acceptodds
Under review as a conference paper at ICLR 2027

A simulation-free diffusion-based method for solving nonlinear PDEs

Abstract

Deep learning methods based on backward stochastic differential equations (BSDEs) provide an effective approach to solving nonlinear partial differential equations (PDEs), particularly in high dimensions. By exploiting stochastic representations of PDE solutions, these methods encode second-order differential operators through the dynamics of diffusion processes, avoiding the explicit computation of second-order derivatives. Their application, however, typically requires sequential path simulation and, on bounded domains, accounting for random exit times, which introduces additional numerical error. A key observation is that the stochastic representation extends beyond the Markovian setting, as Itô's formula remains valid for general continuous semimartingales. We exploit this observation using Brownian bridges, whose endpoints can be prescribed and whose samples can be generated in closed form. Choosing the terminal endpoint on the domain boundary eliminates the need to account for random exit times, while closed-form sampling enables fully parallel generation of the stochastic inputs. This yields a diffusion-based deep-learning method for nonlinear PDEs that avoids sequential path simulation and random exit times, while retaining the advantages of stochastic representations in high dimensions.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
approximation of semilinear PDEs, BSDEs, SDEs, Brownian bridges

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_X6RcOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a simulation-free diffusion-based learning framework for semilinear PDEs.

Using an integral residual derived from It^o's formula, the authors construct training paths through explicitly sampled Gaussian processes and endpoint-conditioned Brownian bridges.

The main algorithmic contribution is a boundary-anchored sampling strategy that incorporates boundary values directly into the interior training loss, together with geometric routing and chained bridges.

Experiments on perforated domains, metastable committor problems, and a high-dimensional HJB equation demonstrate promising computational performance compared with existing PINN, BSDE, and diffusion-based methods.

Critical strengths
  1. Meaningful algorithmic novelty.

The combination of explicitly sampled Gaussian bridges, prescribed boundary endpoints, and geometric path routing provides a useful training strategy for challenging PDEs. Although the stochastic tools are classical, the overall algorithmic construction has independent value.

  1. Well-motivated boundary treatment.

Incorporating known boundary values directly into the interior residual is particularly appealing for problems with complicated geometries or multiple boundary components.

Critical weaknesses
  1. Insufficient justification of boundary-sampling consistency.

The authors acknowledge that a Brownian bridge may leave the domain before reaching its prescribed boundary endpoint and propose anchor placement, node rejection, and truncation as remedies.

However, node rejection only constrains sampled grid points. It neither guarantees that the continuous path remains inside the domain nor preserves the original Gaussian bridge law.

Since the residual representation requires appropriate domain conditions, the consistency of the implemented boundary-handling procedure with the target Dirichlet PDE remains insufficiently established.

The reported low node-freezing rate does not quantify the resulting approximation bias.

This does not prove that the algorithm fails, but it leaves an important gap between the theoretical representation and the implemented method.

  1. Inconsistency in the perforated-square PDE.

Appendix D.2 specifies

\qquad h(x,v,z)=\frac{1}{2}\left(\lVert z\rVert^2-q(x)\right),$$ where $$z=\sigma^\top\nabla V, \qquad q(x)=\lVert\nabla V^\star(x)\rVert^2, \qquad \Delta V^\star=0.$$ These expressions imply $$\mathcal{L}V^\star +h(x,V^\star,\sigma^\top\nabla V^\star) =\frac{1}{2}q(x),$$ which is generally nonzero. The written driver is therefore inconsistent with the stated analytical reference solution. This may be a manuscript error rather than an implementation error. Nevertheless, clarification is essential because this experiment provides important evidence for the proposed boundary-coupling mechanism.
Questions for authors

Q1. Boundary-sampling consistency.

Can the authors justify that the implemented boundary-anchored sampling procedure remains consistent with the original Dirichlet PDE?

In particular, can they quantify the effect of early boundary crossings and node rejection through a suitable consistency argument or a targeted numerical study?

A comparison across bridge discretizations or boundary-handling strategies would be informative.

Q2. Perforated-square implementation.

What is the exact nonlinear driver used in the perforated-square implementation?

Can the authors verify that the analytical reference solution satisfies the PDE corresponding to the implemented training loss?

An explicit formula and a short residual calculation or implementation excerpt would resolve this concern.

6 comments

Thank you for the careful reading, for checking the driver, and for recognizing the algorithmic value of anchoring the bridges on the boundary.

