acceptodds
Under review as a conference paper at ICLR 2027

Local Calibration Is Not Enough: Coherent Spatial Dependence for Query-Level Uncertainty

Abstract

Probabilistic models for spatial fields can appear well calibrated locally while giving unreliable uncertainty for downstream spatial queries. We study what is lost when local predictive uncertainty is preserved but cross-region dependence is discarded. For linear queries, we derive the exact factorization , separating effective spatial participation from cross-region coherence. On Western Australian sea-surface temperature, removing cross-patch sample correspondence reduces nominal- domain-mean coverage from to and underestimates its variance by a factor of . A leading coherent mode explains only of total field variance but of domain-mean variance; removing it changes from to . The effect is also present in stochastic PDE reference distributions and is reproduced by two distinct generative architectures. These results show that query-level uncertainty depends critically on coherent spatial structure, motivating scalable uncertainty representations that preserve local uncertainty alongside coherent large-scale structure.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
spatial uncertainty quantification, spatial dependence, functional uncertainty, probabilistic generative models, ensemble calibration, coherent covariance, stochastic PDEs

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

This paper investigates the limitations of local uncertainty calibration for linear spatial queries. The authors derive a participation-coherence factorisation to explain how spatial dependence affects query-level uncertainty. They demonstrate on several experiments like sea-surface temperature prediction and stochastic PDE settings, and show that ignoring cross-region dependence can substantially distort uncertainty calibration. The authors further analyse coherent covariance modes and explore low-rank corrections to recover query-level uncertainty.

Critical strengths
  • The problem is well motivated, and the paper is clearly written and easy to follow.
  • Experiments are extensive, checked generative models and PDE numerical solvers to demonstrate the limitations they claim. The range of baselines is sufficient.
Critical weaknesses
  • From my understanding the PatchPODiff already generates joint samples and preserves cross patch dependence. So it is a bit unclear how representative this case is practically, perhaps one more experiment on forecasting model actually lose the spatial dependence will be good. And the low-rank correction proposed relies on coherent modes estimated from historical joint samples, which may not be available for case that if the model being used does not generate joint samples. I would appreciate if authors can clarify this point as well.
  • Others see questions.
Questions for authors
  • How sensitive is R to the number of blocks? For example, would grouping the 300 patches into larger blocks change the SST results much?
  • Could the coherent modes for the low-rank correction be estimated without joint samples?
  • Section 4.3 evaluates 500 random spatial queries, but the low-rank recovery experiments in section 6.2 mainly focus on five selected linear queries. Could the authors also evaluate the rank-3 correction on these random queries to better understand how well the recovery generalises?
2 comments
AuthorsOpenReview ↗

Response to Reviewer oFzU: Partition Sensitivity, Residual Modes and Random Queries

We appreciate the reviewer's constructive assessment and questions about practical relevance, partition sensitivity and mode estimation without historical joint samples. The revised manuscript includes additional results on all three points.

W1. Why the PatchPODiff intervention is informative. PatchPODiff generates joint posterior samples; it is not an independently tiled generator. This makes its ensemble a useful controlled reference. We remove only the correspondence of sample identities across patches, retaining predictive means, the trained model and every within-patch empirical distribution. The intervention therefore isolates dependence loss when local uncertainty is stored or sampled without global coupling. We now vary the retained block size and evaluate a compact correction rather than leaving it as future work.

Q1. Does the result depend on 300 patches? We repeat the intervention on finer and coarser partitions of the same held-out 2011 SST ensemble (Table 8, Appendix D.5):

PartitionBlocksDomain RRDomain 90% coverage
Individual ocean pixelsAll valid pixels98791.38%
Native 32×3232\times32 patches30046.6520.4%
Merge 2×32\times3 patches5018.732.0%
Merge 4×34\times3 patches2511.341.3%
Merge 5×55\times5 patches126.854.0%
Merge 10×510\times5 patches63.965.6%
Two regional halves21.7481.8%

Larger blocks preserve more of the original dependence, systematically reducing domain-mean distortion. Query geometry also changes its sign: for the N–S contrast, RR falls from 8.73 with 300 patches to 0.48 with two blocks, where independence overestimates variance. Thus the effect is not an artifact of the native partition.

