{"id":"a8b595bf-3397-49c4-bfb6-873d60ac3445","arxiv_id":"2505.01260","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper reframes geographic dependency as a low-dimensional projection of a higher-dimensional covariate space, using existing theorems to argue that non-stationarity is better interpreted as omitted covariates than as a fundamental property of space.","lead":"A geostatistics paper argues that spatial autocorrelation, trends, and non-stationarity are not intrinsic properties of geographic space but are projections of unobserved covariates, so geographic coordinates are just one predictor among many.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central evidence is circular: Expression (8) learns a latent coordinate to force a stationary variogram, and the learned dimension resembles the response itself, so the demonstration cannot support the claim that meaningful covariates make geographic dependence non-intrinsic.","rationale":"The reader's verdict already identifies the same weak spot, and my read agrees. Section 2's equivalence between Bayesian linear regression and GP regression is standard, so I do not object to the algebraic claim that geographic coordinates can be treated as predictors. The problem is the semantic leap in Section 3. Theorem I is a finite-dimensional representation theorem: it guarantees that a given covariance matrix can be embedded as a stationary covariance in a higher-dimensional Euclidean space. It does not guarantee that the added coordinates are measurable covariates of the underlying process, and the paper provides no continuity or extension argument from finite Gaussian vectors to real data-generating processes. Expression (8) then optimizes the added coordinate against the same semivariances it is meant to explain, so the reported fit of Figure 5 is expected under the null hypothesis that the latent coordinate is pure overfitting. The paper's own observation that the learned coordinate tracks LST values reinforces this: a one-dimensional learned coordinate can simply encode the response. Therefore the Amsterdam case study does not test the central claim. This does not destroy the paper, because the thesis is useful as a modeling perspective, but the claim as stated ('in fact', 'proof') overreaches. The conditional verdict stands: the paper should be read as a conceptual synthesis with an illustrative demonstration, not as a proof of the ontological claim about geographic dependency.","tokens_in":11223,"tokens_out":5136,"duration_ms":59936,"concrete_test":"Re-run the Amsterdam demonstration on 1,000 random permutations of the 20 LST values over the same geographic coordinates, fitting Expression (8) with one extra dimension exactly as in the paper (Gaussian variogram, same starting values). If the optimized sum of squared variogram residuals for permuted data is comparable to or better than the value for the real data, then Expression (8) provides no evidence for meaningful covariates: it only shows that a latent coordinate can be chosen to make any response pattern look stationary. If the real-data fit is substantially better than all or nearly all permutations, the concern is weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The single load-bearing step is the move from Theorem I to the claim that non-stationarity in geographic space is 'essentially' variation along unobserved covariates. Theorem I (Perrin-Schlather, cited at [36]) guarantees that any finite covariance matrix can be embedded as the covariance of a stationary process on R^{s+d}; it says nothing about whether the extra coordinates correspond to real, measurable predictors, nor does it extend from a finite Gaussian vector to an arbitrary data-generating process. The demonstration intended to bridge this gap is Expression (8), which treats the new coordinate Z' as free parameters and minimizes the squared difference between observed semivariances and gamma_phi(d_ij([X,Z'])). Because Z' is optimized to make the variogram stationary, a good fit is built into the objective rather than discovered from data. The paper itself observes (Section 3, Figure 6) that the learned dimension resembles the LST values, confirming that the procedure can produce a covariate that is essentially a transformation of the response. Therefore the Amsterdam experiment cannot distinguish between 'meaningful covariates explain the pattern' and 'an unconstrained latent coordinate absorbs the non-stationarity.' The central claim remains a perspective, not an established result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that geographic dependency, including spatial autocorrelation, heterogeneity, and non-stationarity, should not be understood as an intrinsic property of geographic space. Instead, it proposes that observed geographic patterns are projections of a higher-dimensional covariate space onto the lower-dimensional geographic coordinate space, and that geographic coordinates are just one set of predictors among many. The paper builds this argument by reviewing the equivalence between ordinary linear regression in a feature space and Gaussian process regression, by arranging these modeling choices in a two-by-two matrix with geographic and non-geographic covariates, and by invoking a theorem