{"id":"85e2d650-641f-4b8b-a880-b19a9be0278b","arxiv_id":"2506.09267","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"For Gaussian random fields observed on a fixed domain, the regression slope is consistently estimable when the exposure is not too smooth relative to the confounder, and a spectral ratio condition is necessary.","lead":"This paper characterizes when the regression slope between two spatial processes can be consistently estimated even when an unmeasured confounder affects both. It gives smoothness-based conditions and simple differencing estimators, with sharp results for common covariance families.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Necessity proof of Corollary 1(b) covers only independent X and W; the correlated case, which is the paper's central spatial-confounding claim, is left unsupported.","rationale":"The reader's weakest-assumption assessment is exactly right and identifies the single most load-bearing gap in the paper. The sufficient conditions and the explicit differencing/Laplacian estimators are well developed and represent a substantial contribution. The multivariate equivalence theorem (Theorem 4) is also novel and appears reusable. The problem is that the necessity direction, which is essential for the claimed sharp characterization, is proved only for independent X and W. Corollary 1(b) is stated for general non-degenerate bivariate Matern fields but its proof selects rho_XW = 0 and reduces to Theorem 5, an independence-only result. The same logical gap affects the analogous non-estimability claims for other covariance families. This is not an attack on the truth of the result: a direct application of Theorem 4 to the full Matern spectral matrix with Phi = diag(f_X, f_W) suggests the gap is fillable, because the additional cross-spectral terms are controlled by the same f_X/f_W integrability condition under standard Matern cross-spectrum tail behavior. But that computation is absent from the paper, so the sharpness claim for correlated fields is currently under-supported. Because the concern is addressable by a proof extension rather than a demonstrated counterexample, the appropriate verdict remains conditional acceptance, matching the reader's verdict. No change to the reader's assessment is needed.","tokens_in":71790,"tokens_out":8412,"duration_ms":97933,"concrete_test":"Apply Theorem 4 to the full bivariate Matern spectral density matrix rather than the independent submodel. For beta=0 and beta=1, write F_beta = [[f_X, f_XW + beta f_X], [f_XW + beta f_X, f_W + 2 beta f_XW + beta^2 f_X]], with f_X, f_W, f_XW the Matern spectral densities of X, W, and their cross-spectrum. Take Phi = diag(f_X, f_W) and check condition (13) using the fact that the coherence |f_XW|/sqrt(f_X f_W) is bounded away from 1 for a non-degenerate bivariate Matern field. Then compute the integral in (14). For nu_X > nu_W + d/2, the term from the (1,2) and (2,2) entries is controlled by integrability of f_X/f_W and (f_XW/f_W)^2; for valid bivariate Matern cross-spectra with nu_XW >= (nu_X + nu_W)/2, both integrals are finite. If the integral is finite, equivalence holds for correlated fields and Corollary 1(b) is true but under-proved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline result is an 'if and only if' for spatial confounding, where W is explicitly correlated with X. The sufficient side (Theorems 2 and 3) does handle correlation, but the necessity side does not. Corollary 1(b) states that for a non-degenerate bivariate Matern field with nu_X > nu_W + d/2, beta is not consistently estimable by any estimator. The proof in Section S3.3 considers only the two parameter choices (beta=0, rho_XW=0) and (beta=1, rho_XW=0), and then invokes Theorem 5, which is stated only for independent X and W. Thus the proof establishes non-estimability only on the diagonal submodel where X and W are independent. The same gap appears in Corollary 3(b), the power-exponential equivalence direction, and in the corresponding necessity statements in Corollaries 4, 7, and 8. This is load-bearing because the central claim is that the slope is consistently estimable if and only if the exposure is not more than d/2 degrees smoother than the confounder under correlated confounding. If adding cross-correlation between X and W can make different beta values distinguishable in the regime nu_X > nu_W + d/2, then part (b) is false as stated. If it cannot, then the theorem is true but the proof as written does not support it. The reader's concern is therefore correct: the sharpness of the characterization for correlated fields is not established by the supplied argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies infill (fixed-domain) consistent estimability of the slope β in the spatial regression Y(s) = X(s)β + W(s), where X and W are correlated Gaussian random fields and W is unobserved. The central claim is an 'if and only if' characterization: β is consistently estimable if and only if the exposure is not too smooth relative to the confounder, roughly α_X < α_W + d, together with a condition that the cross-covariance is smoother than the covariance of X (α_XW > α_X). The sufficient direction is established via explicit local-differencing or discrete-Laplacian OLS estimators, whose consistency is proved without parametric covariance knowledge. The necessary direction is approached through a new multivariate Paley-Wiener equivalence theorem, yielding the spectral condition ∫ f_X/f_W < ∞ as implying non-estimability. For Matern, power exponential, generalized Cauchy, and coregionalization families, the paper claims the sufficient and necessary conditions coincide (except at a boundary), giving a sharp characterization. The paper also treats multivariate exposures, measurement error, some non-stationary and non-Gaussian processes, irregular designs, and provides simulations that match the predicted rates.","tokens_in":72050,"tokens_out":8655,"duration_ms":86152,"significance":"If the characterization is correct, this is a substantial and long-needed contribution to the spatial confounding literature. The paper moves beyond estimator-specific analyses and gives a fundamental information-theoretic boundary for when the slope can be recovered at all, even if the entire bivariate field were observed. The main strengths are: (i) explicit, parameter-free estimators (differencing and Laplacian OLS) that do not require fitting covariance parameters; (ii) a genuinely new sufficient condition for equivalence of multivariate Gaussian measures with components of unequal smoothness (Theorem 4), which is of independent interest; (iii) detailed proofs of the variance rates, with simulations that empirically confirm the predicted convergence rates; and (iv) a broad set of extensions (multivariate X, measurement error, non-stationarity, heavy tails, irregular designs). The central weakness is that the necessity direction is proved only for the independent submodel X ⊥ W, whereas the headline claim concerns the spatially confounded setting where X and W are correlated.","major_comments":[{"comment":"The necessity direction for the correlated case is not established. Corollary 1(b) states that for a non-degenerate bivariate Matern field with ν_X > ν_W + d/2, β is not consistently estimable by any estimator. The proof in Section S3.3 explicitly considers only the two parameter choices (β=0, ρ_XW=0) and (β=1, ρ_XW=0), and then invokes Theorem 5, which is stated in the main text for 'independent stationary GRFs' X and W. Thus the supplied argument proves non-estimability only on the diagonal submodel where X and W are independent. The same gap appears in Corollary 3(b) (power exponential), Corollary 4(b) (generalized Cauchy, d=1), Corollaries 7(b) and 8(b), and the second (non-estimability) part of Theorem 6. This is load-bearing because the paper's abstract and Section 1 advertise an if-and-only-if characterization for spatial confounding, which is precisely the setting where W is correlated with X. If adding cross-correlation can make different β values distinguishable in the ν_X > ν_W + d/2 regime (e.g., through a sufficiently smooth cross-covariance), then part (b) is false as stated; if it cannot, the result may be true but the proof as written does not support it. The manuscript should either extend the equivalence argument to the correlated case (for example, by proving a bivariate spectral condition involving the cross-spectrum, or by a coupling argument showing that equivalence under independence implies equivalence under any valid cross-correlation) or explicitly restrict the non-estimability claims to independent X and W, which would substantially weaken the central claim.","section":"§S3.3, proof of Corollary 1(b); §4.3, Theorem 5"}],"minor_comments":[{"comment":"The text says 'generalized Cauchy or powered exponential'; 'powered exponential' should be 'power exponential' for consistency with Section 5.2.","section":"§4.3, after Theorem 5"},{"comment":"In the case analysis, the case where F^{(0)}(ω) ≥ F^{(1)}(ω) pointwise is not explicitly treated; it follows by symmetry but should be stated to avoid a gap in the reader's eyes.","section":"§4.2, proof of Theorem 4"},{"comment":"The boundary case ν_X = ν_W + d/2 is only partially resolved (d=1, integer ν_X) and the general case is left as a conjecture. This is fine, but the wording 'sharp characterization' in the abstract and Section 5 should be tempered to 'sharp except at the boundary' whenever referencing the boundary.","section":"§5.1, Corollary 2 and the following paragraph"},{"comment":"The local-averaging-and-differencing estimator in Theorem 6 depends on the choice of the averaging block size (indexed by ρ in the proof), but the statement does not tell the reader how to choose it in practice; a short remark would help.","section":"§6, Theorem 6 and Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a serious and mostly well-executed contribution to the spatial confounding literature, and the sufficient direction together with the new multivariate equivalence theorem are valuable. However, the necessity direction for correlated X and W — the central claimed contribution — is proved only for the independent submodel. This is a load-bearing gap. The authors should either close it or restrict the non-estimability claims, which would change the paper's scope. No other concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper. The sufficient side is genuine: Theorems 2 and 3 give clean conditions (alpha_11 < alpha_22 + d and alpha_12 > alpha_11) and explicit differencing/Laplacian estimators that are consistent without estimating covariance parameters, and the simulations support the rates. The multivariate equivalence theorem (Theorem 4) allowing different smoothnesses per component is a real contribution and likely reusable beyond this setting.\n\nThe soft spot is the necessity direction, and it is load-bearing. Corollary 1(b) states non-estimability for a general non-degenerate bivariate Matérn field with nu_X > nu_W + d/2, but the proof in Section S3.3 only builds equivalent measures for the independent submodel (rho_XW = 0) and then invokes Theorem 5, which is explicitly stated only for independent X and W. The same gap runs through Corollary 3(b) and the corresponding necessity statements in Corollaries 4, 7, and 8. So the paper's central claim, a sharp characterization under spatial confounding with correlated X and W, is not supported by the supplied argument. If cross-correlation can distinguish beta in that regime, part (b) is false; if it cannot, the proof still doesn't show it. That is not a minor technicality.\n\nThe boundary case is honestly left as a conjecture for non-integer smoothness, and the authors flag that. The GLS inconsistency side result is fine. No other concerns.\n\nWho is this for: anyone working on spatial confounding or on equivalence of multivariate GRFs. They get a strong sufficient theory and a reusable tool, plus a clearly labeled gap on the necessary side. It deserves a serious referee: the sufficient theory alone is worth refereeing, and the necessity claim needs either a real proof for correlated fields or an honest restriction to the independent submodel. I would engage with it, but I would insist the gap be addressed before publication.\n\nRecommendation: send to peer review, with a referee who checks whether the equivalence argument can be extended to correlated fields.","headline":"Sufficient theory and the new multivariate equivalence result are solid, but the paper's central sharp 'only if' for correlated confounders is only proved for independent X and W.","tokens_in":72592,"tokens_out":3286,"would_cite":true,"duration_ms":33154,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M30","60G60","62F12","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"In spatial regression with an unmeasured confounder, the slope is consistently estimable unless the exposure is more than d/2 degrees smoother than the confounder; when estimable, local differencing estimators recover it without knowing…","keywords":["spatial confounding","Gaussian random fields","infill asymptotics","consistent estimability","principal irregular term","Matérn covariance","local differencing","spectral equivalence"],"falsifier":"Simulate a bivariate Matérn field with $\\nu_X>\\nu_W+d/2$ and nonzero cross-correlation between $X$ and $W$, observe $(X,Y)$ on a regular grid with growing density, and compute the likelihood ratio or the maximum likelihood estimate of $\\beta$ for two different slope values; if the likelihood separates or the estimator concentrates as the grid densifies, the claimed non-estimability regime fails for correlated confounding.","tokens_in":71544,"feed_emoji":"📈","tokens_out":10567,"duration_ms":100328,"temperature":0.7,"pith_summary":"This paper asks a basic question that the spatial confounding literature has only answered empirically or estimator-by-estimator: if the full processes $Y$ and $X$ were observed on a fixed spatial domain, with an unmeasured process $W$ entering $Y = X\\beta + W$ and correlated with $X$, when does any estimator of the slope $\\beta$ exist? The answer given is a smoothness comparison: $\\beta$ is consistently estimable under infill asymptotics when the exposure is not too smooth relative to the confounder and the cross-covariance is smoother than the exposure covariance. The claim matters because it turns a widely debated bias problem into a checkable condition on local behavior of covariances, and because the proof supplies explicit non-parametric estimators—discrete differences or Laplacians of suitable order—that need no covariance parameters. For the Matérn, power exponential, generalized Cauchy, and coregionalization families the paper shows the