Q2. Can the modes be estimated without historical joint predictive samples? Yes. We form observed-minus-predictive-mean residual fields from 257 days in 2010, project them into the fixed patchwise POD coordinates and extract empirical orthogonal functions. We freeze this basis before the 2011 evaluation. Its leading three directions have absolute cosines 0.987/0.974/0.963 with modes estimated from the 2010 predictive ensemble, eigenvalue ratios 0.98–1.02 and mean squared principal-angle cosine 0.956. The maximum change in canonical-query coverage is 0.6 percentage points. The residual-derived correction has a 2.1 pp mean absolute coverage gap from the retained joint ensemble, compared with 2.3 pp for posterior-derived modes (Appendix F.7). This empirical agreement need not hold for every system, but shows that historical observations and predictive means can suffice here.

Q3. Does rank three recover the 500 random queries? We use the same 250 non-negative aggregates and 250 zero-mean contrasts from Section 4.3:

ResultPositiveContrasts
Rank-1 median recovery97.2%13.1%
Rank-3 median recovery99.4%98.7%
Rank-5 median recovery100.1%99.8%
Rank-3 recovery IQR96.1–103.8%92.4–106.9%
Within 80–120% recovery99.6%97.6%
Rank-3 nominal-90% coverage88.8%91.7%
Joint coverage (original benchmark)88.9%91.9%

Rank one largely recovers positive averages but only 13.1% of contrast variance; adding Modes 2–3 restores the sign-changing structure. This demonstrates recovery beyond the five canonical queries. The results appear in Section 6.2 and Appendix D.3 of the revised manuscript; partition sensitivity is in D.5 and residual-based estimation in F.7.

I thank the authors for the detailed response and additional experiments. I will maintain my rating.

Reviewer_sRxnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies what happens to the uncertainty of downstream spatial quantities when a predictive ensemble keeps the distribution within each spatial block but loses the correspondence of samples across blocks ("block independence"). The paper does not propose a new method. The block-independent ensemble is described as a controlled intervention, and the low-rank correction is presented as motivation for future scalable representations.

Critical strengths
  • The intervention is clean: the model, the query and every block's marginal are kept fixed, and only cross-block sample correspondence is removed. This isolates the information that pointwise calibration cannot see.
  • RwR_w, NeffN_\text{eff} and κ\kappa can be computed directly from samples (or latent samples, Eq. 7) without forming a dense covariance, so the check is cheap for any ensemble and any linear query.
Critical weaknesses

W1 Presentation and clarity.

  • Many terms are used before (or without) being defined: "query", "query-aligned", "mode", "coherent mode", SST, PatchPODiff, POD, ROMS, DDPM, MACE. PatchPODiff is named in the introduction but never introduced, and "mode" is only defined in §5.1 as an eigenvector of the latent predictive covariance.
  • The conditioning variable yy in p(u∣y)p(u \mid y) for the SST experiment is never stated. Since the backbone is a super-resolution model, this matters for interpreting why a domain-wide mode carries so much uncertainty.
  • The summary of contributions is hard to map to the paper. It is not clear which part is the method, which is the analysis, and which is the empirical finding. The same message ("participation bounds, coherence determines") is repeated in the abstract, §1, Fig. 1, §2.3, §4.2 and §7, while key definitions (query weights, how intervals are built, how daily results are aggregated) sit only in the appendix.

W2 Novelty, positioning and significance.

