acceptodds
Under review as a conference paper at ICLR 2027

Brownian Smoothing via the Driving Noise

Abstract

Amortized posterior sampling offers a promising route to accurate and efficient smoothing. Learned smoothers can model complex trajectory posteriors, while drawing a sample still often requires full-trajectory modeling or a large population of particles. In this paper, we introduce Brownian Smoothing, which generates complete posterior trajectories in one rollout by conditioning the initial state and shifting the simulator's Gaussian driving noise. To maintain accuracy under one-rollout sampling, Brownian Smoothing is trained by conditional maximum likelihood on model simulations. And we prove that this objective is a forward-KL projection and that the optimal drift correction converges to the one induced by Doob's -transform under grid refinement. Empirically, Brownian Smoothing closely matches exact posterior statistics and achieves the best marginal and joint scores on nonlinear, high-dimensional benchmarks. It also substantially reduces sampling time and memory compared with global flow and score-based baselines.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian smoothing, amortized inference, stochastic differential equations, simulation-based inference, diffusion processes

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_MBYYOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes using an amortized method for estimating the smoothing posterior of stochastic differential equation with known dynamics given some noisy observations. They use an EM rollout to map the initial state and Gaussian increments to a trajectory. They then learn a control applied to the Gaussian increments which using targets from the simulations. After training it produces the approximate posterior trajectory with a single rollout. The paper also analyzes the approximation error of the fixed-covariance model and relation to the conditional trajectory law.

Critical strengths

The paper tackles and interesting problem. I find the strongest parts of the work to be:

  1. Shifting the means of the Brownian increments εk\varepsilon_k to incorporate the information of y≥ky_{\geq k} is interesting.
Critical weaknesses

While the paper tackles an interesting problem I have some concerns.

  1. The notation is often vague and ambiguous making it harder to read and understand what the authors did. For example the authors use vkv_k to denote the learned control but also later use vθv_\theta in Alg 1. The lack of consistent notation makes it harder to follow what Alg 1 does for example.
  2. My understanding reading the paper is that we learn this stochastic flow map Φ\Phi which is a deterministic function of the initial condition X0X_0 and Gaussians {εk}\{\varepsilon_k\}. It's not clear to me where this map is used in Algorithm 1. Also the the practical implementation still does an Euler-Maruyama rollout so I'm confused by the utility of this deterministic map from the initial condition and discretized Brownian path. Additionally, several works have explored learning such stochastic flow maps in law [1-3] and in path [4-5]. Of the former [1] focuses on the traditional neural SDE setting and [2-3] focus on generative-time models; and for the later two [4] use a shifted Legendre polynomial to model the Brownian path following [6] and [5] uses the classic Karhunen-Loeve sinusoidal approximation of the Brownian path. I'm a bit concerned by the first-order discretization employed here as this encoding seems difficult to use.
  3. There is also quite a bit of work on learning these latent SDEs for example [7-8] and it is not clear to me how this work differs and how it fits into the broader literature.

References

[1] Kiyohara, Naoki, Johns, Edward, and Li, Yingzhen (2025). “Neural Stochastic Flows: Solver-Free Modelling and Inference for SDE Solutions”. In: The Thirty-ninth Annual Conference on Neural Information Processing Systems. URL: https://openreview.net/forum?id=PrYDDxphym

[2] Holderrieth, Peter, Chen, Douglas, Eyring, Luca, Shah, Ishin, Anantharaman, Giri, He, Yutong, Akata, Zeynep, Jaakkola, Tommi, Boffi, Nicholas Matthew, and Simchowitz, Max (2026). “Diamond Maps: Efficient Reward Alignment via Stochastic Flow Maps”. In: arXiv preprint arXiv:2602.05993

[3] Potaptchik, Peter, Saravanan, Adhi, Mammadov, Abbas, Prat, Alvaro, Albergo, Michael S, and Teh, Yee Whye (2026). “Meta Flow Maps enable scalable reward alignment”. In: arXiv preprint arXiv:2601.14430

[4] Pan, Zhengkai, Potaptchik, Peter, Yao, Wenxi, Albergo, Michael S, and Pidstrigach, Jakiw (2026). “Itô maps for any-step SDEs”. In: arXiv preprint arXiv:2606.11156

[5] McCallum, Sam, Blasingame, Zander W., Herschell, Timothy, Rindtorff, Niklas, Tong, Alexander, and Foster, James (2026). “Strong Stochastic Flow Maps”. In: ICML 2026 Workshop on Structured Probabilistic Inference & Generative Modeling.