of Perrin and Schlather to claim that any stochastic process can be viewed as stationary in higher dimensions. A demonstration with land surface temperature data from Amsterdam learns one additional coordinate for each sample by minimizing the mismatch between observed semivariances and a stationary Gaussian variogram, and the paper interprets the resulting stationary fit as support for its central claim.","tokens_in":11496,"tokens_out":3654,"duration_ms":42083,"significance":"If the central claim were established, the paper would provide a useful conceptual reframing of the first law of geography and of the role of stationarity in geostatistics, connecting geographic dependency with covariate-based modeling and dimension-expansion methods. The paper has genuine strengths: the equivalence between the weight perspective of linear regression and the function perspective of Gaussian process regression is standard and correctly presented; the two-by-two modeling matrix in Figure 2 is a clear pedagogical device; and the paper draws attention to an important and often unstated distinction between modeling choices and substantive interpretations of geographic pattern. The attempt to connect the Perrin-Schlater dimension-expansion result to a concrete geographic example is also valuable as an illustration. However, the strong ontological conclusion of the paper is not supported by the cited theorem or by the demonstration in Section 3.","major_comments":[{"comment":"The paper states that 'any random process is a sample of a stationary process in high dimension' and then uses this to conclude that non-stationarity in geographic space is essentially unexplained variation from other covariates. This overstates the cited result of Perrin and Schlather, which applies to any finite real-valued covariance matrix, equivalently to a finite Gaussian vector, and guarantees existence of a stationary Gaussian process on a higher-dimensional Euclidean space; it does not apply to arbitrary data-generating processes with continuous index sets, nor does it guarantee that the extra coordinates correspond to real, measurable predictors. This distinction is load-bearing because the paper's central claim depends on the extra dimensions being interpretable as covariates rather than as a purely mathematical embedding. The authors should restate the theorem accurately for finite Gaussian samples and explicitly discuss what additional assumptions are needed to extrapolate to real spatial processes.","section":"Section 3, Theorem I paragraph"},{"comment":"The Amsterdam demonstration is circular as evidence for the central claim. Expression (8) optimizes the new coordinates Z' to minimize the squared difference between observed semivariances and a stationary variogram gamma_phi, so the stationary fit shown in Figure 5 is a property of the objective rather than a discovery from the data. The paper itself observes that the learned dimension resembles the LST values, confirming that an unconstrained latent coordinate can simply absorb the response variation. To support the claim that meaningful covariates make geographic dependence non-intrinsic, the authors would need to validate the learned dimensions out of sample, compare against known covariates, constrain the number of extra dimensions and variogram hyperparameters, or test whether the learned coordinates correspond to independently measurable variables. As written, the experiment cannot distinguish the hypothesis 'unobserved covariates explain the pattern' from the hypothesis 'a flexible latent coordinate absorbs the non-stationarity.'","section":"Section 3, Expression (8) and Figures 5-6"},{"comment":"The paper conflates a modeling equivalence with an ontological conclusion. The fact that ordinary linear regression with basis functions and Gaussian process regression can be formulated in equivalent predictive terms shows that geographic coordinates and other covariates can play formally similar roles as predictors, but it does not establish that geographic dependency is 'essentially' a projection of high-dimensional covariate variation or that such dependency is not a property of geographic space. The authors should state explicitly that the equivalence is a modeling equivalence and then provide a separate, falsifiable argument for the stronger claim that geographic dependency is not intrinsic.","section":"Section 2, Expressions (4)-(7) and Figure 2"},{"comment":"The number of extra dimensions is a free parameter that the authors set to one without a selection criterion, and the variogram hyperparameters gamma_0 and a are also fitted rather than independently fixed. Because the demonstration is purely in-sample and admits these free parameters, a good fit is expected and does not by itself provide evidence for the proposed framework. The authors should report sensitivity to these choices and, if possible, a model-selection or validation procedure.","section":"Section 3, 'number of x must be chosen arbitrarily'"}],"minor_comments":[{"comment":"The phrase 'there is alway a hyperplane' contains a typo and should read 'there is always a hyperplane'.","section":"Section 3, first paragraph"},{"comment":"The sentence describing the four-dimensional space says 'the two dimensions of