
  1. Combining distribution-dependent nonlinearity with a branching representation has potential value.

    The method combines distribution-dependent semilinear coupling, a time-reflected branching representation with two particle types, and MMD-based joint learning of terminal potentials.

    If the theoretical and algorithmic consistency concerns below can be resolved, this combination could broaden the constraints expressible in generative transport and provide a computational approach to bridges with nonlinear interaction costs.

  2. The experiments cover varied transport phenomena and provide some transparency about the results.

    Experiments cover hard and soft constraints, nonlinear penalties, transport around potential-field obstacles, and latent-space image transport. Figure 3 on page 8 and Figure 7 on page 9 already compare with methods including DSBM, LightSB, DeepGSB, and BranchSBM. Figure 3 reports multiple distribution distances and moment statistics, and Appendix O.2.1 explains that the error bars come from sampling bootstraps with fixed trained models.

    The discussion acknowledges nonlinear hyperparameter tuning and branching-variance issues, and does not claim systematic superiority over mature DSB methods in the linear setting.

    These comparisons and limitations provide a basis for assessing the method's scope.

Critical weaknesses
  1. The direction of the soft terminal KL cost is inconsistent with the key derivation, affecting the subsequent optimization objective.

    Remark 2.4 on page 4 explicitly states g(m)=λKL(μT∥m)g(m)=\lambda\mathrm{KL}(\mu_T\Vert m), but Equation (6) gives an exponential-form conjugate, and Equation (7) then yields a power-law boundary corresponding to G−1(v)=μTexp⁡ ⁣(vλ)G^{-1}(v)=\mu_T\exp\!\left(\frac{v}{\lambda}\right).

    Under the paper's definition G=δgδmG=\frac{\delta g}{\delta m}, the variational derivative of the stated reverse KL should be −λμTm-\frac{\lambda\mu_T}{m}, with an additional spatially constant multiplier if total mass is constrained. The resulting boundary should have a reciprocal form rather than the stated exponential form. The exponential relationship is more consistent with the opposite KL direction.

    This issue propagates to Equation (8), Equations (14)–(15), and the finite-λ\lambda experiments. The paper must therefore clarify whether the cost is misstated, the derivation is incorrect, or the implementation uses another definition. Qualitative trends in soft-constraint experiments cannot by themselves validate the current proposition.

  2. The paper extends unbiasedness of potential estimation to the log-gradient, but Algorithm 2 does not support this inference.

    Proposition 3.1 on page 6 provides a stochastic representation of φ\varphi, while Section 4.2 on page 7 claims that it yields unbiased estimates of ∇log⁡φ\nabla\log\varphi.

    Algorithm 2 actually averages finitely many branching trees to obtain φ^\widehat{\varphi}, then computes:

    =\frac{\nabla\widehat{\varphi}}{\widehat{\varphi}}.$$ Even if numerator and denominator are individually unbiased, their ratio is generally biased. No additional debiasing mechanism or sufficient special conditions are provided. Page 8 also attributes high-dimensional experimental behavior to unbiased drift estimation, so the discrepancy affects both the theoretical claim and interpretation of the experiments. Tree count, stability when the potential approaches zero, and drift error need analysis; the potential-representation proposition alone is insufficient.
  3. A factor in the sample-based MMD loss is inconsistent with the boundary-matching objective, and its correspondence to the implementation needs checking.

    The second matching term in Equation (13) on page 6 and Equation (27) in Appendix B involves φˉ(0,⋅)φ(T,⋅)\bar{\varphi}(0,\cdot)\varphi(T,\cdot). Its branching estimate at a reference location ZZ should therefore contain A^(Z,∙)φθ(Z)\widehat{A}(Z,\bullet)\varphi_\theta(Z).

    The quadratic term in Equation (15) on page 7 uses this factor, but the corresponding cross term instead contains A^(Z,∙)φˉη(Z)\widehat{A}(Z,\bullet)\bar{\varphi}_\eta(Z). The two terminal potentials have different boundary roles, so this substitution generally changes the loss and its gradient rather than merely changing notation.

    This may only be a typographical error, and does not establish an implementation error. However, the manuscript does not currently allow readers to verify whether training optimizes the stated pair of boundary-matching objectives.

    This requires a separate check alongside correction of the terminal KL definition, including confirmation that the reported experimental results remain valid.

  4. The extent to which the high-dimensional experiments support the distribution-dependent nonlinear extension is unclear.