[6] Foster, James, Lyons, Terry, and Oberhauser, Harald (2020). “An optimal polynomial approximation of Brownian motion”. In: SIAM Journal on Numerical Analysis 58.3, pp. 1393–142

[7] Grigory Bartosh, Dmitry Vetrov, and Christian A. Naesseth. SDE Matching: Scalable and simulation- free training of latent stochastic differential equations. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pp. 3054–3070. PMLR, 2025.

[8] Henry D Smith, Brian L Trippe, and Scott W Linderman. Closing the approximation gap in simulation-free latent SDEs. arXiv preprint arXiv:2606.16138, 2026.

Questions for authors
  1. Why pass the guidance term via εl\varepsilon_l rather than additively to the drift as currently you multiplicatively scale by the diffusion coefficient σ(t,Xt)\sigma(t, X_t)? I suppose the argument is via Prop 3 that vv ideally is the score of hh but this wasn't very clear to me.
  2. Where do you use Φ\Phi your deterministic map from the joint initial condition and Brownian path?
Reviewer_7jKXOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes an amortized smoother for a supplied stochastic simulator. It learns a posterior initial distribution and observation-conditioned Gaussian mean shifts, allowing the simulator to generate trajectories consistent with the observations. The paper proves a forward-KL decomposition, bounds the approximation error introduced by fixed-covariance Gaussian increments, and connects the population-optimal control to Doob's h-transform. Experiments assess posterior accuracy and sampling efficiency across stochastic dynamical systems, including Lorenz-96 and Kolmogorov flow.

Critical strengths
  • The introduction and methodology are clearly written. The SDE sampling rollout formation and the training objective has a clear statistical interpretation.

  • The proposed Brownian Smoothing sampler is evaluated on scientific problems, including Lorenz and Kolmogorov flow, and shows promising improvements in posterior reconstruction and sampling efficiency over the evaluated neural baselines.

  • The evaluation examines not only marginal reconstruction but also temporal dependence. Exact posterior references, cross-time covariances, crossing probabilities, hitting-time distributions, and temporal-shuffling tests provide useful evidence about the quality of the sampled trajectories.

Critical weaknesses
  • Section 5 is difficult to follow and does not provide a clear and self-contained description of the experimental setup. The exact model architecture and training procedure are missing. The notation DD and 64264^2 is also inadequately introduced with the nonlinear benchmark results.

  • The paper does not establish its methodological novelty relative to standard score-based diffusion. A direct mathematical comparison is missing, despite the close correspondence between the proposed sampler and the standard controlled-diffusion formulation. Writing both samplers over an increasing sampling-time interval t∈[0,1]t\in[0,1], with bb denoting the respective base drift, makes this correspondence explicit.

    For standard diffusion, t=0t=0 corresponds to noise and t=1t=1 to data:

    Z0∼N(0,I),dZt=[b(t,Zt)+g(t)2sθ(Zt,t)]dt+g(t)dBt,0≤t≤1Z_0\sim\mathcal N(0,I),\qquad dZ_t=\left[b(t,Z_t)+g(t)^2s_\theta(Z_t,t)\right]dt+g(t)dB_t,\qquad 0\le t\le1

    For Brownian Smoothing, t=0t=0 and t=1t=1 correspond to the beginning and end of the physical trajectory, with time normalized for this comparison:

    X0∼qϕ(⋅∣y),dXt=[b(t,Xt)+σ(t,Xt)vθ(t,Xt,y>t)]dt+σ(t,Xt)dBt,0≤t≤1X_0\sim q_\phi(\cdot\mid y),\qquad dX_t=\left[b(t,X_t)+\sigma(t,X_t)v_\theta(t,X_t,y_{>t})\right]dt+\sigma(t,X_t)dB_t,\qquad 0\le t\le1

    The structural correspondence is

    σ=gI,vθ=gsθ,σvθ=g2sθ.\sigma=gI,\qquad v_\theta=gs_\theta,\qquad \sigma v_\theta=g^2s_\theta.