The paper’s central finding that removing cross-block dependence while preserving block marginals distorts the variance of spatial aggregates follows directly from Var⁡(w⊤u)=w⊤Σw\operatorname{Var}(w^\top u) = w^\top\Sigma w and is acknowledged by the authors as well recognized (§1). The participation–coherence factorization is algebraically the classical design effect 1+(Neff−1)ρˉ1+(N_\text{eff}-1)\bar\rho (Kish, 1965, which the paper cites), and the mode analysis restates that a query’s variance is dominated by the eigenmodes aligned with its weights. The block-independent ensemble is constructed by the authors from joint models and does not correspond to any method evaluated or cited as used in practice, so it is unclear what decision or method the analysis informs. The most actionable element, the rank-rr correction (Eq. 5), is presented as motivation rather than evaluated as a method under the same 2010-only protocol as the baselines. I would encourage the authors to either (a) apply the diagnostic to a scalable approximation that practitioners actually use, or (b) develop and fairly evaluate the local-plus-low-rank representation.

Questions for authors

What decision or method does the intervention inform? Is there a method used in practice that generates or stores block-independent uncertainty, and can you apply RwR_w, NeffN_\text{eff}, κ\kappa to it (e.g., tiled or overlapping-patch diffusion, per-pixel quantile products, models trained on marginals such as Alet et al., 2025)?

2 comments
AuthorsOpenReview ↗

Response to Reviewer sRxn: Evaluated Low-Rank Correction and Practical Relevance

We appreciate the emphasis on a methodological contribution beyond diagnosing dependence loss. Following option (b), we evaluate local-plus-low-rank covariance against alternative calibrations on held-out SST queries.

Novelty beyond the design effect. We do not claim novelty for Var(w⊤u)=w⊤Σw\mathrm{Var}(w^\top u)=w^\top\Sigma w or classical design-effect theory. The exact factorization Rw=Neff(w)κ(w)R_w=N_{\rm eff}(w)\kappa(w) diagnoses query-specific participation and coherence. Our contribution is identifying the coherent dependence that controls error and testing a compact representation that restores it.

Three findings matter. (i) With 300 independent patches, domain R=46.65R=46.65 but N–S/E–W contrast ratios are 8.73/8.16. Pixel independence gives domain R=9879R=9879; with two blocks the N–S contrast instead has R=0.48R=0.48, so independence overestimates variance. (ii) Mode 1 explains only 9.99% of field variance but 99.70% of domain-mean variance; removing it changes domain RR from 46.65 to 0.14. Modes 2–3 dominate contrasts; the first three modes explain just 14.63% of field variance. (iii) Frozen 2010 modes recover near-joint uncertainty on 2011, including unseen queries.

Practical setting. PatchPODiff samples patches jointly. Its ensemble provides a controlled reference: shuffling member identities across patches preserves the model, means and block marginals, removing only cross-patch dependence. This models locally stored ensembles, tile distributions and marginal products without shared member identities. Domain nominal-90% coverage falls from 89.5% joint to 20.4% (300 independent patches) and 1.38% (independent pixels). For 50/25/12/6/2 blocks, domain RR falls to 18.7/11.3/6.8/3.9/1.74. This quantifies the dependence needed for downstream queries.

Evaluated correction and baselines. We preserve each 2011 within-block covariance and restore only cross-block contributions of rr coherent modes estimated on 2010, subtracting their within-block terms to avoid double counting. The rank-three covariance satisfies the PSD condition on all 108,900 SST day–block pairs; a shrinkage safeguard guarantees PSD for other inputs while keeping block covariances unchanged (Appendix B.6). Table 15 compares five canonical 2011 queries; lower gap and score are better:

MethodNew-query reuseGap (pp)Score/joint
Independent blocks—54.01.620
Stationary MatérnYes25.81.340
Historical ECC-styleYes16.71.165
Global scalingYes13.01.245
Matérn + commonYes12.31.170
Query-specific scalingNo11.71.095
Functional conformalNo8.71.085
Rank-3 posterior modesYes2.31.018
Rank-3 residual modesYes2.11.027

Rank-three posterior coverage for domain/north/south/N–S/E–W is 84.0/90.6/84.8/86.8/95.6%, against joint 89.5/90.4/88.7/88.4/95.6%. The E–W-fitted scale gives only 41.6% domain coverage, while the domain scale gives 99.7% E–W coverage. A single common mode also fails to capture sign-changing contrasts.

