acceptodds
Withdrawn from ICLR 2027

A covariate-conditioned Ornstein-Uhlenbeck bridge model for reconstructing probability field in sparse domain

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A natural strategy for inferring the latent state at an unobserved location in space, especially when target observations are sparse, is to use information from neighboring target observations and its auxiliary information. We consider this problem in a setting where target observations are sparse, while auxiliary covariate information is available throughout the space domain. In order to solve the problem, we propose a covariate-conditioned Ornstein-Uhlenbeck (COU) bridge model for reconstructing probability field at unobserved location. The proposed model decomposes the probability field into a deterministic component induced by global covariates and a stochastic residual modeled by an OU process. The local decorrelation structure of the residual is modulated by the covariates, allowing the model to reflect spatial heterogeneity and structural discontinuities. Given observations at two flanking locations, the resulting COU bridge provides probabilistic inference at interior locations. We first perform experiment on spatial data, and then for more generalization we evaluate the model on multiple spatio-temporal datasets. The results show that the proposed model is effective when sparse observations are complemented by informative covariates and the underlying dependence structure is consistent with the bridge formulation.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Ornstein-Uhlenbeck bridge, Covariate, Stochastic, Residual, Probability, Decomposition

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_icAvOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a covariate-conditioned Ornstein-Uhlenbeck (COU) bridge model for reconstructing probability field along a curve. The main idea is to model the latent field using a covariate-dependent OU process. The paper derives the formulas and an algorithm for fitting the probability field and interpolation at new locations. The paper further presents numerical results for the proposed method for synthetic data and real data.

Critical strengths

The paper proposes an algorithm that uses Laplace approximation to evaluate the intractable likelihood for the proposed probability field.

Critical weaknesses
  1. The paper claims that the COU process is new for modeling the latent probability field. However, this contribution is rather incremental, for several reasons.

First, the paper has completely missed out a vast literature on Gaussian process regression (kriging) in spatial statistics that studies the modeling and theoretical properties of stationary and nonstationary Gaussian random fields. The stationary OU process has been studied extensively, with its properties in parameter estimation and prediction fully understood. In addition, it is well known that the stationary OU process is essentially the Gaussian process with exponential covariance function, which is also the Mat'ern covariance function with smoothness parameter ν=1/2\nu=1/2. The Mat'ern covariance functions have been extensively studied in spatial statistics (Stein 1999, Porcu et al. 2024). The Markov property of OU process in the spatial context is already very well understood (Ying 1991, Lindgren et al. 2011). The construction of nonstationary Gaussian processes has been studied in Paciorek and Schervish (2006), which has also been cited in the current paper.

In this regard, the formula of OU covariance function in (3) is elementary and not new, since it is a basic combination of the stationary OU process with exponential kernel exp⁡(−∣x−x′∣)\exp(-|x-x'|) (the Mat'ern kernel with ν=1/2\nu=1/2) with the nonstationary construction in Paciorek and Schervish (2006). Equation (4) also follows from the standard formula of kriging.

Second, even if we restrict our attention to only the OU process on a curve, such a process has also been thoroughly studied in spatial statistics. For example, the recent work Wen et al. (2021) studies the parameter estimation in general Mat'ern covariance functions on a curve. Although the current paper studies a slightly more involved probability field, the construction of latent process has very limited difference. Furthermore, the current paper has not investigated the theory of parameter estimation or prediction, which has been studied in Wen et al. (2021).

  1. The paper is inconsistent in its exposition, as the primary model under consideration is not consistently defined. At the beginning of Section 2, it seems that the primary goal is to model the YY process. However, at the bottom of page 2, the focus is changed to modeling the ZZ process. The YY process is never mentioned again until Section 4.2. The remaining Section 2 seems to study the modeling of a probability random field pp which is linked to ZZ through the logistic link. The likelihood approximation in Section 3.3 also works only for this logistic model for pp. For the experimental results in Section 4, the simulation in Section 4.1 is for the pp process, while the simulation in Section 4.2 instead uses a Gaussian process YY, not the probability process pp. The mismatch and inconsistency severely undermine the motivation and narrative.