    Thus, the sampling-SDE form does not itself establish a new methodological contribution. The manuscript does not sufficiently explain what substantive advance follows beyond applying this familiar SDE path and score matching loss. The authors need to explain which part of the resulting algorithm constitutes the claimed novelty.

I am willing to increase my score if the weakness and questions can be addressed.

Questions for authors
  • Could the authors provide visual comparisons of sampled Lorenz-96 trajectories and Kolmogorov flow fields against the observations, simulated ground truth, and baseline samples?

  • What is the total rollout and time step resolution for Lorenz and Kolmogorov flow? How robust is the sampler against error accumulation for the long horizon rollout?

  • How applicable is this sampler to a standard pretrained diffusion model?

4 comments
Reviewer_7jKXOpenReview ↗

Follow up questions from Reviewer 7jKX

Thank you for your response and for conducting the additional experiments. I have also reread the revised paper, and my concerns regarding points 1–4 have been addressed. I have a few suggestions for the revision and two follow-up questions.

Suggestions for the revision

  • Could you highlight the revised text in a different color so that we can easily identify the changes?
  • The notation vθv_\theta in Equation (6) is unclear. Please introduce and define both vθv_\theta and qϕq_\phi in Section 3.2. Section 3.3 can then focus on the inference.

Follow-up questions

  • The training set requires ϵ0:K−1\epsilon_{0:K-1}. Since the method uses noise as a supervision signal, does this mean that you also need to record the Brownian noise term γdWt\gamma dW_t added to the Lorenz system? If so, this would significantly limit the method's applicability, because the noise is generally not directly observable in real-world settings.

  • I am still confused by the explanation in point 5: "Run the frozen model to generate a high-resolution image, record its initial noise and the noise used at each step, and downsample the image." Isn't this procedure already part of standard diffusion models? Could you clarify what you intend to illustrate by presenting the diffusion and Brownian smoothing equations side by side?

Thank you for rereading the revision and for confirming that your concerns regarding points 1–4 have been addressed. We will highlight the revised text in color and explicitly define both the conditional initial-state density, qϕq_\phi, and the observation-conditioned control in the driving-noise space, vθv_θ, in Section 3.2 before Eq. (6). Section 3.3 will then focus on inference.

1. Are Brownian noise realizations required from the physical system?

No. The Gaussian innovations are required only during offline training, where they are generated and recorded by the supplied simulator, not observed from the real physical system.

BS is trained on simulated trajectories and their corresponding simulated observations. Each simulation supplies the initial state, the Gaussian innovations used to advance the dynamics, and the resulting states and observations. For Lorenz–96, the recorded innovation is the standard Gaussian draw used in each discretized Brownian increment. These internally available quantities provide the supervision in Eq. (6).

At inference, BS takes a new observation record together with the supplied model parameters. It samples an initial state from the learned conditional distribution and generates fresh Gaussian innovations with learned mean shifts. Neither the true states nor the Brownian noise realization underlying the observed physical trajectory is required.

Thus, the requirement is access to the simulator's random inputs during training; it does not impose an additional noise-measurement requirement on the physical system. We will make this distinction explicit in Section 3.2.

2. What was the pretrained-diffusion example intended to illustrate, and why compare the equations?

The example illustrates how the same BS construction could condition an already trained stochastic generator using supervision supplied by the generator itself. The key distinction is between the noise predicted by the pretrained denoiser and the fresh random innovations drawn by its sampler.