Unseen queries. Across 250 smooth positive aggregates and 250 zero-mean contrasts, rank-three median variance recovery is 99.4% / 98.7%. 99.6% / 97.6% of queries recover 80–120% of joint variance; coverage is 88.8% / 91.7% versus original joint benchmarks 88.9% / 91.9%. Rank one recovers 97.2% for positives but only 13.1% for contrasts; Modes 2–3 restore contrast dependence (Section 6.2, Appendix D.3).

No historical joint ensemble required. We estimate modes from 2010 observed-minus-predictive-mean residuals in fixed POD coordinates. Their absolute directional cosines with predictive modes are .987/.974/.963, eigenvalue ratios .98–1.02, and mean squared principal-angle cosine .956. Canonical coverage changes by at most 0.6 pp. This agreement is specific to SST, not universal.

Cost, generality and clarity. Three modes plus local covariance use 0.626 MB, compared with 3.48 MB for 100 paired latent samples (5.6× smaller, excluding shared means), a storage rather than sampling-speed gain. Independent conditional PDE references show the same phenomenon: advection–diffusion R=4.27R=4.27–10.5810.58 versus PatchPODiff 4.11–10.94; Darcy R=2.06R=2.06–5.085.08 versus DLL (11.1% median relative error). Removing DLL correspondence drops coverage 95.1% to 78.8%. The correction targets linear-query covariance, not the full joint law; nonlinear exceedance-area coverage falls from 87.1% to 58.4% after shuffling. The revision defines queries and coherent modes in the Introduction and Table 1, clarifies SST conditioning and intervals in Section 3, and evaluates the correction in Section 6.2.

Thanks for the additional illustration of a methodological framing building on the analysis. My concerns regarding the clarity and scope of the paper remain, and it is not clear how this can be addressed with a revised draft without another round of review.

Reviewer_eHxEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies uncertainty for downstream spatial queries when cross-region dependence is removed while each block’s empirical predictive distribution is preserved. For linear queries, it derives an exact factorization of the full-to-block-independent variance ratio into effective participation and coherence. SST experiments show large, query-dependent variance distortion and undercoverage after dependence removal. Frozen-mode analyses identify large-scale structures that contribute modest field variance but dominate particular query variances. A correction using a few modes estimated on 2010 substantially recovers uncertainty on 2011. Stochastic advection–diffusion and Darcy experiments provide additional evidence using conditional PDE reference ensembles and two generative architectures.

Critical strengths
  1. A clear intervention with substantial empirical effects: The intervention preserves within-block samples and changes only their cross-block correspondence, providing a clear way to isolate dependence information. In the SST domain-mean example, nominal-90% coverage falls from 89.5% to 20.4%, with a full-to-block variance ratio of 46.65. The moving-block bootstrap, season-matched evaluation, and ensemble-size checks support the robustness of this effect. These results make a consequential failure mode tangible rather than merely restating that covariance matters.

  2. A useful explanation of query dependence: The comparison between positive averages and sign-changing contrasts, together with mode-removal interventions, explains why broad spatial support alone does not determine variance distortion. The leading SST mode accounts for 9.99% of field variance but 99.7% of domain-mean variance. The PDE reference experiments and the second architecture strengthen the evidence that dependence loss is not solely an artifact of one learned model.

Critical weaknesses
  1. I find the paper to be very hard to follow. Many of the symbol definitions are introduced in the Appendix, instead of the main paper. For example, in Line 74, the author does not define F_w, u, and w. I have to read them multiple times to understand this. In Line 87, the N_{eff} and k are not defined. In Line 151, the A_c is not defined. I have read the very long appendix to understand the paper. I feel like this paper require significant revision before it can be published.

  2. Key experimental and evaluation details remain unclear: Appendices A.2–A.3 report grids, dataset sizes, and ensemble sizes, but do not explicitly specify the stochastic PDEs, random inputs, conditioning variables, or the procedure for obtaining repeated conditional reference solutions. These details are important for interpreting what uncertainty the reference and learned distributions represent. For SST, Appendix A.1 states that nominal-90% intervals are constructed from daily predictive means and variances, but does not give the interval formula or distributional assumption. Appendix F also leaves the scale-fitting objectives, conformal score and quantile, and block-level ECC template construction insufficiently specified. These omissions limit reproducibility and make the baseline comparisons harder to assess. The concern is that essential details remain unclear after consulting the appendix, rather than that supplementary results are placed there.