  2. The reason for the choice of two endpoints is not sufficiently explained. To start with, in the standard Gaussian process (GP) regression, it is not common to assume that we have observations at two endpoints. Even if we do, the GP posterior should condition on all the observed data, not just the two endpoints. It is unclear why the likelihood in Section 3.3 only involves the two endpoints.

  3. The numerical results in Section 4 do not convincingly demonstrate the advantage of COU. For example, for the experiments in Section 4.1, the value of nin_i is not given. In addition, the data in Section 4.1 are only generated from COU, which is an unfair comparison and is expected to give COU an advantage. Even this is the case, Table 2 shows that the performance improvement from COU compared to the stationary kriging is very limited under almost all the metrics. The numerical experiments in Section 4.2 are for a totally different Gaussian model. Since standard errors are not reported, it remains unclear whether the marginal improvement from COU is statistically significant.

References

Lindgren, F., Rue, H., and Lindstr"om, J. (2011). An explicit link between Gaussian fields and Gaussian Markov random fields: the stochastic partial differential equation approach. {\it Journal of the Royal Statistical Society: Series B} 73: 423-498.

Paciorek, C. J. and Schervish, M. J. (2006). Spatial modelling using a new class of nonstationary covariance functions. {\it Environmetrics} 17: 483-506.

Porcu, E., Bevilacqua, M., Schaback, R., and Oates, C. J. (2024). The Mat'ern model: A journey through statistics, numerical analysis and machine learning. {\it Statistical Science} 39: 469-492.

Stein, M. L. (1999). {\it Interpolation of Spatial Data: Some Theory for Kriging.} Springer Series in Statistics. Springer, New York.

Wen, J., Sun, S., and Loh, W.-L. (2021). Smoothness estimation of nonstationary Gaussian random fields from irregularly spaced data observed along a curve, {\it Electronic Journal of Statistics} 15: 6071-6150.

Ying, Z. (1991). Asymptotic properties of a maximum likelihood estimator with data from a Gaussian process. {\it Journal of Multivariate Analysis} 36: 280-296.

Questions for authors

None.

Reviewer_cDFSOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a bridge based solution, COU, that tackles spatial inference prediction problem given globally available spatial covariates and sparse observations. The method is validated on a 1-D synthetic binomial probability field, where it has lower RMSE and CRPS than a covariate-mean model, linear interpolation, a stationary GP and its own stationary variant (Table 2), and on held-out sensor reconstruction for the METR-LA and PEMS-BAY traffic datasets, where it has the lowest RMSE among these baselines (Table 3).

Critical strengths
  1. The inference recipe is short and self-contained. Prediction is cheap: a 2-D concave Newton problem plus a 1-D quadrature per target point.

  2. The conclusion states several limitations openly (sparse data, uninformative covariates, linear mean, no temporal model).

Critical weaknesses

W1. Notation is unclear, and the mathematical framework is very hard to follow.

  • It is never said that uu is only the coordinate along the path, i.e. Z(u)Z(u), x(u)x(u), m(u)m(u) mean Z(γ(u))Z(\gamma(u)), x(γ(u))x(\gamma(u)), m(γ(u))m(\gamma(u)). The paper then switches from uu (Sec. 2–3) to ss (Sec. 4). I first read uu as a second index next to ss.
  • The random-variable definitions are sloppy:
    • Y:S×Ω→DY:S\times\Omega\to D and YA,YB,YMY_A, Y_B, Y_M are defined but never used again.
    • The observation model linking YY to pp or ZZ (e.g. Binomial) is never stated in Sec. 2–3; it only appears as Ki∼Binomial(ni,p(si))K_i\sim\mathrm{Binomial}(n_i,p(s_i)) in Sec. 4.1.
    • DA,DBD_A, D_B are never defined, and DD is also the codomain of YY.
  • LA(r)=p(DA∣RA=r)L_A(r)=p(D_A\mid R_A=r) hides its dependence on the mean m(0)m(0). It reads as if rr were a quantity shared along the path.
  • Overloaded symbols:
    • KK is the covariance KABK_{AB}, the binomial count KiK_i, and the number of quadrature nodes.
    • γ\gamma is both the path and the coefficient vector of log⁡ν2\log\nu^2.
    • c\mathbf c is both KAB−1kMK_{AB}^{-1}k_M and the covariates c(s)c(s) in Algorithm 1.
  • Algorithm 1 points to the wrong section and equations.
  • τ(u)=∫0uθ(s) ds\tau(u)=\int_0^u\theta(s)\,ds is integrated in path coordinate, while θ\theta is specified per unit of physical distance (Table 4: ℓ0=5\ell_0=5). As written, the correlation between AA and BB does not depend on their distance. A factor ∣γ′(u)∣|\gamma'(u)| (path length) seems to be missing.
  • Sec. 4.2 reuses Yt(si)=mt(si)+Rt(si)Y_t(s_i)=m_t(s_i)+R_t(s_i) for a Gaussian quantity, so YY now plays the role of ZZ in Sec. 2.