W1/Q1: boundary-sampling consistency. Reviewer Sjqj raised this too (see W4 there); in short:

  • Rejection changes the path law, but this alone does not bias the loss. Prop. 2.2 is a pathwise identity: by Itô's formula, R=∫t0Tρφ(Zs,s)dsR=\int_{t_0}^{\mathcal T}\rho_\varphi(Z_s,s)\mathrm{d}s holds along every path that stays in Ω‾\overline\Omega and has quadratic variation σσ⊤\sigma\sigma^\top, regardless of the law from which the path is drawn. In particular, the exact solution, for which ρV≡0\rho_V\equiv0, has zero residual on every such path. Changing the path law, as rejection does, therefore only reweights the paths in the loss, similar to changing the distribution of collocation points in a PINN.
  • The continuous target depends on the form of rejection. As the grid is refined, whole-path rejection heuristically approaches the first-passage bridge, which is Brownian motion conditioned to leave Ω\Omega for the first time at time rr and at the point zrz_r; averaged over the exit law, it is the stopped process (new Remark C.3). Per-node rejection, which our experiments use, heuristically approaches the Gaussian bridge (16) reflected at ∂Ω\partial\Omega, which keeps the bridge drift inside Ω\Omega and along which Itô's formula gives the same identity. Following Reviewer Sjqj, we corrected Prop. 2.4, the characterization of zero loss. It covers the first-passage bridge, and the new Remark B.6 extends it to the reflected bridge. We do not formally prove that the discrete sampler converges to either law.
  • The new Fig. 6 quantifies the effect of crossings and rejection. At the exact solution and with rejection, the loss decreases roughly like n−1/2n^{-1/2} (per-node) and n−1n^{-1} (whole-path) as the bridges are refined: at N=64N=64, for n=5→20→80→320n=5\to20\to80\to320 sub-steps per hop (time steps per bridge), from 0.027 to 0.0087 to 0.0040 to 0.0018 with per-node rejection and from 0.023 to 0.0041 to 0.0015 to 0.00033 with whole-path rejection, with the same trend at every NN. Without rejection it does not decrease.
  • New comparison, as you suggest. We trained with per-node, whole-path and no rejection, at 5 and 20 sub-steps, for N=16N=16, 6464 and 144144 (3 seeds). The two rejection schemes agree to within about 25%, and no rejection is never better than per-node rejection and up to 1.75× worse.

The table gives the relative L2L^2 error after training at the paper's budget, as the mean over 3 seeds:

NNnnper-nodewhole-pathnone
1650.0330.0310.039
16200.0290.0230.051
6450.0470.0480.051
64200.0630.0630.073
14450.0210.0260.022
144200.0340.0410.040

W2/Q2: driver. You are right that the formula in App. D.2 is inconsistent. It is a typo in the paper; the code and the reference solution agree. Like the main-text PDE (25), the implementation uses the raw gradient:

h(x,v,∇V)=λ(∥∇V∥2−q),i.e.h(x,v,z)=λ(∥z∥2/σ2−q)  for z=σ⊤∇V,h(x,v,\nabla V)=\lambda\bigl(\Vert \nabla V\Vert ^2-q\bigr),\quad\text{i.e.}\quad h(x,v,z)=\lambda\bigl(\Vert z\Vert ^2/\sigma^2-q\bigr)\ \text{ for } z=\sigma^\top\nabla V,

with λ=12\lambda=\tfrac12 and σ2=2\sigma^2=2, so the generator is Δ\Delta. Since ΔV⋆=0\Delta V^\star=0 and q=∥∇V⋆∥2q=\Vert \nabla V^\star\Vert ^2, the residual at V⋆V^\star is 0+12(q−q)=00+\tfrac12(q-q)=0.

We evaluated the implemented residual σ22ΔV⋆+h(x,V⋆,∇V⋆)\tfrac{\sigma^2}{2}\Delta V^\star+h(x,V^\star,\nabla V^\star) in float64 at 20,000 uniform points of Ω\Omega. Its absolute value is at most 1.8⋅10−121.8\cdot10^{-12} for N=16,36,64,144N=16,36,64,144. The formula as printed leaves q/2q/2, as you computed. The revision corrects App. D.2, including the typo "L=ΔtL=\Delta t", which should read L=ΔL=\Delta.

Thank you again for your review. Your check of the driver and your question on boundary sampling led to the corrected App. D.2 and to the new analysis of rejection in App. C.3, and they have significantly improved the manuscript. Please let us know whether this resolves your concerns about boundary sampling, or whether you have further questions. We would be glad to discuss them.