geographic coordinates slat,slon, and the dimension of their observed LST values x', while x was previously used for the new covariate dimension; this notation is confusing and should be clarified.","section":"Section 3, paragraph after Expression (8)"},{"comment":"The sentence 'f is explained by some predictor x1 while autocorrelated with itself as a combination of existing noisy observations weighted by covariance function defined on x2' is grammatically and conceptually unclear and should be rewritten.","section":"Section 2, Expression (7)"},{"comment":"The paper does not describe the optimization algorithm, initialization, or software used for Expression (8), nor does it provide code or data; adding these details would improve reproducibility and allow readers to assess the stability of the learned dimension.","section":"Section 3, Expression (8) implementation"},{"comment":"The final remarks raise the question 'Whether non-stationarity, trends or even dependency should be considered as properties intrinsic to geographic space?' but the paper does not provide a clear operational definition of 'intrinsic'; defining this term would sharpen the argument.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is best read as a conceptual essay proposing a reframing of geographic dependency rather than as a fully established empirical result. The central mathematical-equivalence material is sound, but the load-bearing interpretive step from that equivalence to the claim that geographic dependency is not intrinsic is not supported. The demonstration in Section 3 is circular and cannot, in its current form, provide independent evidence. I believe the manuscript can be revised within its own scope by clearly separating the modeling equivalence from the interpretive claim, correcting the statement of Theorem I, and either replacing the demonstration with a validation exercise or explicitly presenting it as an illustration of a possibility rather than as a proof of the claim. A major revision is therefore appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful thing about this paper is that it names a real interpretive problem and gives it a clean frame: spatial autocorrelation, heterogeneity and non-stationarity are not intrinsic to geographic space but are projections of a higher-dimensional feature space of all potential predictors. The authors are honest that the mathematical content is a compilation of existing results — GP/kriging equivalence, GP as Bayesian linear regression, and Perrin-Schlather's dimension-embedding theorem — and they cite those properly. As a conceptual synthesis, it works.\n\nWhat it does not do is establish the claim as a new result, and the paper's own demonstration is the weakest part. Theorem I holds for finite Gaussian vectors; the authors extend it to 'any random process' without proof. The Amsterdam LST example learns an extra coordinate by minimizing variogram misfit in Expression (8), so a good stationary fit is built into the objective rather than discovered. The learned dimension ends up resembling the response values themselves, which the authors notice and acknowledge. That means the demonstration cannot distinguish 'meaningful covariates explain the pattern' from 'an unconstrained latent coordinate absorbs the non-stationarity.' The stress-test note is right on this point.\n\nThe central claim remains defensible as a perspective, not established as a theorem. That is not fatal — the paper would be fine as a perspective piece if the authors reframed the demonstration as illustrative and disclosed the finite-sample scope of the embedding theorem. They also don't ship code or data, which is a missed opportunity for a paper whose whole point is interpretive.\n\nWho should read it: geographers and spatial statisticians who are tired of the trend-vs-stationarity debate and want a unified way to talk about omitted covariates. It deserves a serious referee, because the framing is genuinely useful and the authors are careful about attribution. I would send it out, with the clear expectation that the demonstration be reworked or downgraded, and the overgeneralization of Theorem I fixed. I wouldn't cite it in the next year, but I'd be happy to see it in the literature.","headline":"A clear conceptual synthesis that reframes spatial dependence as omitted covariates, but the demonstration is circular and the theorem is overextended; fine as a perspective piece, not as a new result.","tokens_in":11986,"tokens_out":2394,"would_cite":false,"duration_ms":23961,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M30","60G60"],"pacs":[],"model":"deepseek-v4-flash","headline":"Geographic dependency is a projection of hidden covariates","keywords":["geographic dependency","spatial autocorrelation","non-stationarity","Gaussian process regression","variogram","dimension expansion","geostatistics","covariate equivalence"],"falsifier":"Look for a geospatial dataset in which non-stationarity persists after a generous set of meaningful covariates is added and after many extra dimensions are learned: if the augmented variogram still shows clear distance-dependent structure, or if the learned coordinates fail to match any plausible driver and fail to predict held-out measurements, the claim