W2. The bridge construction brings little; the proposal is incremental.

  • The "bridge" (Eq. 4) is ordinary two-point Gaussian conditioning, i.e. kriging of a latent GP from two sites.
  • A logit link, binomial likelihood, Laplace approximation and a 1-D predictive integral is the standard latent-GP recipe.
  • Covariate-dependent range and variance via log-linear models already exist and are not cited:
    • Ingebrigtsen, Lindgren & Steinsland (2014);
    • Bolin, Riera-Segura & Simas (2025), on metric graphs, which covers 1-D paths and road networks;
  • Conditioning only on the two flanking observations throws away information. The OU process is Markov in the latent RR, but RA,RBR_A, R_B are only seen through noisy binomial counts, so p(RA,RB∣all data)≠p(RA,RB∣DA,DB)p(R_A,R_B\mid\text{all data})\neq p(R_A,R_B\mid D_A,D_B). A Kalman smoother uses all observations at O(n)O(n) cost (Nickisch, Solin & Grigorevskiy, 2018), so the bridge has no computational advantage either.
  • The positioning against diffusion bridges is superficial. I2SB, DDBM, GOUB and stochastic interpolants are bridges in diffusion time between two distributions; here the bridge is in space between two points. None of them is compared against.

W3. The scope is limited to 1-D paths, with no extrapolation.

  • γ\gamma is a curve inside SS, so the model only predicts at points lying on a chosen path between two observed sites. In every experiment ss is effectively 1-D: s∈[0,200]s\in[0,200] in the synthetic data, and sensors "ordered along a corridor" in the traffic data.
  • For a 2-D domain or a branching graph, there is no rule for choosing the path or the pair of sites.
  • Each path defines its own process starting from R(0)∼N(0,V(0))R(0)\sim\mathcal N(0,V(0)). Different paths through the same location can give different predictions, so it is unclear whether a single random field on SS is defined, as Sec. 2.2 promises.
  • The method needs an observed site on both sides of the target, so it cannot predict beyond the first or last observed site. Ordinary kriging handles this naturally by reverting to the mean. The one-sided OU transition is written in Sec. 3.2 but never used for prediction.

W4. Limited Evaluation

  • The synthetic evaluation is for illustration only, and key details are missing.
  • The real-data evaluation is thin and compares with no other baselines. The traffic model is Gaussian, Yt(sq)∣Dt∼N(μtq,vtq)Y_t(s_q)\mid D_t\sim\mathcal N(\mu_{tq},v_{tq}), with no logit link, binomial likelihood, Laplace step or quadrature. The main probability-field method is never tested on real data. There are only two datasets, both traffic, with time ignored ("each timestamp as a separate spatial reconstruction problem"). The baselines are only the covariate mean, a stationary GP and the paper's own stationary variant.
Questions for authors
  1. Notation and model definition.

    • Can you confirm that uu is only the coordinate along the path, so that Z(u)Z(u), x(u)x(u) and m(u)m(u) mean Z(γ(u))Z(\gamma(u)), x(γ(u))x(\gamma(u)) and m(γ(u))m(\gamma(u)), with no second index? Please use one coordinate consistently in Sec. 2–4.
    • What is the observation model of YY given the latent field? Please state it explicitly, e.g. k(x)∣Z∼Binomial(n(x),σ(Z(x)))k(x)\mid Z\sim\mathrm{Binomial}(n(x),\sigma(Z(x))), conditionally independent across sites. Please define DA,DBD_A, D_B (e.g. DA=(kA,nA)D_A=(k_A,n_A)) and the codomain DD, and either use YA,YB,YMY_A, Y_B, Y_M or remove them.
    • Could you write out the full probabilistic model with coherent notations?
  2. What is the role of the bridge here?