Reviewer_X6RcOpenReview ↗

Follow-up on Boundary Sampling in the Committor Experiment

Thank you for the additional analysis and experiments. I understand that Remark B.6 establishes the residual identity for an ideal reflected bridge under suitable assumptions, while the convergence of the implemented per-node rejection scheme to such a process is not formally established.

My remaining question concerns the committor experiment specifically. Since the Gaussian bridges are anchored on ∂A\partial A or ∂B\partial B, could the authors clarify whether they can provide a consistency argument, or a bound on the effect of node rejection and freezing, for the discrete loss used in this experiment? In particular, under what conditions would refining the time grid preserve the true committor as a zero-loss solution in the limit?

If a formal result is outside the scope of this revision, a limited grid-refinement or boundary-handling comparison on one representative committor instance would also help assess the practical effect.

Thank you for the follow-up question. We agree with your summary of Remark B.6. Below, we first clarify how the committor experiment handles the absorbing sets and then answer your question for the continuous objective and for the discrete loss. Changes since our previous revision are marked in red.

Boundary handling in the committor experiment. This experiment uses neither node rejection nor freezing. Each chained bridge is stopped at its first node in A∪BA\cup B, and the exit residual evaluates gg at that node. Since gg equals 00 on all of AA and 11 on all of BB, this is the exact boundary value of the set in which that node lies. Detecting also the entries between two nodes, with a crossing test, changes the loss at the reference committor by less than 1% (next comment). We now state this explicitly in App. D.3.

Continuous objective. The continuous counterpart of this sampler is the chained Gaussian bridge stopped at its first exit from Ω\Omega, with gg evaluated at the exit point, i.e., the stopped construction of Def. B.3. Its terminal anchor lies on ∂Ω\partial\Omega, but the path may leave Ω\Omega before reaching it. The proof of Prop. 2.4 needs only the weaker requirement of Remark B.6, Qa(Zra−δ∈Ω)>0\mathbb Q_a(Z_{r_a-\delta}\in\Omega)>0 for every δ\delta, which this law satisfies: given that the path has not left Ω\Omega by ra−δr_a-\delta, its law on [t0,ra−δ][t_0,r_a-\delta] is the Gaussian bridge conditioned on staying in Ω\Omega, whose support is given by Lemma B.5, and in the elliptic case Step 3 needs only T>t0\mathcal T>t_0. A new paragraph at the end of Remark B.6 states this. Hence, for isotropic σ\sigma, smooth ∂A\partial A and ∂B\partial B, and a boundary term with full support, a function has zero continuous loss if and only if it solves the committor equation.

Discrete loss. Each chained bridge ends at its terminal anchor on ∂A\partial A or ∂B\partial B, so a path that stays in Ω\Omega until then needs no exit detection. Truncation matters only for paths that enter A∪BA\cup B earlier, and the implemented loss detects such an entry only at the nodes. It therefore misses entries between nodes and places the final node of a truncated path up to one step inside A∪BA\cup B, the O(Δt)O(\sqrt{\Delta t}) bias of exit times that are monitored only at discrete times; in the interior, the time discretization adds the O(Δt)O(\Delta t) term of Prop. C.1. A formal convergence result for the discrete objective is outside the scope of this revision. As you suggest, we instead report a grid-refinement and boundary-handling study on the committor instance of Sec. 3.2 (next comment).

Reviewer_SjqjOpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes neural losses for semilinear elliptic and parabolic PDEs that avoid explicit Hessians and numerical time stepping of the original SDE. It evaluates an Itô residual along continuous semimartingales whose quadratic variation matches the PDE diffusion, with the PDE drift included as an explicit first-order term. When the diffusion coefficient is independent of the state, Brownian motion and Gaussian bridges with prescribed endpoints provide directly sampled training nodes. Interior-boundary and boundary-boundary variants substitute known boundary values into the residual. The stochastic integral still requires discretization.

The theoretical results include a residual identity, a zero-loss characterization, short-window limits, and discretization analysis. Experiments compare learning losses on a perforated square, a metastable committor problem, and a 100-dimensional Hamilton-Jacobi-Bellman (HJB) equation.

Critical strengths

Strength 1: The residual addresses concrete computational bottlenecks

Equation (9) encodes the second-order operator through quadratic variation and explicitly handles the PDE drift. Equations (16)-(18) and Algorithm 1 provide directly sampled Gaussian nodes, with a residual requiring only values and first derivatives. With state-independent diffusion and computable covariance, this avoids Hessian evaluation and numerical integration of the original SDE, offering clear implementation value.