In standard DDPM, the denoiser is trained to predict the noise used to corrupt data. During generation, the sampler also draws fresh Gaussian innovations at its stochastic steps; these are separate random variables from the denoiser's prediction (Ho et al., 2020, Algorithms 1–2). BS would learn how a new observation changes the distribution of these sampling inputs. Specifically, it would fit their conditional initial distribution and the conditional mean of each subsequent innovation given the current sampling state and the observation. The pretrained denoiser's usual noise prediction targets a different quantity.

This is why recording the generation process matters. Each frozen-generator run provides its initial input, intermediate states, stepwise innovations, and final image. Downsampling the image provides an observation paired with those inputs. Eq. (6) can then use these joint records to train the BS conditioning components directly, without reference posterior samples or evaluations of the observation likelihood. At deployment, a new low-resolution image would determine the conditional initial-input law and the stepwise mean shifts, while the original denoiser would continue to perform every generation step. Thus, the generated image supplies the observation, while the recorded random inputs supply the supervision for conditioning the generator.

The equations side by side explain both the mathematical relationship and how this adaptation would fit into it. They exhibit a common drift-plus-correction structure, with different learning targets: the perturbed-data marginal score in standard reverse diffusion, and the observation-conditioned simulator innovation mean in BS. In the pretrained model's application, the existing score would remain inside the supplied simulator's drift; BS would learn an additional correction for the new observation task. Under the paper's continuous-time assumptions, this correction has the classical Doob conditioning interpretation.

The practical motivation is to amortize observation conditioning while reusing the pretrained generator. Unlike likelihood-guided posterior sampling, the proposed construction would not require observation-likelihood gradients at each denoising step for every new record. Instead, the observation-dependent corrections would be learned offline and evaluated directly during sampling, without retraining the original denoiser.

The example therefore carries over the paper's central construction: learning a conditional law of simulator inputs from the simulator's own records, then generating an approximate conditional sample through the same simulator. It would require training the BS conditioning components for the observation task while retaining the pretrained denoiser. Empirical evaluation of this application is future work; our current experiments validate the construction for physical-system smoothing.

Reviewer_7jKXOpenReview ↗

Follow up questions from Reviewer 7jKX

No. The Gaussian innovations are required only during offline training, where they are generated and recorded by the supplied simulator, not observed from the real physical system.
BS is trained on simulated trajectories and their corresponding simulated observations. Each simulation supplies the initial state, the Gaussian innovations used to advance the dynamics

This is what I am saying. In real-world settings, we have an offline training dataset consisting of observations of the form

x~t=xt+σϵt,ϵt∼N(0,I),\tilde{x}_t = x_t + \sigma \epsilon_t, \quad \epsilon_t \sim \mathcal{N}(0, I),

where tt denotes the time index.

BS requires access to both xtx_t and ϵt\epsilon_t separately for training. However, in most real-world cases, we only have x~t\tilde{x}_t, and it is impossible to separate the noise from the underlying signal. Therefore, I think BS has significant limitations. For this reason, I will retain my score.

We agree that the clean states and noise realizations generally cannot be recovered separately from physical measurements. However, BS does not require such a decomposition of real-world data.

BS generates its own synthetic training set by sampling initial states and process innovations, running the dynamics, and simulating noisy observations. It does not require splitting physical measurements into unknown clean states and noise.

Moreover, for clarity, in your additive observation model, ϵt\epsilon_t represents measurement noise. By contrast, the innovations supervised in Eq. (6) are the random inputs to the Euler state transition in Eq. (2). They are recorded during simulation, not recovered by decomposing measured signals.

There is also direct evidence that this kind of approach can support real-data inference. DINGO (Dax et al., Physical Review Letters, 2021) trains on simulated gravitational-wave signals with randomly generated noise, whose statistics are informed by detector data, and analyzes eight real LIGO–Virgo events. It does not require the true signal and noise of each detected event to be separately observed as training labels. This example illustrates why simulator-generated supervision can remain useful even when only the combined noisy signal is measured experimentally. For BS, the additional supervision consists of random inputs already generated when running the supplied stochastic simulator. Recording these inputs requires access to the simulator's internals, not additional measurements of the physical system.