    The main text on pages 8–9 labels the 50-dimensional mixture, MNIST, and FFHQ experiments as F=0,λ=∞F=0,\lambda=\infty, corresponding to classical bridges. Appendices O.4/O.5 on pages 31–32 instead call the first two “nonlinear bridges” with c=0.1c=0.1.

    Having c>0c>0 or using a branching simulator does not establish distribution-dependent nonlinearity in the running cost. When F=0F=0, potential propagation remains governed by linear heat equations. The two-dimensional Gross–Pitaevskii and potential-field experiments illustrate different costs, but do not automatically establish numerical stability or computational costs for density-dependent nonlinearity in high dimensions.

    The discussion already acknowledges difficulties with nonlinear tuning and branching variance. This makes it particularly important to distinguish high-dimensional computability of classical bridges from the demonstrated scope of the nonlinear extension.

    This issue concerns the strength of evidence for the method's applicability, rather than negating the existing high-dimensional classical-bridge results.

Questions for authors
  1. Definition clarification:

    Is the implemented terminal penalty KL(μT∥m)\mathrm{KL}(\mu_T\Vert m) or KL(m∥μT)\mathrm{KL}(m\Vert\mu_T)?

    Please provide the variational derivative, conjugate, and normalization constraints consistent with the implementation, and identify which propositions, loss terms, and finite-λ\lambda results require changes after correction.

  2. Theoretical and numerical validation of drift estimation:

    For finite branching-tree count MM, does Algorithm 2 include an additional debiasing mechanism or special conditions sufficient to guarantee an unbiased log-potential gradient?

    If only consistency under appropriate conditions is available, please state those conditions and qualify the claim accordingly.

    On a bridge with an analytic solution, could you report drift bias, variance, and sampling error as functions of MM, and explain how near-zero potential estimates are handled?

  3. Checking the MMD formula against the training implementation:

    Should the cross term of the second MMD in Equation (15) use φθ(Z)\varphi_\theta(Z) or φˉη(Z)\bar{\varphi}_\eta(Z) at ZZ?

    Please give an expansion consistent with Equation (13) and the boundary conditions in Appendix B, and identify which factor is used in the training code.

    If this is only a typographical error, please confirm that all reported results use the correct implementation. If the implementation also requires correction, please explain the effects on the loss, boundary matching, and experimental results.

  4. Experimental settings and scope of the nonlinear conclusions:

    Please reconcile, for each of the 50-dimensional mixture, MNIST, and FFHQ experiments, the actual FF, λ\lambda, polynomial coefficients, and branching settings, and explain the difference between F=0F=0 in the main text and the “nonlinear bridge” descriptions in the appendices.

    If all high-dimensional experiments are classical bridges, please clarify which applicability conclusions follow from those results and which follow from the two-dimensional distribution-dependent nonlinear experiments. If high-dimensional results with density-dependent FF already exist, please identify their settings and the evidence for stability and computational cost.

    The priority is accurately characterizing the existing experiments; extending to a new task is unnecessary for this clarification.

1 comment
AuthorsOpenReview ↗

Clarifications on KL convention, score estimator, MMD term and experiment settings

We thank the reviewer for the detailed review and careful reading, which helped us clarify several points.

W1/Q1 – KL direction. We acknowledge that Remark 2.4 does not state the ff-divergence convention and that the label "reverse" mismatches the equations. This changes no equation, implementation or result. With the convention used in Eqs. (6)–(8) and in the code, g(m)=λ∫m(log⁡mμT−1)=λ(KL(m∥μT)−1)g(m)=\lambda\int m(\log\frac{m}{\mu_T}-1)=\lambda(\mathrm{KL}(m\Vert\mu_T)-1) since ∫m=1\int m=1: the KL up to an additive constant, as the remark states. Then G=λlog⁡(m/μT)G=\lambda\log(m/\mu_T), so G−1(−1clog⁡φT)=μTφT−aG^{-1}(-\frac1c\log\varphi_T)=\mu_T\varphi_T^{-a} with a=1/(λc)a=1/(\lambda c) (Eq. 7), and g∗(v)=λEμT[ev/λ]g^{\ast}(v)=\lambda\mathbb E_{\mu_T}[e^{v/\lambda}] (Eq. 6). The reviewer's −λμT/m-\lambda\mu_T/m is the derivative of KL(μT∥m)\mathrm{KL}(\mu_T\Vert m), not used in the paper. Imposing ∫m=1\int m=1 only rescales the condition by a constant, absorbed by the classical gauge invariance (φ,φˉ)↦(sφ,φˉ/s)(\varphi,\bar\varphi)\mapsto(s\varphi,\bar\varphi/s) of Schrödinger potentials; mm and the drift are unchanged. We will state the convention and fix the label.