  3. The frozen-mode construction needs a more precise statistical account: The paper estimates modes using 2010 and evaluates them on 2011, while computing the test variance quantities separately for each day. It should explain how the training covariance is pooled and centered across days. Pooling samples without centering within each day could mix variation in conditional means with within-condition predictive uncertainty. The manuscript does not provide enough detail to determine which construction was used. Equation (5) combines current block variances with cross-block contributions from frozen modes. This does not automatically guarantee a positive semidefinite covariance for arbitrary new block covariances. The authors should state sufficient assumptions or describe safeguards. This is a request to establish the correction’s validity as a reusable covariance representation; it does not show that the reported five-query variances are negative or incorrect.

Questions for authors
  1. Please specify the conditioning information and resampled random quantities used to generate the conditional reference solutions in each PDE experiment.

  2. Appendices A.1 and F describe the evaluation and calibration procedures, but please provide the exact SST interval formula, scale-fitting objectives, conformal score and quantile, and ECC template construction.

  3. Given that the modes are estimated on 2010 and frozen, please clarify how the covariance is centered and pooled across days, whether eigenvalues are also frozen, and which period and aggregation procedure are used to compute Table 2’s variance fractions.

  4. Please clarify the conditions or safeguards ensuring that Equation (5) defines a positive semidefinite covariance, and assess the rank-three correction on the existing random-query collection.

1 comment
AuthorsOpenReview ↗

Response to Reviewer eHxE: Notation, PDE Specifications, Frozen Modes and PSD Validity

We thank the reviewer for the technical questions. The revision clarifies notation, conditional PDE references, calibration and frozen modes, and adds recovery on 500 random queries.

W1. Notation and presentation. The Introduction now defines u∈Rdu\in\mathbb R^d, w∈Rdw\in\mathbb R^d and Fw(u)=w⊤uF_w(u)=w^\top u on first use. Table 1 collects NeffN_{\rm eff}, κ\kappa, RwR_w and the latent notation. Section 2.2 defines u=b+Dau=b+Da, q=D⊤wq=D^\top w and centered samples Ac∈RM×PKA_c\in\mathbb R^{M\times PK}; for SST it is 100×8700100\times8700. Coherent modes are introduced in the Introduction.

W2/Q1. Conditional PDE references. For advection–diffusion, we solve ∂tu+v⋅∇u=ηΔu\partial_tu+v\cdot\nabla u=\eta\Delta u on a periodic 128×128128\times128 domain with random smooth Gaussian-blob initial fields, at Pe∈{20,80,320}\mathrm{Pe}\in\{20,80,320\}. Each conditional case fixes the coarse observation and PDE regime; repeated solves resample unresolved fine-scale variability. We use 16×1616\times16 patches, 6000/120/240 train/calibration/test cases and 100 model samples. At Pe=80\mathrm{Pe}=80, east–west RR is 10.58 for the numerical reference versus 10.94 for PatchPODiff; an independent-patch reference gives mean R=0.992R=0.992.

For Darcy, −∇⋅(a∇u)=f-\nabla\cdot(a\nabla u)=f on (0,1)2(0,1)^2 with zero Dirichlet boundary conditions. The fixed conditioning field aa is thresholded Gaussian permeability (values 12 and 3). The resampled forcing is f=λσlneGln+(1−λ)σgpGgpf=\lambda\sigma_{\rm ln}e^{G_{\rm ln}}+(1-\lambda)\sigma_{\rm gp}G_{\rm gp}, with λ=0.1\lambda=0.1, (σln,ℓln)=(10,0.2)(\sigma_{\rm ln},\ell_{\rm ln})=(10,0.2) and (σgp,ℓgp)=(10,0.5)(\sigma_{\rm gp},\ell_{\rm gp})=(10,0.5). For each of 32 test permeabilities, 64 solves resample the forcing; DLL generates 100 samples. Reference and DLL use the same 64×6464\times64 grid, 4×44\times4 blocks and queries. These are empirical conditional PDE references, not analytic oracles.