Reviewer_Q2c7OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a methodology to address the smoothing problem. That is, the problem of sampling from the conditional distribution of a stochastic dynamical system over a fixed time interval given noisy partial observations. In this article, the stochastic system is a stochastic differential equation with known drift and diffusion matrix. The proposed method consists in parametrizing admissible posterior distributions as the laws of a discretized diffusion process whose drift can be controlled and then to choose an optimal control optimizing the loss function (6). Once an optimal control has been chose, simulation from the posterior can be carried out using Euler's method. The main theoretical results are Theorem 1 and Proposition 1. In Theorem 1 the authors connect the loss function (6) with the KL divergence up to an error term. In Proposition (9) the quantity the error term under assumptions, showing in particular that it vanishes as the time-discretization gets narrower. Numerical studies complement the theoretical findings.

Critical strengths
  • The paper tackles a natural problem proposing an algorithm that is simple to implement and not computationally intensive when compared to competitors. Numerical experiments seem to indicate that the method performs well across different benchmarks.
Critical weaknesses
  • A major issue is that the authors do not explain why the optimal vkv_k in (6) is connected to the conditional law of XX given YY. There is no theoretical result in this direction and no intuition is provided. At equation (7) they provide a form of the optimal control. At the very least, I would expect to see a statement saying that simulating (2) with εk\varepsilon_k given by (4)(4) is connected to the posterior law

  • The connection between conditional distributions of diffusion processes and stochastic control in continuous-time is well known, starting at least from the paper by Mitter and Newton [1]. This body of literature is ignored here. Even if we work in discrete-time here, my feeling is that all the theoretical principles underlying the method proposed here are developed there, thus attenuating the novelty of the manuscript under consideration. I may be wrong, and I would be glad to receive explanations about the difference between this approach to the smoothing problem and the approach therein.

  • Ultimately, the proposed methodology rests on solving a stochastic control problem in discrete time via a gradient descent method. Though I am not an expert, I would be surprised if this problem had not been considered before. I would have expected to see a discussion around this topic in the article.

  • The theoretical findings of Theorem 1 and Proposition 1 do not seem very informative about the properties of the proposed algorithm. Again, they appear to me as discretized versions of well-known results about the KL between the laws of two diffusion processes in continuous time.

Another problem is that, despite having a decent understanding of the mathematical aspects of filtering and smoothing theory as well as stochastic differential equations, I find the exposition rather challenging.

  • I don't quite get what the maps Φ\Phi and equation (3) mean. Is Φ\Phi the strong solution map associated to the the law of the optimally controlled process? Is (3) an equivalent way of saying that posterior sampling is carried out through an Euler scheme?
  • At line 82, I find the following definition for the observation path Y∼gm(⋅∣Xtm)Y\sim g_m(\cdot|X_{t_m}). Is gmg_m a map? Is it a conditional distribution or a kernel? The following sentence does not help: "their dependence on the supplied parameter \lambda is suppressed throughout" does not help. What is λ\lambda? Where has it been supplied?

[1] S.K.Mitter and N.J. Newton

Questions for authors
  • Why optimizing (6) should give back the conditional distribution?
3 comments

1 I thank the authors for clarifying the role of vkv_k, which indeed acts as a control. I agree that Φ\Phi is not learned.

I also thank the authors for explaining the content of Proposition 1, which provides some form of theoretical guarantees in KL between the learned law and the correct conditional law. I still don't understand well what the constant rr is in equation (9). The assumption on the hessian of the future log-likelihood in Proposition 1 is very strong. In fact, no indication is given on how to bound this function. My strong feeling is that this assumption reads as a bounded Hessian condition for a Hamilton-Jacobi equation. It is well known that two-sided such bounds are challenging to obtain.