W2/Q2 – Score. We average MM trees, φ^=1M∑jΦj\widehat\varphi=\frac1M\sum_j\Phi_j, and autograd gives ∇φ^/φ^\nabla\widehat\varphi/\widehat\varphi. Numerator and denominator are unbiased, but their ratio is not: E[∇φ^/φ^]=∇φ/φ+b1/M+O(M−2)\mathbb E[\nabla\widehat\varphi/\widehat\varphi]=\nabla\varphi/\varphi+b_1/M+O(M^{-2}), with b1=∇φVarΦ/φ3−Cov(∇Φ,Φ)/φ2b_1=\nabla\varphi\mathrm{Var}\Phi/\varphi^3-\mathrm{Cov}(\nabla\Phi,\Phi)/\varphi^2, while the standard deviation is O(M−1/2)O(M^{-1/2}). Under conditions (i)–(ii) of Prop. 3.1 (Φ∈L2\Phi\in L^2) and φ>0\varphi>0, the estimator is consistent, with bias M\sqrt M times smaller than its noise (M=5000M=5000 in all high-dimensional runs). Prop. 3.1 is unaffected; only the wording on the drift changes. We replace "unbiased" by "consistent with O(1/M)O(1/M) bias", remove the p. 8 sentence on variance collapse, and will report drift bias, variance and sampling error against MM on the analytic Gaussian bridge. Near-zero potentials: for F=0F=0, φ^>0\widehat\varphi>0 since φθ>0\varphi_\theta>0 (scaled sigmoid). Otherwise, rows with φ^≤0\widehat\varphi\le0 get zero drift; this never occurred in our runs.

W3/Q3 – MMD cross term. Expanding MMD2(ν,ρ)\mathrm{MMD}^2(\nu,\rho) with ν=A(⋅,∙)φθ\nu=A(\cdot,\bullet)\varphi_\theta and ρ=μTφθ−a\rho=\mu_T\varphi_\theta^{-a}, both ν\nu-terms contain A^(Z,∙)φθ(Z)\widehat A(Z,\bullet)\varphi_\theta(Z), as in Eqs. (13) and (27). The φˉη(Z)\bar\varphi_\eta(Z) in the cross term of Eq. (15) is a typo: the supplementary code uses φθ(Z)\varphi_\theta(Z) (WT = B_Z * phi_Z / mu_Z in sbnl_toy.py). Implementation details:

  • Importance weights W(Z)=A^(Z,∙)φθ(Z)/μ∗(Z)W(Z)=\widehat A(Z,\bullet)\varphi_\theta(Z)/\mu^{\ast}(Z) with a characteristic Gaussian-mixture kernel, so Prop. 4.1 applies.
  • ZZ and Z′Z' are drawn independently with independent trees: no diagonal bias in the quadratic term.
  • Weights are self-normalised, W/∑∣W∣W/\sum\vert W\vert, matching ν/∫ν\nu/\int\nu to μT\mu_T (a consistent ratio estimator). For linear runs (F∈{0,V}F\in\lbrace 0,V\rbrace, λ=∞\lambda=\infty) this is exact up to a constant: m=φφˉm=\varphi\bar\varphi conserves mass (Remark 2.3), so ∫ν0=∫νT\int\nu_0=\int\nu_T, and φ↦sφ\varphi\mapsto s\varphi rescales both while leaving ∇log⁡φ\nabla\log\varphi invariant.

We will state the implemented objective in Algorithm 1.