W2/Q2. SST intervals and baselines. For each query and day, all covariance representations use nominal-90% Gaussian intervals μw,t±1.6449Vw,t\mu_{w,t}\pm1.6449\sqrt{V_{w,t}}, evaluated on 363 days in 2011. On 257 days in 2010, variance multipliers are the smallest values attaining 90% empirical coverage for the block-independent intervals: global 3.2573.257 (pooled queries); query-specific 4.44,3.52,3.15,2.39,2.174.44,3.52,3.15,2.39,2.17 (domain, north, south, N–S, E–W). The scalar split-conformal baseline instead uses Sw,i=∣yw,i−μw,i∣S_{w,i}=|y_{w,i}-\mu_{w,i}| and the 233rd ordered score, giving a query-specific but time-constant half-width. We do not assume exchangeability across years. Historical-template ECC-style reordering matches each 2011 date to its nearest available 2010 calendar day, then reorders current patch members by the patch-mean anomaly ranks of 100 jointly indexed historical members. Current within-patch distributions are preserved; unlike conventional ECC, the template is historical.

W3/Q3. Frozen modes and Table 3. Each day's 100 posterior latent samples are centered by that day's mean before pooling the 257 within-day covariances, Σ^2010=1257∑t199∑m(at(m)−aˉt)(at(m)−aˉt)⊤\widehat\Sigma_{2010}=\frac1{257}\sum_t\frac1{99}\sum_m(a_t^{(m)}-\bar a_t)(a_t^{(m)}-\bar a_t)^\top. Eigenvectors and eigenvalues are frozen on 2010. Revised Table 3 (original Table 2) reports 2010 modal shares using hj=Dgjh_j=Dg_j, qw=D⊤wq_w=D^\top w: field share λj∥hj∥2/∑kλk∥hk∥2\lambda_j\|h_j\|^2/\sum_k\lambda_k\|h_k\|^2, query share λj(gj⊤qw)2/∑kλk(gk⊤qw)2\lambda_j(g_j^\top q_w)^2/\sum_k\lambda_k(g_k^\top q_w)^2. Mode removal (revised Table 4) and coverage use 2011. Between-day mean variation is just 1.37% of uncentered pooled variance; the first three mode cosines are .993/.989/.983. Domain Mode-1 share changes only 99.70% to 99.46%, and rank-three recovery 97.7% to 97.6%.

W3/Q4. PSD validity. With current block covariance Cp,tC_{p,t} and frozen L=GrΛr1/2L=G_r\Lambda_r^{1/2}, the corrected covariance is Σ^r,t=blockdiag⁡p(Cp,t−LpLp⊤)+LL⊤\widehat\Sigma_{r,t}=\operatorname{blockdiag}_p(C_{p,t}-L_pL_p^\top)+LL^\top. It is PSD if Cp,t⪰LpLp⊤C_{p,t}\succeq L_pL_p^\top for every block. This holds on all 108,900 SST day–block pairs (worst normalized minimum eigenvalue 1.7×10−31.7\times10^{-3}). For other covariances, replace LL's contribution by γtLL⊤\gamma_tLL^\top and subtract γtLpLp⊤\gamma_tL_pL_p^\top within each block, choosing the largest γt∈[0,1]\gamma_t\in[0,1] making every residual block PSD. This always exists and preserves each current block covariance. Here γt=1\gamma_t=1, so the results are unchanged.

Q4. Recovery on 500 random queries. The 250 positive aggregates and 250 zero-mean contrasts give rank-three median variance recovery of 99.4% / 98.7%, IQR 96.1–103.8% / 92.4–106.9%, and 99.6% / 97.6% of queries within 80–120% recovery. Nominal-90% coverage is 88.8% / 91.7%, versus original joint benchmarks 88.9% / 91.9%. Rank one recovers 97.2% for positives but only 13.1% for contrasts.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.