2 I thank the authors for their response. I agree that no posterior samples are needed. It also would appear that the author's methodology has not been considered before in the literature. However, I still confirm my previous impression that ultimately the proposed methodology is that of solving a discretized version of Mitter-Newton via gradient descent. It may still be that this is a relevant methodological contribution in the field, though I am not the best placed to assess this.

3 Ok understood.

Overall, I better appreciate the paper now, and I may consider raising my score.

We thank the reviewer for rereading the paper. We appreciate your thoughtful follow-up. We would like to clarify the remaining points.

1. Meaning of rr in Eq. (9).

The notation rr is defined in Section 2, immediately after Eq. (1), where we state that WW is an rr-dimensional Brownian motion and σ(t,x)∈Rd×r\sigma(t,x)\in\mathbb R^{d\times r}. Thus, rr is the dimension of the driving Brownian noise. For example, in our Kolmogorov experiment, d=4096d=4096 while r=32r=32. We will restate this definition near Eq. (9) for clarity.

2. Scope of the curvature assumption in Proposition 1.

We agree that obtaining uniform Hessian bounds can be challenging for general nonlinear dynamics. However, our assumption is more specific than a uniform bound on the full state-space Hessian.

More precisely, the curvature assumption concerns the discrete future-observation log-likelihood along the Brownian driving directions. The chain-rule identity in Eq. (35), makes this restriction explicit:

\qquad z=F_k(x,\varepsilon),\quad \sigma_k=\sigma(t_k,x).$$ where $\Psi_{k+1}$ is the discrete future-observation likelihood. Since the Euler transition is affine in each innovation, this gives $$\nabla_\varepsilon^2\log\Psi_{k+1}(F_k(x,\varepsilon))= \Delta_k\sigma_k^\top \nabla_x^2\log\Psi_{k+1}(F_k(x,\varepsilon)) \sigma_k.$$ Therefore, the assumption controls curvature only in directions accessible through the driving noise, rather than requiring a bound in every state-space direction. We nevertheless agree that this is a sufficient regularity condition, not an unconditional convergence guarantee for arbitrary nonlinear systems. Establishing such bounds in general can indeed be difficult. A concrete case where the condition can be verified is the scalar linear-Gaussian smoother: $$dX_t=dW_t,\qquad Y=X_T+\eta,\quad \eta\sim\mathcal N(0,\tau^2),$$ with $\tau>0$. The future-observation likelihood satisfies $$\partial_x^2\log h(t,x) =-\frac{1}{\tau^2+T-t}.$$ The discrete potential $\Psi_{k+1}$ has the corresponding quadratic form, so the curvature is uniformly bounded by $M=\tau^{-2}$. Appendix C.6 further shows that the resulting $O(|\Delta|)$ residual rate is sharp even in this simple setting. Importantly, the curvature assumption is needed for the explicit $O(|\Delta|)$ residual bound in Proposition 1, not for the exact finite-grid forward-KL decomposition in Theorem 1. The latter holds under its separately stated moment conditions. We will make this distinction and the scope of Proposition 1 more explicit in the revision.

3. Relationship to Mitter–Newton and the role of gradient descent.

We appreciate this question, as it highlights an important distinction between a variational characterization and a learning procedure.

We agree that the connection between smoothing and stochastic control is classical. Indeed, Propositions 2–3 explicitly establish the relationship between our optimal control and the Doob correction. However, sharing the same optimal posterior control does not imply that two methods solve the same optimization problem.

In Mitter and Newton's variational characterization (Proposition 2.1, Eq. (2.8)), for a fixed observation record yy, the posterior path law is characterized by minimizing a functional of the form