    • Is there anything in the construction beyond two-point Gaussian conditioning (kriging) of a latent GP with a binomial likelihood?
    • Why condition only on the two flanking observations, when the endpoints are seen through noisy counts? How does the method compare with the full latent-GP posterior using all observations, e.g. Kalman smoothing (Nickisch et al., 2018)?
Reviewer_PJy7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a covariate-conditioned Ornstein–Uhlenbeck (COU) bridge model for reconstructing sparsely observed spatial fields. The model decomposes the latent field into a covariate-driven mean component and a nonstationary OU residual whose local decorrelation rate and variance depend on covariates. The resulting bridge distribution enables probabilistic interpolation between neighboring observations, allowing spatial heterogeneity and structural discontinuities to be reflected. The effectiveness of the proposed method is demonstrated on synthetic data and two traffic datasets.

Critical strengths
  • The proposed model is conceptually simple and interpretable. The decomposition into a deterministic covariate-induced mean and a stochastic residual provides a clear interpretation of the two sources of spatial variation. In particular, the parameters that control the local decorrelation rate and residual amplitude have intuitive interpretations.
  • The covariate-dependent effective distance via the integrated decorrelation rate is an appealing modeling approach.
  • The Gaussian Markov structure of the OU process yields a simple bridge distribution between two endpoints, leading to a tractable and computationally attractive inference.
Critical weaknesses
  • The practical interpretation of the proposed probability-field formulation is unclear. In the title and abstract, the authors insist that the purpose is to reconstruct a probability field, which is more clearly defined in the main text as a Binomial success probability field. This implicitly assumes multiple exchangeable trials sharing exactly the same pointwise success probability. It is not clear what realistic spatial observation mechanism this is intended to represent. In many applications, observations are either associated with individual locations, aggregated over spatial regions, or arise as spatial events, in which case an areal model or a spatial intensity model may be more natural. The manuscript should provide a concrete real-world interpretation of the assumed pointwise Binomial sampling mechanism.
  • A natural alternative to the proposal would be to model the latent field directly as a function of the available covariates, e.g. Z(u)=g(x(u))Z(u) = g(x(u)), or more generally Z(u)=g(u,x(u))Z(u) = g(u, x(u)), with a GP prior or an RKHS/kernel estimator on g(⋅)g(\cdot). The paper does not convincingly demonstrate what is gained over such established covariate-aware or nonstationary GP/kernel approaches beyond locality, interpretability, and computational convenience. A comparison with them would be essential for determining whether the improvements arise from the bridge structure itself or simply from using a covariate-aware smoothing approaches.
  • The synthetic experiments use a latent Binomial success-probability field, whereas the METR-LA and PEMS-BAY experiments reconstruct continuous-valued traffic measurements under a Gaussian field. This means that the central probability-field setting is not directly validated on real data. If probability-field estimation is the main claim of the paper, the authors should provide real-data experiments under the same setting. Alternatively, if the essential contribution is the reconstruction of a more general latent field, the title, abstract, and claims throughout the manuscript should be revised accordingly.
  • The claimed general spatial applicability is broader than the actual pathwise formulation. The method is defined along a path γ:[0,1]→S\gamma:[0,1] \to \mathcal{S} connecting two observed endpoints and exploits the one-dimensional Markov structure of the OU process. It is not clear how the method generalizes to genuinely multidimensional spatial domains with multiple relevant paths and more than two informative neighboring observations. This limitation is important given the method's broad framing as a general approach to sparse spatial reconstruction.
  • While the model parameterization for the mean, decorrelation rate, and residual amplitude is given, the main text and the prediction algorithm focus primarily on inference conditional on these parameters. The manuscript would benefit from a more explicit description of the training objective and estimation procedure for (β,α,γ)(\beta, \alpha, \gamma), including optimization details.
Questions for authors

N/A

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

Related papers

Discussion (0)

Sign in to comment.