condition is sharp, with a $d/2$ smoothness gap in $d$ dimensions; beyond that gap no estimator can be consistent, even in principle. A by-product is a new spectral criterion for equivalence of multivariate Gaussian random fields whose components have different smoothness.","feed_headline":"Spatial slope recoverable unless exposure is too smooth","feed_subtitle":"For Matérn fields the dividing line is sharp: β is identified whenever ν_X < ν_W + d/2.","key_machinery":"The load-bearing object is the principal irregular term of a covariance function: the leading non-analytic term $c\\,t^{\\alpha}$ in the near-zero expansion, whose exponent $\\alpha$ measures the smoothness of the process. Under the estimability conditions, the slope $\\beta$ appears as the ratio of the principal irregular coefficients of the cross-covariance and of the covariance of $X$. The constructive estimators use local differencing or discrete Laplacians of order high enough to annihilate the analytic part of the covariance, leaving the principal irregular term to dominate; the OLS ratio of differenced $Y$ to differenced $X$ then converges to $\\beta$ through variance calculations built on stationarity. For the impossibility direction, the machinery is spectral equivalence in multivariate Paley-Wiener spaces: comparing the spectral density matrices of $(X,Y)$ under two values of $\\beta$, the paper proves a sufficient condition for equivalence of multivariate Gaussian measures with component fields of different smoothness, which yields the $\\int f_X/f_W<\\infty$ criterion.","core_discovery":"The paper's central claim is that for $Y(s)=X(s)\\beta+W(s)$ with $(X,W)$ a bivariate Gaussian random field on a fixed domain in $\\mathbb{R}^d$, the slope $\\beta$ is consistently estimable from the paths of $(X,Y)$ exactly when the exposure is not too smooth relative to the confounder. Writing $\\alpha_{11}$, $\\alpha_{22}$, and $\\alpha_{12}$ for the exponents of the principal irregular terms of the covariance of $X$, of $W$, and of the cross-covariance, the sufficient conditions are $\\alpha_{11}<\\alpha_{12}$ and $\\alpha_{11}<\\alpha_{22}+d$; in that regime $\\beta$ equals the limiting ratio of the principal irregular coefficients and is recovered by regressing $p$-th order differences ($d=1$) or $m$-th order discrete Laplacians ($d>1$) of $Y$ on those of $X$. The necessary condition is spectral: when $X$ and $W$ are independent and $\\int f_X/f_W<\\infty$, the measures on $(X,Y)$ for two different $\\beta$ values are equivalent, so no consistent estimator exists even with complete observation. For Matérn, power exponential, generalized Cauchy, and coregionalization covariances the sufficient and necessary conditions coincide except at a boundary, giving a sharp characterization: $\\beta$ is consistently estimable if $\\nu_X<\\nu_W+d/2$ and $\\nu_{XW}>\\nu_X$, and not estimable if $\\nu_X>\\nu_W+d/2$.","pith_inferences":["If the paper's unstated premise survives—that cross-correlation cannot restore identifiability where independence already prevents it—the results imply a practical diagnostic: estimate the local smoothness exponents of $X$ and $Y$ from data, and when $\\alpha_{11}\\ge\\alpha_{22}+d$ abandon slope estimation on these variables alone and look for instruments or additional covariates.","The new multivariate spectral equivalence theorem is a transferable tool; it should apply to identifiability of cross-covariance parameters and, as the authors conjecture, to Gaussian fields on manifolds, where the dimension gap $d/2$ reappears through eigenvalue growth.","The boundary $\\nu_X=\\nu_W+d/2$ is where the interesting rates live: the paper shows consistency survives on the boundary in one dimension only at a logarithmic rate and only for integer smoothness, so higher-dimensional boundary cases likely need estimators beyond plain differencing or Laplacians."],"forward_implications":["If the sharp characterization holds, consistent estimation of $\\beta$ requires no knowledge of the covariance family or its parameters: only the order of differencing or Laplacian chosen from the smoothness of $X$.","Under spatial confounding, $\\beta$ is consistently estimable even when the exposure is up to $d/2$ degrees smoother than the confounder, not only when it is rougher.","For Matérn, power exponential, generalized Cauchy, and coregionalization processes, $\\nu_X>\\nu_W+d/2$ is an impossibility boundary: no estimator based on the observed $(X,Y)$ paths on the fixed domain can be consistent.","Measurement error does not change the estimability boundary; with noise, a local-averaging-then-differencing estimator is consistent under the same smoothness conditions.","With multivariate exposures, the condition must hold for every nonzero linear combination of the exposures, which restricts how collinear or smooth the exposure components may be relative to the confounder."],"supporting_citations":[{"why":"Supplies