W4/Q4 – Scope. The main-text "F=0F=0" for the 50D mixture is a typo, now corrected: as stated in App. O.4, it uses the Gross–Pitaevskii cost. We also align the MNIST wording in App. O.5. Settings:

ExperimentFFλ\lambdacoefficientsMM
50D mixtureκm\kappa m, c=0.1c=0.1∞\inftya0=1a_0=1, a1=−κa_1=-\kappa, K=1K=15000
MNIST (d=32d=32)0∞\inftya0=1a_0=1, K=0K=05000
FFHQ (d=512d=512)0∞\inftya0=1a_0=1, K=0K=05000

The 50D run is high-dimensional evidence for a density-dependent cost: training is stable, and the mean, standard deviation and skewness of Law(XT)\mathrm{Law}(X_T) match μT\mu_T within 50 epochs. This agrees with the theory: the expected tree size (Lemma L.1) and condition (ii) of Prop. 3.1 do not depend explicitly on dd. We will add κ\kappa, endpoint MMD, tree size and runtime to App. O.4.

The image experiments use F=0F=0 on purpose. The SOTA SB baselines (DSBM, LightSB, LightSBB-M, DeepGSB, BranchSBM) do not implement law-dependent running costs, so a fair comparison requires the classical bridge. For F≡0F\equiv0, Eq. (10) has K=0K=0, so Φ=φθ(x+WT−t)\Phi=\varphi_\theta(x+W_{T-t}) and cc plays no role.

Reviewer_XE56OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a novel generative modeling framework based on nonlinear Schrödinger bridges (dubbed SBNL), which extend classical optimal transport to handle complex interactions where the running cost depends on the evolving probability distribution (marginal law). To solve the resulting coupled forward-backward partial differential equations (PDEs), the authors derive a simulation algorithm using marked branching diffusions, assigning "white" and "black" particles to the two potentials. By leveraging a Maximum Mean Discrepancy (MMD) objective, the model trains these coupled potentials jointly using stochastic gradient descent rather than expensive alternating refits. The method is shown to be embarrassingly parallel, highly scalable to high dimensions, and capable of recovering classical Schrödinger bridges while incorporating mean-field interactions like convex penalties.

Critical strengths
  1. the paper is well written and easy-to-follow
  2. The use of branching diffusions allows for unbiased Monte Carlo estimation that scales well to high-dimensional problems, avoiding the curse of dimensionality that plagues grid-based PDE solvers.
  3. By training the coupled potentials jointly using an MMD-based objective via stochastic gradient descent, the method bypasses the computationally heavy and error-prone iterative proportional fitting (IPF) typically used in classical bridge models.
Critical weaknesses
  1. While the branching representation is elegant, it still relies on a truncated power series (polynomial nonlinearity) to make the arbitrary PDE analytically tractable for the branching mechanism, which might limit expressivity for highly complex, non-analytic cost functions.
  2. Generating samples requires simulating numerous branching trees (Monte Carlo particles) and using automatic differentiation to estimate the score field, which may be significantly slower than a standard feedforward pass in conventional diffusion models.
  3. The training and sampling procedures involve intricate machinery (marked branching, Feynman-Kac representations, time-reversed coupling, and reweighted kernels) that are substantially more complex to implement and tune compared to standard score-matching models.
Questions for authors
  1. How does the computational time required for sampling with branching diffusions compare to standard diffusion model sampling, and does it scale efficiently to very high dimensions 64x64?
  2. How sensitive is the training stability and final generation quality to the choice of the reference density and the truncation order of the polynomial nonlinearity?
  3. Why did you choose intricate machinery (marked branching, Feynman-Kac representations, time-reversed coupling, and reweighted kernels) that are substantially more complex to implement? Is there any variants?
  4. Could you provide the plot of convergence depedence on time to realize hwo much time is necessary to converge?
2 comments

We thank the reviewer for the constructive assessment. We address each point below.

W1 – Polynomial truncation. (i) Several standard mean-field costs are exactly polynomial, so no truncation occurs: potential costs F=V(x)F=V(x) (K=0K=0) and the quadratic interaction f(m)=κ2∫m2f(m)=\frac{\kappa}{2}\int m^2, i.e. F=κmF=\kappa m (K=1K=1). All experiments of Sec. 5 are in these cases. (ii) For analytic FF, truncation is natural: the branching tree is a perturbative Feynman–Galton–Watson expansion where a diagram with nn branchings has weight O(cn)O(c^n) (App. M). FF is evaluated at the density m=φtφˉT−tm=\varphi_t\bar\varphi_{T-t}, which after whitening (App. N) is of order (2π)−d/2(2\pi)^{-d/2}, so the terms akmka_k m^k decay fast in kk and low orders dominate. (iii) Low-order costs are interpretable (e.g. κm\kappa m penalises concentration), which helps analyse the learned dynamics. We will add a remark on branching for non-polynomial nonlinearities (Henry-Labordère & Touzi, 2021).