\mathrm{KL}(Q\Vert P_0) -\mathbb E_Q[\log L(y\mid X)],$$ where $P_0$ denotes the prior path law and $L(y\mid X)$ is the observation likelihood. Up to a constant independent of $Q$, this is the reverse KL $\mathrm{KL}(Q\Vert P_y)$, where $P_y$ is the exact posterior path law. **Our Eq. (6) is not obtained by discretizing this variational objective.** Instead, we optimize $$\min_{\phi,\theta} \mathbb E_{(Z,Y)\sim p} [-\log q_{\phi,\theta}(Z\mid Y)],$$ where $Z=(X_0,\varepsilon_{0:K-1})$ consists of the recorded random inputs of the supplied simulator. This is conditional maximum likelihood, equivalent up to an additive constant to $$\min_{\phi,\theta} \mathbb E_Y \left[ \mathrm{KL}\bigl( p(Z\mid Y)\Vert q_{\phi,\theta}(Z\mid Y) \bigr) \right].$$ These formulations differ in two important respects. First, they optimize different objectives over different representations. The classical variational formulation minimizes a reverse KL over candidate path laws for a fixed observation record, whereas BS minimizes an expected forward KL over conditional simulator-input laws. Although their unrestricted optima represent the same posterior, the corresponding projections generally differ under restricted model families. In particular, the forward-KL projection onto our fixed-covariance Gaussian innovation family selects the conditional innovation means, as established in Eq. (7). This property does not generally follow from the reverse-KL projection. Second, and more importantly, the BS objective admits direct supervision from simulations. Although its interpretation involves the true conditional law, evaluating Eq. (6) requires only joint samples $(Z,Y)$ generated by the supplied simulator. The Gaussian innovation loss reduces to supervised regression on recorded innovations. Thus, training requires neither evaluating the observation likelihood nor optimizing an expected control cost under a candidate controlled process. It also does not require differentiating through a controlled simulator rollout. Consequently, even an amortized implementation of the classical variational-control objective would not automatically yield our conditional maximum-likelihood objective or its regression targets via innovation. We fully acknowledge the shared theoretical characterization. Our contribution is not a new characterization of the optimal control, but a different statistical construction for learning an amortized posterior sampler from ordinary simulator runs, together with a finite-grid analysis of its approximation error. We hope this clarifies why BS should not be viewed simply as gradient descent applied to a discretized Mitter–Newton control problem. We appreciate the opportunity for this substantive technical discussion and will make the distinction clearer in the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Closing summary of the discussion

Brownian Smoothing (BS) learns how to steer the simulator for new measurements from ordinary forward runs alone. The noise that drove a simulated step is already a training label for the inverse problem. Its mean given the current state and later observations is the optimal Gaussian mean shift. This lets us learn the control without computing an optimal control or reference posterior for any training record. Once trained, the same simulator produces a full approximate posterior trajectory for new measurements in one rollout. On 320-site Lorenz–96, BS and the matched flow baseline are both amortized. BS generated 64 paths in 1.69 seconds versus 127.25 seconds, with lower posterior-mean RMSE (1.79 versus 2.06, Table 1).

Q2c7 asked whether Eq. (6) simply discretizes the Mitter–Newton objective. The two formulations agree on the unrestricted posterior. However, what matters for the algorithm is which control they select when the sampler can shift only Gaussian means. Mitter–Newton minimizes reverse KL over candidate path laws for a fixed record. Discretizing time preserves that KL direction. BS minimizes expected forward KL over conditional simulator-input laws across records. These projections can select different controls. The forward projection selects the conditional innovation mean, which we can learn by regression on ordinary simulator draws, without evaluating the observation likelihood or differentiating through a controlled rollout. Moreover, even if the Mitter-Newton method is also made into an amortization algorithm, the discretization and amortization steps still cannot derive the noise regression objective Eq. (6).

7jKX asked how BS can be trained if a real sensor records only a noisy signal. We agree that the clean signal and measurement noise generally cannot be recovered separately from that reading. The training set of BS is generated by the simulator itself, so in reality, only mixed signals can be measured, and this record can also be used for inference. The simulator first draws and records the process noise that drives a trajectory, then generates noisy sensor readings along that trajectory. The regression target in Eq. (6) is therefore known before the sensor reading is even generated. When a real record replaces the simulated readings at inference, it serves only as the observation on which the smoother conditions.

We thank the reviewers for rereading the paper and for their questions, and the Area Chair and Program Chairs for their consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.