Strength 2: Boundary data participate directly in interior training

The interior-boundary and boundary-boundary bridges in Section 2.2 insert known boundary data into path residuals. Proposition 2.7 gives short-window analyses under the corresponding assumptions, and Figures 6-7 examine boundary-regression controls and penalty-weight sensitivity. This offers a concrete approach to reducing reliance on separate boundary penalties.

Strength 3: Experimental controls help interpret the results

Appendix D.1 specifies wall-clock budgets, D.2 and D.4 explain reference constructions, and Figures 6-7 provide boundary-regression controls and weight scans. Equation (58) in D.3 discloses the committor's intrinsic two-dimensional structure; its reference is a two-dimensional finite-difference solution lifted to the ambient dimension. The paper also reports cases where PINN is more accurate and HJB performance is comparable to the unbiased BSDE baseline. These details help assess accuracy and cost.

Critical weaknesses

Weakness 1: Admissibility does not imply the general zero-loss characterization

Definition 2.1 leaves finite variation unrestricted, and the coefficient assumptions do not require nondegenerate diffusion. Yet Proposition 2.4 claims zero-loss equivalence for every admissible process. Appendix B.2 (p. 14) infers a pointwise zero PDE residual from a zero integral over a fixed window, without justification.

For a counterexample, take the domain (-1,1), terminal time 1, zero drift, diffusion, and nonlinearity, and zero terminal and boundary data. The PDE then requires a zero time derivative and has only the zero solution. Set φ(x,t)=(1−t)sin⁡(πx)\varphi(x,t)=(1-t)\sin(\pi x). For full-support anchors with x in the domain and a between 0 and 1, use Zs=xcos⁡(2πs−a1−a)Z_s=x\cos\left(2\pi\frac{s-a}{1-a}\right) on [a,1], keeping the path at x beforehand. This continuous, finite-variation path stays inside the domain and is admissible for zero diffusion. The residual integrals over the two half-periods cancel, and the terminal and boundary losses vanish, although the function is not a PDE solution. The general converse therefore fails; additional conditions must cover the implemented samplers.

Weakness 2: Terminal supervision differs in the HJB comparison

In Section 3.3, Figure 5, and Appendix D.4 (p. 26), the proposed method receives terminal value-and-gradient supervision with batch size 512, while PINN receives value-only supervision with batch size 50. The accuracy gap in Figure 5 therefore cannot be attributed to the interior loss alone. Terminal supervision and sampling volume need to be matched to isolate its benefit.

Weakness 3: The perforated-domain BSDE baseline misses exits

Appendix D.2 (p. 22) reports that, at N = 64, 32% of BSDE paths cross a hole without stopping at least once. The baseline detects exits only at discrete nodes; bisection of detected crossing segments cannot recover missed crossings. The accuracy and cost gap in Figure 3 may therefore include exit-handling error. Reliable detection or a targeted time-step convergence check is needed to distinguish its effect.

Weakness 4: The bounded-domain discrete sampler is insufficiently connected to the continuous objective

Proposition 2.2 requires paths to remain in the domain during integration, but Section 2.2 and Appendix C.3 acknowledge that a boundary endpoint cannot exclude earlier leave-and-return excursions. Appendix D.2 uses five substeps per hop and up to eight retries for out-of-domain nodes. Nodewise rejection changes the Gaussian sampling distribution and cannot exclude crossings between accepted nodes; the frozen-node fraction measures only exhausted retries. Interpreting Figures 2-3 requires identifying the continuous objective approximated by this discrete loss or quantifying sampling error at the actual settings.

Questions for authors
  1. What additional conditions exclude path-integral cancellation and make Proposition 2.4's converse valid? Can validity be established for the implemented Gaussian samplers, window distribution, and training objective without an unproved transition from a fixed-window objective to shrinking-window objectives?
  2. With terminal value-and-gradient supervision, terminal batch size, and wall-clock budget matched, do the windowed residual's accuracy and convergence advantages over PINN persist? If controlled results are available, please report them to distinguish interior-loss and terminal-supervision effects.
  3. How do BSDE accuracy and cost in Figure 3 change after improving hole-crossing detection or performing a targeted time-step convergence check? This would quantify the effect of the reported missed exits on the current gap.
  4. Which continuous objective does the discrete loss after nodewise rejection approximate? Can consistency analysis or diagnostics at the implemented settings quantify how crossings between nodes, rejection, and freezing affect the loss and Figures 2-3?
5 comments

Thank you for a very careful review, and for the clear account of what the bridge loss offers. All four points are correct; here is what we changed.