W2 – Sampling cost. (i) For F=0F=0 and F=VF=V, K=0K=0: each tree is a single Brownian path with a multiplicative weight, i.e. plain Feynman–Kac Monte Carlo. (ii) For nonlinear FF, The tree-size lemma (App. L) bounds the expected tree size by 1+c(T−t)(1+2Kˉ)1−cTKˉ1+\frac{c(T-t)(1+2\bar K)}{1-cT\bar K}, independently of dd. The cost per Euler step is linear in dd and the MM trees are independent, so they run as one batched GPU call. Training takes ~30 s and generation ~3 s for 50 steps on a T4 (Sec. 5.1), the same order as diffusion samplers. (iii) Score-based models are a special case: App. H.2 shows DDPM is the exact F≡0F\equiv0 bridge with φˉ≡1\bar\varphi\equiv1 and μ0=μT∗N(0,TI)\mu_0=\mu_T\ast\mathcal N(0,TI). They only transport from noised data, and for general SBs they correspond to the first IPF iteration of DSB (De Bortoli et al., 2021). SBNL transports between arbitrary μ0,μT\mu_0,\mu_T without IPF and adds law-dependent costs, which have no score-matching counterpart.

W3 – Complexity. The components simplify in the common cases. With K=0K=0, branching reduces to Feynman–Kac Monte Carlo, and the time-reflected coupling is active only when FF depends on mm. The reweighted kernel k~=μ∗kμ∗\tilde k=\mu^{\ast}k\mu^{\ast} is importance sampling for the MMD between unnormalised densities. Training is plain SGD on a U-statistic. The general formulas are specialised in remarks (remarks of Sec. 2; App. H, K, M).

Q1 – Time and scaling. Cost is linear in dd; our largest setting is d=512d=512 (FFHQ ALAE). For images we work in latent space (App. N.2), as do the SB baselines, since pixel-space whitening is ill-conditioned. We will add a table of training and sampling wall-clock times for SBNL and all baselines, and sampling time as a function of dd.

Q2 – Sensitivity. The reference μ∗\mu^{\ast} only needs μ∗>0\mu^{\ast}>0 a.e. so that k~\tilde k stays characteristic; a Gaussian covering both whitened marginals worked as a default. Generation was stable over λ∈{102,…,105}\lambda\in\lbrace 10^2,\dots,10^5\rbrace, c∈{0.01,0.1,0.3}c\in\lbrace 0.01,0.1,0.3\rbrace, κ∈{1,10,100}\kappa\in\lbrace 1,10,100\rbrace and five potential landscapes (Sec. 5.2; App. O.3, P.2). The experiments use exact polynomials, so we will add an ablation over μ∗\mu^{\ast} and over the truncation order KK for a non-polynomial analytic FF. As stated in Sec. 6, the nonlinear regime needs tuning of cc and the kernel bandwidth; the remark at the end of App. L gives the conditions on c,T,κc,T,\kappa that keep the variance finite.

Q3 – Why this machinery. Each component answers one obstacle. The main theorem characterises the optimum by coupled semilinear PDEs and a nonlinear Sinkhorn system. Grid solvers (Guéant, 2012) suffer from the curse of dimensionality, and IPF/IMF assume linear propagation. Branching gives a mesh-free, unbiased estimator of φ\varphi even with the reflected product φtφˉT−t\varphi_t\bar\varphi_{T-t}. The MMD turns the Sinkhorn fixed point into one loss with LMMD=0\mathcal L_{\rm MMD}=0 iff it is solved (MMD-objective proposition, Sec. 4), trained jointly. Variants exist for individual blocks: other discrepancies in place of MMD, a Wasserstein soft terminal cost (App. F), closed-form heat kernels when F=0F=0, or score-matching for the linear part with branching only for nonlinear corrections (Sec. 6). We know of no alternative unbiased, mesh-free estimator for the coupled system that keeps full generality.

Q4 – Convergence. The paper reports Law(XT)\mathrm{Law}(X_T) and both potentials over epochs (App. P.1), and the 50D moments converge within 50 epochs (App. O.4). We will add loss and terminal-distance curves against wall-clock time for every experiment.

Reviewer_XE56OpenReview ↗

Answer

Thank you, you for your answers. Some answers I found well, however I'd like to keep my score

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.