that such dimensions always exist loses its edge. A sharper check is the paper's own setup: learn one or several extra coordinates on the Amsterdam LST samples, then test whether those coordinates correspond to land-surface material and generalize to unheld locations, rather than merely re-encoding the observed temperatures.","tokens_in":11034,"feed_emoji":"🗺️","tokens_out":7587,"duration_ms":74223,"temperature":0.7,"pith_summary":"This paper argues that the familiar geostatistical properties of geographic dependency—spatial autocorrelation, similarity, heterogeneity, and non-stationarity—are not intrinsic features of geographic space. Instead, they are a lower-dimensional projection of variation that unfolds along many predictor dimensions, and geographic coordinates play the same role as any other covariate in a regression or Gaussian process model. The authors support this by assembling known equivalences between weight-space linear regression and kernel-based Gaussian process regression, by invoking a theorem stating that any Gaussian vector can be viewed as a sample from a stationary process in a higher dimension, and by a demonstration in which land-surface-temperature samples are made to fit a stationary variogram after adding one learned extra coordinate. If the view is right, then choosing a trend, a stationary covariance, or a set of covariates is a modeling choice that changes interpretation, not a measurement of a geographic law.","feed_headline":"Geographic dependency is a projection of hidden covariates","feed_subtitle":"Non-stationarity shrinks to missing predictors when geographic coordinates are treated as ordinary covariates.","key_machinery":"The argument runs on three connected pieces. First, the equivalence between weight-space Bayesian linear regression with a basis expansion $\\phi(x)$ and function-space Gaussian process regression with covariance kernel $k(x,x')$, which lets geographic coordinates $s$ and other covariates $x$ be swapped freely as predictors in the paper's mixed model. Second, Theorem I (reference [36]): any real-valued covariance matrix on $s$ points can be realized as the restriction of a positive-definite stationary function on $\\mathbb{R}^{s+d}$, so every finite non-stationary geographic sample can be embedded in a higher dimension where stationarity holds. Third, the dimension-expansion objective of Expression (8), which learns new coordinate values $Z'$ for each sample by minimizing the squared mismatch between observed semivariances $v^*_{ij}$ and a stationary Gaussian variogram $\\gamma_\\phi$ evaluated at distances $d_{ij}([X,Z'])$ in the augmented space; the fitted variogram then becomes the load-bearing evidence that the projection picture is real.","core_discovery":"The central discovery this paper is trying to establish is that geoinformation dependency properties are the projection, onto the low-dimensional geographic plane, of a high-dimensional feature space formed by all potential predictors. More specifically, any observed pattern—whether it looks stationary, trending, heterogeneous, or purely random in geographic space—can be reframed as a pattern in a higher-dimensional covariate space in which geographic coordinates are just two or three of the predictors; non-stationarity is then a compressed summary of unexplained variation along missing covariate dimensions. Drawing on Theorem I of the literature (cited as [36]), the paper asserts that with properly specified extra dimensions a stationary description can always be found, and it demonstrates this with a dimension-expansion fit to urban land-surface-temperature data, where a stationary Gaussian variogram fits the augmented distances once one extra coordinate is learned. The paper therefore concludes that treating autocorrelation as a law of geography, and drawing a sharp line between trend and random variation, reflects model choice rather than an intrinsic geographic property.","pith_inferences":["A testable extension not developed in the paper: in a controlled dataset with known omitted drivers, adding those drivers to a Gaussian-process model should visibly reduce residual variogram noise and non-stationarity; if it does not, the projection claim needs qualification.","The single learned extra dimension in the Amsterdam demonstration may simply re-encode observed temperature values rather than a meaningful driver; a stronger check would learn several dimensions, inspect their correspondence to land cover or other covariates, and validate them on held-out locations.","The equivalence also suggests a practical diagnostic routine: compare variograms before and after dimension expansion as a scanner for missing covariates, rather than treating spatial correlation as irreducible noise.","If the view generalizes, it extends beyond geography: any low-dimensional projection of a high-dimensional process risks being misread as an intrinsic law of that low-dimensional space."],"forward_implications":["Non-stationarity and heterogeneity in a geographic dataset become diagnostic signals: they point to missing covariates rather than to properties of space itself.","Geographic coordinates and other covariates can be put on equal footing in the modelling matrix of the paper, so a spatial model, a