W1/Q1: zero-loss characterization (Prop. 2.4). Your counterexample is right, and so is the objection to App. B.2: for a fixed sampler, "divide by ℓ\ell" is not justified. The revised proposition states explicit conditions on the sampling measure (Def. B.3, explained in the new Remark 2.5) and is proved without any limit in ℓ\ell:

  • (a) Support. Given its anchors and its path after any time r−δr-\delta before the window end rr, the support of the path law on [t0,r−δ][t_0,r-\delta] contains every continuous path through the anchors that stays in Ω\Omega and joins the given path. A new lemma (B.5) shows this when σσ⊤⪰λI\sigma\sigma^\top\succeq\lambda I for Brownian motion, and for bridges and chains of bridges on Rd\mathbb R^d or conditioned to stay in Ω\Omega. Without the conditioning, these processes are Gaussian, and by a standard result on Gaussian measures (Bogachev, 1998, Thm. 3.6.1), their support contains every continuous path through the anchors; conditioning on staying in Ω\Omega keeps the paths of the support that stay in Ω\Omega. Condition (a) also holds for the first-passage bridge and, by the new Remark B.6, for the reflected bridge, the two laws that our exit bridges approximate (see W4). All experiments use σσ⊤=σ2I\sigma\sigma^\top=\sigma^2 I. Your example has σ=0\sigma=0 and violates (a).
  • (b) Windows, parabolic case only. All windows have the same length ℓ\ell, truncated at TT: T=(t0+ℓ)∧T\mathcal T=(t_0+\ell)\wedge T. Our proof uses this. The start times must also cover [0,T][0,T], as Def. 2.1 already requires: if no window reaches TT, adding to φ\varphi an ℓ\ell-periodic function of tt that vanishes at TT leaves the loss unchanged on Rd\mathbb R^d (new Remark B.4).

The proof is in App. B.2. It also covers bridges, whose drift becomes unbounded as they approach their terminal anchor. The short-window limit to the PINN loss is now stated under an explicit condition (c): over windows of length ℓ→0\ell\to0, the paths must stay close to their start. This holds for Brownian motion and for bridges whose terminal anchor lies within O(ℓ)O(\sqrt\ell) of the start, as in Sec. 3.3; exit bridges have the boundary limits of Prop. 2.8 instead.

In the submitted Fig. 5, the HJB runs drew t0≤T−ℓt_0\le T-\ell, so no window reached TT, which violates the full-support requirement. The revised Fig. 5 draws t0∈[0,T]t_0\in[0,T] and truncates the windows at TT. This improves the bridge at every time step of the right panel: at Δt=10−3\Delta t=10^{-3} from 2.8e-4 to 1.4e-4 with clipping and from 3.2e-4 to 1.4e-4 without (3 seeds), and by a factor of 16 at Δt=10−2\Delta t=10^{-2} (from 5.5e-3 to 3.5e-4). App. D.4 reports this comparison.

W2/Q2: HJB terminal supervision. You are right that PINN matched only the terminal value, on 50 points, while all other losses matched the value and the gradient on 512 points (64 for Heun). The revised Fig. 5 uses the matched PINN.

With this experiment (Sec. 3.3), we want to highlight that our windowed loss is more accurate than recent BSDE methods designed to reduce the time-discretization error, namely the method of Seo et al. (2026) and the Heun scheme of Park & Tu (2025), at a ten times finer time step that only the windowed residual can afford. The right panel of Fig. 5 shows this as converged error against Δt\Delta t. PINN has no time discretization at all and serves as a reference, not as the comparison the experiment is about. Sec. 3.3 now states this purpose and the result explicitly.

The ablations below show that neither the terminal term nor the other settings we varied changes the comparison. The table reports the error after 8 h with the protocol of Fig. 5, as the geometric mean over 3 seeds (* marks a single seed):

LossTerminal term (batch)ClippedUnclipped
Bridge / diffusion, windows truncated at TT (revised Fig. 5)value + gradient (512)1.4e-41.4e-4
Bridge / diffusion, windows ending before TT (submitted Fig. 5)value + gradient (512)2.8e-43.2e-4
Bridge / diffusion, windows ending before TTvalue only (512)2.6e-43.9e-4*
PINN (submitted Fig. 5)value only (50)8.6e-51.7e-2 (diverges)
PINNvalue only (512)7.7e-5*1.7e-2* (diverges)
PINN (revised Fig. 5)value + gradient (512)8.7e-51.7e-2 (diverges)
BSDE, Seo et al. (2026)value + gradient (512)2.2e-42.6e-4
BSDE, 5-samplevalue + gradient (512)1.3e-31.4e-3
BSDE, Euler–Maruyamavalue + gradient (512)3.8e-33.8e-3
BSDE, Heun (Park & Tu, 2025)value + gradient (64)1.7e-3*2.1e-3

Across seeds, every three-seed entry varies by less than a factor of 1.7. Unlike PINN, the bridge needs no gradient clipping.

