- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies the problem of nonparametric drift estimation for stochastic differential equations (SDEs) from multiple independent trajectories corrupted by additive observation noise. VEST alternates between Gaussian Markov smoothing (E-step) and neural drift learning (M-step), using conditional posterior-increment targets and deterministic cubature integration to incorporate posterior uncertainty into drift regression. Both updates are safeguarded by an Armijo-type acceptance criterion on a shared variational objective. The core motivation is to simultaneously address the issues of "observation noise amplification" and "discarding latent-state uncertainty".
Critical strengths
- Good ablation experiment design. The "Without safeguards" ablation clearly demonstrates the critical role of the safeguard mechanism for numerical stability.
- The paper provides a well-structured and thorough theoretical analysis in Section 4 and Appendix E. Table 1 shows that VEST achieves great drift estimation accuracy, often by a substantial margin.
- The paper correctly identifies a critical flaw in naive increment regression (Section 2): the -scaled conditional noise term that biases the estimate when regressing noisy increments. By formulating a marginal-likelihood objective and optimizing an ELBO, VEST addresses this issue from principles.
Critical weaknesses
- Theorem 2 assumes exact gradients of the variational objective for its convergence guarantees. However, the actual M-step uses gradients of the cubature-approximated loss (Eq. 10), which is only an approximation of the true objective. Whether the theoretical guarantees remain valid for the practically?
- VEST's computational cost per outer iteration is likely higher than SING and SDE Matching, since its outer iterations each require running IEKS on every trajectory, multiple neural inner steps, and Armijo backtracking for both updates. However, the paper does not report runtime, parameter counts, or wall-clock time for convergence.
Questions for authors
Please refer to 'Critical Weaknesses'.
Convergence guarantees for the cubature objective and comparisons of runtime, parameter counts, and time to target accuracy
Thank you for your careful reading and constructive feedback! We respond to your concerns below.
W1. FΔ denotes the discrete ELBO defined using exact Gaussian expectations, whereas FΔ denotes the deterministic objective obtained by replacing the expected transition terms with finite weighted sums over cubature points. The loss in Eq. (10) indeed uses a cubature approximation. Our convergence analysis is formulated for the resulting cubature objective FΔ.
Assumption (A4) requires exact gradients of the cubature objectives—FΔ,n for the Gaussian update and FΔ(ϕk+1,⋅) for the neural update. With ϕk+1 fixed, the M-step loss in Eq. (10) satisfies −∇θLk(θ)=∇θFΔ(ϕk+1,θ). Here, the exact-gradient requirement in Theorem 2 applies to the cubature objective corresponding to the loss in Eq. (10). We apologize that this distinction was not sufficiently clear in the original presentation.
W2. We first compare VEST and SING on the same NVIDIA L40 across seven systems in Table 1, using five seeds per system. Taking the one-dimensional OU system as an example, a complete run takes 8.34 s for VEST and 42.68 s for SING (Table F.9). To measure the cost of each step, we run a separate timing experiment on the same OU system. Both methods run for 16 outer iterations, with a budget of 10 E step iterations and 10 M-step updates per round. Initial JAX compilation takes 5.77 s for VEST and 31.67 s for SING. Over rounds 2–16, the E step takes 0.0080 s per round for VEST and 0.0342 s for SING. This difference in E-step runtime can be partly explained by the update mechanisms: VEST linearizes the dynamics and computes posterior moments directly by Kalman smoothing, whereas SING's update requires gradient evaluations for both the likelihood and transition terms. The M step takes 0.0013 s per round for SING and 0.0030 s for VEST, including approximately 0.0023 s for Armijo backtracking (Table F.11).
Moreover, we compare VEST and SDE Matching on the same GPU across all ten Table 1 settings, using five seeds per setting. Under Table 1 configurations, mean runtime ranges from 8.33 to 25.01 s for VEST and from 40.72 to 42.15 s for SDE Matching with 4,000 training updates (Table F.9). On the 1D OU system, VEST and SING each use a drift network with 24 hidden units and 73 trainable parameters. SDE Matching uses 100 hidden units and 301 trainable parameters in its drift network, with a further 51,402 parameters in its posterior network (Table F.11).
Finally, we compare the time needed to reach a drift RMSE of 0.1 or lower on the 1D OU system. VEST and SDE Matching reach this target on all five seeds, taking 8.09 s and 9.81 s on average, respectively. SING reaches it on four of five seeds within 400 rounds, averaging 48.56 s over those four.