Gaussian process on non-spatial predictors, and a land-use regression are variants of one regression framework rather than competing paradigms.","The distinction between a deterministic trend and a random spatially correlated component depends on model specification; the same data can be described either way without changing the underlying explanation.","If enough covariate dimensions can be found and learned, a single stationary variogram in the augmented space should capture the dispersion of observed values, making trend removal and non-stationary covariance functions unnecessary.","Interpreting spatial autocorrelation as a law of geography reopens: 'near things are more related' is a compressed statement about the covariates that happen to be included in the model."],"supporting_citations":[{"why":"Supplies the Gaussian-process / Bayesian linear regression equivalence (Expressions 2-6) that lets geographic coordinates be treated as ordinary predictors.","marker":"[27]"},{"why":"States the theorem that any multivariate Gaussian vector can be interpreted as a sample from a stationary process in higher dimension, the formal core of the projection claim.","marker":"[36]"},{"why":"Provides the dimension-expansion method used in Expression (8) and in the Amsterdam land-surface-temperature demonstration.","marker":"[38]"},{"why":"Extends nonstationarity in $\\mathbb{R}^n$ to second-order stationarity in $\\mathbb{R}^{2n}$, backing the theorem used for the high-dimensional embedding.","marker":"[37]"},{"why":"Frames the long-standing debate on stationarity and trends that the paper claims to resolve by moving to higher dimensions.","marker":"[16]"},{"why":"Supplies the observation that one person's nonstationarity is another person's random correlated variation, which motivates the interpretation of model choice.","marker":"[42]"}],"fun_headline_variants":["Non-stationarity is just missing predictors in disguise","Geographic dependency: a low-dim view of hidden covariates","Stationarity restored by adding one more covariate","Where's the trend? It's in the unobserved predictors","Geographic space is a projection of covariate space"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the premise that for every observed non-stationary geographic pattern there exist properly specified extra covariates that make the pattern stationary in a higher dimension, and that these covariates can actually be learned from the data rather than merely being a mathematical possibility.","fun_headline_variants_meta":{"raw":{"variants":["Non-stationarity is just missing predictors in disguise","Geographic dependency: a low-dim view of hidden covariates","Stationarity restored by adding one more covariate","Where's the trend? It's in the unobserved predictors","Geographic space is a projection of covariate space"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1406,"prompt_tokens":954,"completion_tokens":452,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":374}},"tokens_in":570,"tokens_out":452,"duration_ms":4557,"temperature":1.0,"reasoning_tokens":374,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:21:39.533474+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Look for a geospatial dataset in which non-stationarity persists after a generous set of meaningful covariates is added and after many extra dimensions are learned: if the augmented variogram still shows clear distance-dependent structure, or if the learned coordinates fail to match any plausible driver and fail to predict held-out measurements, the claim that such dimensions always exist loses its edge. A sharper check is the paper's own setup: learn one or several extra coordinates on the Amsterdam LST samples, then test whether those coordinates correspond to land-surface material and generalize to unheld locations, rather than merely re-encoding the observed temperatures.","supporting_citations":[{"cited_title":"Can any multivariate gaussian vector be interpreted as a sample from a stationary random process? Statistics & probability letters, 77(9):881–884, 2007","cited_arxiv_id":null,"evidence_quote":"States the theorem that any multivariate Gaussian vector can be interpreted as a sample from a stationary process in higher dimension, the formal core of the projection claim."},{"cited_title":"Modeling nonstationary processes through dimension expansion","cited_arxiv_id":null,"evidence_quote":"Provides the dimension-expansion method used in Expression (8) and in the Amsterdam land-surface-temperature demonstration."},{"cited_title":"Nonstationarity in rn is second-order stationarity in r2n","cited_arxiv_id":null,"evidence_quote":"Extends nonstationarity in $\\mathbb{R}^n$ to second-order stationarity in $\\mathbb{R}^{2n}$, backing the theorem used for the high-dimensional embedding."},{"cited_title":"To be or not to be","cited_arxiv_id":null,"evidence_quote":"Frames the long-standing debate on stationarity and trends that the paper claims to resolve by moving to higher dimensions."},{"cited_title":"Kriging nonstationary data","cited_arxiv_id":null,"evidence_quote":"Supplies the observation that one person's nonstationarity is another person's random correlated variation, which motivates the interpretation of model choice."}],"review_version":1}