W3/Q3: BSDE missed exits. We agree that bisection places a detected exit on ∂Ω\partial\Omega but cannot recover a missed one.

  • The exit point was already placed in the BSDE's favor. When a node lands outside Ω\Omega, we compute the point where the straight segment from the previous node crosses ∂Ω\partial\Omega. The path ends at this point, and the last time step is shortened accordingly, so that gg and the stochastic integral are evaluated at the same boundary point. The textbook alternative keeps the full last step and evaluates gg at the boundary point nearest to the first node outside Ω\Omega; at the same budget it is 1–8% worse (new runs, 3 seeds per NN).
  • Missed exits. We added the standard Brownian-bridge crossing test between nodes (Gobet, 2000; 2001), now stated in App. D.2. It makes the BSDE baseline considerably stronger: at the same Δt\Delta t and wall clock, its error drops by a factor of 2.5–5. The bridge remains the most accurate method where the geometry is hardest, by a factor of 1.14 at N=64N=64 and 1.76 at N=144N=144, and is on par for N≤36N\le36 (dashed curve in Fig. 3). It needs no such test for its boundary value: its exit point is prescribed, so gg is always evaluated at the correct boundary point. Excursions of the bridge between nodes affect only the residual, and their effect vanishes as the step size goes to zero (W4).
  • Where BSDE fails for another reason, the test does not help. On the committor of Sec. 3.2, the BSDE loss needs a realized exit from a metastable potential, 2.5⋅1042.5\cdot10^4 sequential steps per update, and its error stays above 0.25 at every dimension, against 0.08 for the bridge, which draws its nodes in closed form. A crossing test would not shorten these trajectories.
  • Time step. Refining the node-only BSDE lowers the bias of its objective at the exact solution, L(V⋆)/L0\mathcal L(V^\star)/\mathcal L_0, from 0.7–1.3 at Δt\Delta t to 0.02–0.03 at Δt/16\Delta t/16, but each update costs proportionally more. At the same budget, the crossing test at Δt\Delta t is therefore 2.5–2.7× more accurate than the node-only BSDE at any of these time steps.

The table gives the relative L2L^2 error at the budget of Fig. 3, as the mean over 3 seeds (8 for the bridge at N=144N=144):

MethodN=16N=16N=36N=36N=64N=64N=144N=144
BSDE, node-only, Δt\Delta t (submitted Fig. 3)0.1780.1520.2810.078
BSDE, node-only, Δt/4\Delta t/40.0900.0890.1440.077
BSDE, node-only, Δt/16\Delta t/160.0850.1280.1940.174
BSDE, crossing test, Δt\Delta t (revised Fig. 3)0.0340.0340.0540.031
Bridge (Fig. 3)0.0330.0330.0470.018

Finally, note that Sec. 3.1 mainly compares the bridge with PINN. The crossing test and the time-step study change only the BSDE curve of Fig. 3 and leave this comparison unchanged.

W4/Q4: which continuous objective. You are right that the submitted paper does not identify the continuous objective of the discrete loss after rejection. The revision does so in App. C.3, which we summarize here.

  1. The residual (9) depends only on the sampled path: it evaluates φ\varphi, ∇φ\nabla\varphi and the PDE coefficients bb, hh along it. How the path was generated, by the drift of the bridge (16) or by the redraws of rejection, does not enter. By Itô's formula, the identity of Prop. 2.2 therefore holds for every path law in Ω‾\overline\Omega with quadratic variation σσ⊤\sigma\sigma^\top. Rejection changes which paths the loss averages over, but not this identity.
  2. The continuous objective depends on the form of rejection (new Remarks C.3 and B.6). Whole-path rejection heuristically approaches the first-passage bridge, Brownian motion conditioned to leave Ω\Omega for the first time at time rr and at the point zrz_r, in the same way as a random walk conditioned to stay positive approaches the Brownian meander in one dimension (Iglehart, 1974; Bolthausen, 1976). Per-node rejection, which our experiments use, heuristically approaches the Gaussian bridge (16) reflected at ∂Ω\partial\Omega, which keeps the bridge drift inside Ω\Omega. Prop. 2.2 holds along both laws, and the corrected Prop. 2.4 covers both: the first-passage bridge satisfies condition (a) (Lemma B.5), and Remark B.6 extends the proposition to the reflected bridge, which is not stopped at ∂Ω\partial\Omega. We do not formally prove that the discrete sampler converges to either law.
  3. Two discrete effects remain: crossings between accepted nodes, and frozen nodes, i.e., nodes whose redraws all fail (at most 2⋅10−52\cdot10^{-5} of all nodes at the paper's n=5n=5).