the principal-irregular-term formalism and the univariate spectral equivalence theory the paper extends to multivariate fields.","marker":"[38]"},{"why":"Provides the spectral-density condition for equivalence of univariate Gaussian fields that underlies Theorem 4 and the necessary condition in Theorem 5.","marker":"[37]"},{"why":"States the prior multivariate equivalence condition requiring equal smoothness, which the paper's diagonal-Phi condition generalizes.","marker":"[3]"},{"why":"Establishes the $d/2$ threshold and GLS consistency when $X$ is independent of $W$, the baseline setting the paper extends to correlated confounding.","marker":"[46]"},{"why":"Shows GLS can be inconsistent when $X$ is smooth relative to $W$, a precursor to the impossibility claims.","marker":"[43]"},{"why":"Analyzes GLS consistency through RKHS membership of a fixed $X$, framing the estimator-specific results this paper moves beyond.","marker":"[5]"},{"why":"Defines the valid bivariate Matérn cross-covariance family used in the sharp Corollary 1 characterization.","marker":"[16]"}],"fun_headline_variants":["Spatial slope recoverable when exposure rougher than confounder","Spatial confounding: slope identifiable unless exposure too smooth","Exposure smoothness decides spatial slope estimability","Sharp smoothness condition for spatial slope identification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The impossibility half of the sharp Matérn characterization is proved only for the special case where $X$ and $W$ are independent; the paper relies on the unstated premise that adding cross-correlation between $X$ and $W$ cannot make two different slopes distinguishable when they are already indistinguishable under independence.","fun_headline_variants_meta":{"raw":{"variants":["Spatial slope recoverable when exposure rougher than confounder","Spatial confounding: slope identifiable unless exposure too smooth","Exposure smoothness decides spatial slope estimability","Sharp smoothness condition for spatial slope identification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000388,"raw_usage":{"total_tokens":2160,"prompt_tokens":1174,"completion_tokens":986,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":790,"completion_tokens_details":{"reasoning_tokens":921}},"tokens_in":790,"tokens_out":986,"duration_ms":11965,"temperature":1.0,"reasoning_tokens":921,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:54:40.809024+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a bivariate Matérn field with $\\nu_X>\\nu_W+d/2$ and nonzero cross-correlation between $X$ and $W$, observe $(X,Y)$ on a regular grid with growing density, and compute the likelihood ratio or the maximum likelihood estimate of $\\beta$ for two different slope values; if the likelihood separates or the estimator concentrates as the grid densifies, the claimed non-estimability regime fails for correlated confounding.","supporting_citations":[{"cited_title":"Springer Science & Business Media, 1999","cited_arxiv_id":null,"evidence_quote":"Supplies the principal-irregular-term formalism and the univariate spectral equivalence theory the paper extends to multivariate fields."},{"cited_title":"Equivalence of Gaussian measures of multivariate ran- dom fields.Stochastic Environmental Research and Risk Assessment, 29:325–334, 2015","cited_arxiv_id":null,"evidence_quote":"Provides the spectral-density condition for equivalence of univariate Gaussian fields that underlies Theorem 4 and the necessary condition in Theorem 5."},{"cited_title":"Asymptotically equivalent prediction in multivariate geostatistics.Bernoulli, 28(4): 2518–2545, 2022","cited_arxiv_id":null,"evidence_quote":"States the prior multivariate equivalence condition requiring equal smoothness, which the paper's diagonal-Phi condition generalizes."},{"cited_title":"PhD thesis, University of Maryland, College Park, 2022","cited_arxiv_id":null,"evidence_quote":"Establishes the $d/2$ threshold and GLS consistency when $X$ is independent of $W$, the baseline setting the paper extends to correlated confounding."},{"cited_title":"Springer Science & Business Media, 2003","cited_arxiv_id":null,"evidence_quote":"Shows GLS can be inconsistent when $X$ is smooth relative to $W$, a precursor to the impossibility claims."},{"cited_title":"Spatial self-confounding: Smoothness-related estimation bias in spatial regression models.Biometrika, page (In press), 2025","cited_arxiv_id":null,"evidence_quote":"Analyzes GLS consistency through RKHS membership of a fixed $X$, framing the estimator-specific results this paper moves beyond."},{"cited_title":"Mat´ ern cross-covariance functions for multivariate random fields.Journal of the American Statistical Association, 105(491):1167–1177, 2010","cited_arxiv_id":null,"evidence_quote":"Defines the valid bivariate Matérn cross-covariance family used in the sharp Corollary 1 characterization."}],"review_version":1}