Their effect on the loss at the exact solution vanishes under refinement (new Fig. 6). At the exact solution and with rejection, L(V⋆)/L0\mathcal L(V^\star)/\mathcal L_0 decreases roughly like n−1/2n^{-1/2} (per-node) and n−1n^{-1} (whole-path) in the number nn of time steps per bridge. At N=64N=64, it falls from 0.027 at n=5n=5 to 0.0018 at n=320n=320 with per-node rejection, and from 0.023 to 0.00033 with whole-path rejection. Without rejection, it does not decrease, because nodes then land inside the holes, close to their centers, where V⋆V^\star is singular; on the paths that stay away from the centers, it decreases as well.

After training at the paper's budget (3 seeds, N=16/64/144N=16/64/144), per-node and whole-path rejection agree to within about 25%. Without rejection, the error is never smaller than with per-node rejection and up to 1.75× larger.

Thank you again for your detailed review. Your points led to the corrected zero-loss characterization, the identification of the continuous objective of our sampler, and the revised baseline comparisons, and they have significantly improved the manuscript. Please let us know whether the revision addresses your concerns, in particular on Prop. 2.4 and on the sampler, or whether you have further questions. We would be glad to discuss them.

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

This paper proposes a simulation-free diffusion loss for nonlinear parabolic and elliptic PDEs. Deep BSDE and diffusion losses encode the second-order operator through Itô’s formula along a diffusion, so they avoid explicit second derivatives, but they still need sequential path simulation and, on a bounded domain, a random exit time. The paper uses the fact that Itô’s formula holds for any continuous semimartingale whose quadratic variation matches σσ⊤\sigma\sigma^\top, while the drift can be chosen freely. When σ\sigma is state-independent, Brownian-bridge states are Gaussian and can be drawn in closed form, in parallel, with both endpoints prescribed. Placing the terminal endpoint on ∂Ω\partial\Omega inserts the boundary value into the interior residual and removes the need to approximate a hitting time. The resulting loss is zero exactly when the network solves the boundary-value problem, and in a short-window limit it recovers the PINN residual. On a perforated square, the exit-bridge error grows by less than a factor of 1.51.5 as the number of holes increases from 1616 to 144144, while the PINN error grows by a factor of four. On a metastable committor, the relative L2L^2 error stays near 0.080.08 from dimension 22 through 100100. On a 100100-dimensional Hamilton–Jacobi–Bellman equation, the interior bridge coincides with the diffusion loss but is sampled without an SDE solver. The construction requires state-independent σ\sigma, and an exit bridge is exact only when the path does not leave the domain and return before the prescribed horizon.

Critical strengths

I found the theory especially clear and instructive. The loss is derived for any continuous semimartingale whose quadratic variation matches σσ⊤\sigma\sigma^\top, so the drift can be chosen freely and the sampling process need not solve the original SDE. Brownian bridges then give closed-form Gaussian states, and anchoring an endpoint on ∂Ω\partial\Omega puts the boundary datum inside the interior residual instead of treating the exit time as a random hitting time. The validity result, the short-window limit that recovers the PINN residual, and the zero-variance property at the true solution make the construction easy to follow. The experiments are modest in scope, and the results that are included are presented and analyzed clearly, including the boundary-weight sweeps, the scaling with the number of holes and with dimension, and the effect of the time step.

Critical weaknesses

The empirical comparisons are limited to PINNs, BSDE losses, and the diffusion loss, together with the recent BSDE time discretizations of Park and Tu (2025) and Seo et al. (2026) on the Hamilton–Jacobi–Bellman example. The paper does not compare with more recent deep-learning PDE solvers, in particular operator-learning methods such as DeepONet (Lu et al., 2021) and the Fourier neural operator (Li et al., 2021). These methods target parametric solution maps rather than a single boundary-value problem, so a comparison would be most informative on a parametric family, for example the perforated domain as the holes vary or the committor as the potential varies. Can the authors add such a comparison on a recent benchmark, or explain where a single-solution diffusion loss is the appropriate baseline instead?

  • Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. "Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators." Nature Machine Intelligence, 3(3):218–229, 2021.
  • Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. "Fourier neural operator for parametric partial differential equations." International Conference on Learning Representations, 2021.
Questions for authors

Please see the concerns in the weakness.

1 comment

Thank you for the positive assessment, and in particular for finding the theory clear.

On operator learning, we explain why we do not compare rather than adding a comparison:

  • Different tasks. DeepONet and FNO learn a parametric solution map, usually from solver data on a grid. Our contribution is a loss for a single boundary-value problem in high dimension or on complicated domains.
  • Complementary, not competing. In a physics-informed DeepONet (Wang, Wang & Perdikaris, 2021), the PINN residual can be replaced by our windowed residual without other changes. The bridge loss would then bring its advantages, no second derivatives and boundary data inside the residual, to operator learning.
  • No data in our main setting. Grid-based solvers suffer from the curse of dimensionality: their cost grows exponentially with the dimension, so none exists for the 100-dimensional HJB equation. Training data for an operator network, and thus a data-driven comparison, are therefore not available there. The appropriate single-problem baselines are residual losses: PINN, BSDE and the diffusion loss.

The revision adds a paragraph on this scope and on the operator-learning extension to App. D.5 (Scope and open directions).

Thank you again for your review. Your question on operator learning led us to state the scope of the method more clearly and to describe how the bridge loss could be used in physics-informed operator learning. Please let us know if you have further questions or suggestions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response (to all reviewers and the AC)

We thank the reviewers for careful and constructive reviews. We are glad that all three see value in the bridge loss for the ICLR community: Reviewer cumL found the theory "especially clear and instructive", Reviewer Sjqj notes that the residual "addresses concrete computational bottlenecks" and that boundary data "participate directly in interior training", and Reviewer X6Rc sees "meaningful algorithmic novelty" and a "well-motivated boundary treatment".

Aim of the method. Residual losses for PDEs pay at least one of three costs: second derivatives (PINN), sequential path simulation (diffusion and BSDE losses), and random exit times (BSDE). The bridge loss avoids all three. Its residual needs only the network and its gradient, its nodes are drawn jointly in closed form, and anchoring the endpoint on ∂Ω\partial\Omega replaces the random exit time by a prescribed horizon, so that the boundary datum enters the interior residual itself. Our experiments test this in three settings:

  • Many boundary components. On the perforated square, the bridge error grows by less than a factor of 1.5 from 16 to 144 holes, while the PINN error grows fourfold.
  • Metastability in high dimension. On the committor, the bridge error stays near 0.08 from d=2d=2 to d=100d=100 and is flat in the boundary weight over six decades. PINN degrades by an order of magnitude, and BSDE stays above 0.25.
  • Time discretization in high dimension. On the 100-dimensional HJB equation and with the same time budget, the windowed residual is more accurate than the strongest recent BSDE method (Seo et al., 2026), at a ten times finer time step.

What we changed (all changes in blue):

  1. Zero-loss characterization, Prop. 2.4 (Sjqj). We closed the gap in the proof. The proposition now holds for a stated class of samplers, defined by a support condition that our Gaussian samplers and the first-passage bridge satisfy. It also covers bridges, whose drift blows up as they approach their terminal anchor. On the HJB equation, the windows now reach the terminal time, as the corrected proposition requires, which halves the bridge error (revised Fig. 5).
  2. Discrete sampler and continuous objective on bounded domains (Sjqj, X6Rc). New remarks identify the continuous objectives that whole-path and per-node rejection approximate, the first-passage and the reflected bridge, and the corrected Prop. 2.4 covers both. A new figure shows that the loss of the exact solution under our sampler decreases steadily as the bridges are refined, and does not decrease without rejection.
  3. Baselines (Sjqj). On the HJB equation, PINN now uses the same terminal term as every other loss. On the perforated square, we reran BSDE with a Brownian-bridge crossing test (App. D.2). The reply to Reviewer Sjqj adds an ablation table for the HJB equation and a time-step study for BSDE.
  4. Driver in App. D.2 (X6Rc). The formula in the paper contained a typo, which we corrected; the code and the reference solution agree (the residual at V⋆V^\star is below 2⋅10−122\cdot10^{-12}).
  5. Scope relative to operator learning (cumL). A new paragraph in App. D.5 explains why operator learning addresses a different task and how our loss could be used within it.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.