{"id":"5dd797ad-6fc1-437c-830d-00e8cf2096f6","arxiv_id":"2509.00742","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A factor-augmented spatial autoregression can be estimated by a three-step procedure that is uniformly selection consistent over all response components.","lead":"This paper introduces factor-augmented spatial autoregression (FSAR), a model for high-dimensional multivariate spatial responses that estimates latent common factors in the errors. The three-step estimator (componentwise MLE, diversified projections, factor-augmented MLE with SCAD) comes with uniform selection consistency guarantees, verified through simulations and a 50-indicator Chinese urban economics dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Condition (C9)(2) on the projection matrix is not verified for the data-dependent random-partition M, and for an unaligned M its eigenvalue rate is false; the implemented procedure is not covered by the proof.","rationale":"The reader's weakest assumption—Condition (C9)(2)—is indeed load-bearing, and I agree that the paper does not verify it from data. My concern sharpens this: the eigenvalue condition λ_min(H^T H) ≫ 1/p is stronger than a generic 'not orthogonal' assumption; it requires H to have singular values diverging above the 1/√p scale. A purely random independent projection fails this at the stated rate, so the theorem only applies to M deliberately aligned with B*. Since the implementation estimates M from 10% of the data, the fixed-M theorem is not directly applicable, and the missing Appendix material does not fill the gap. This is a real correctness risk for the central FMLE and uniform-selection claims, but it is not a demonstrated contradiction; the data-driven M may well satisfy the condition in the simulations. Hence the reader's CONDITIONAL verdict remains appropriate. I set verdict_should_be to UNCHANGED because my analysis does not move the verdict, and agreement_with_reader is partial because I add the eigenvalue-rate/data-dependence point rather than merely restating the reader's orthogonality concern.","tokens_in":20808,"tokens_out":15284,"duration_ms":197974,"concrete_test":"Reproduce the exact simulation design of Section 3.2 using the GitHub code, with known B*, and record for every replication the realized H = M^T B*/p from the random-partition M. Check whether rank(H)=d and whether p λ_min(H^T H) diverges as p increases. If it does, the data-driven M does satisfy the eigenvalue rate in the tested settings, and the concern is reduced to a missing proof. If p λ_min stays O_p(1) or H loses rank, (C9)(2) is violated by the very procedure used in the simulations, and Theorems 2–4 do not explain the reported results. A direct contrast is to rerun with M chosen as independent Gaussian or ±1 projections; under that M, p λ_min should remain O_p(1), confirming that the condition is not automatic.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorems 2–4 all assume Condition (C9)(2): for H = M^T B*/p, rank(H)=d and λ_min(H^T H) ≫ 1/p. In Section 2.3, M is treated as a pre-specified matrix. In the simulations (Section 3.2) and real data example (Section 3.3), however, M is obtained by a 'random partition method' that uses 10% of the data to estimate M. No theorem states that this data-dependent M satisfies (C9)(2). This matters because (C9)(2) is not a mild generic condition: if M were an independent random projection with bounded entries and B* had bounded entries, each entry of H would be O_p(p^{-1/2}), so the eigenvalues of H^T H would be O_p(1/p) and p λ_min(H^T H) would be O_p(1), not diverge. The condition can hold only if M is strongly aligned with the loading space, e.g., M is a consistent PCA-type estimate of B*. But proving that alignment from the first subsample is absent from the arXiv text. If (C9)(2) fails, Theorem 2 fails, bZ is inconsistent for ZH^T, the FMLE likelihood in (2.5) is misspecified, and the asymptotic expansion leading to Theorem 3—and hence the uniform selection consistency in Theorem 4—does not follow. The omitted Appendices A–C also prevent verification of Eq. (2.7) and of the matrices A_j, D_j in Condition (C10).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a factor-augmented spatial autoregressive (FSAR) model for high-dimensional multivariate response spatial data. The error vector in a componentwise SAR model is given a fixed-dimensional factor structure. Estimation proceeds in three steps: componentwise QMLE residuals, factor estimation via diversified projections (Fan and Liao 2022), and a per-component factor-augmented QMLE; SCAD penalization with a BIC criterion is used for variable selection. The main theoretical claims are uniform consistency of the componentwise estimator (Theorem 1), consistency of the projected factor estimator (Theorem 2), sqrt(n)-consistency and asymptotic normality of the FMLE (Theorem 3), and uniform selection consistency of SCAD and BIC (Theorems 4-5). Simulations on three network structures and a 287-city macroeconomic dataset illustrate the method.","tokens_in":21211,"tokens_out":8158,"duration_ms":99023,"significance":"If correct, the paper offers a computationally attractive route to MSAR with many responses: per-component parallel estimation, a parsimonious factor-augmented error structure, and a uniform selection guarantee stronger than usual pointwise model selection. The use of the diversified projections idea in a spatial-autoregression context is a sensible extension, and the authors are explicit about tail, network, and identification conditions. The provision of code on GitHub is a plus. However, the proof of the central asymptotic results is not contained in the arXiv submission, and the connection between the theoretical Condition (C9) on the projection matrix and the data-dependent random-partition implementation is unproved; these issues need to be resolved before the claims can be accepted.","major_comments":[{"comment":"Condition (C9)(2) requires rank(H)=d and lambda_min(H^T H) >> 1/p with H = M^T B*/p. In the theory M is pre-specified, but in the numerical implementation (Section 3.2 and 3.3) M is obtained from 10% of the data by a random partition. No theorem proves that this estimated M satisfies (C9)(2). For a generic bounded random projection, entries of H are O_p(p^{-1/2}) and eigenvalues of H^T H are O_p(1/p), violating the condition; the condition requires M to be strongly aligned with B*. If it fails, Theorem 2 fails and the FMLE likelihood (2.5) is misspecified, so Theorems 3 and 4 do not cover the implemented procedure. A lemma showing that the random-partition construction yields (C9)(2) with high probability, or a reformulated condition on the estimated M, is needed.","section":"Section 2.3, 3.2; Theorems 2-4"},{"comment":"The arXiv version states that all proofs are in Appendices A-C, but these appendices are not included. The expansion (2.7) is central to Theorem 3, and the matrices A_j and D_j in Condition (C10) are defined only in the omitted (A.30) and (A.33). Without these materials the covariance formula in Theorem 3, the identifiability condition, and the uniform selection proofs in Theorems 4-5 cannot be checked. The authors should include the appendices or point to a publicly accessible supplement.","section":"Eq. (2.7), Condition (C10); Appendix"},{"comment":"The random-partition method itself is not fully specified: it is described only as estimating M from 10% of the data. The reader cannot determine whether M is a PCA loading estimate, an eigenvalue-thresholded estimate, or some other construction. This matters because the required alignment with B* in (C9)(2) is precisely what must be demonstrated for the method actually used in the simulations and real-data analysis.","section":"Section 2.3 and Section 3.2"}],"minor_comments":[{"comment":"The definition of the induced norms contains an undefined 'm': 'max_{1<=i<=m}' should presumably be 'max_{1<=i<=n}'.","section":"Condition (C6)"},{"comment":"The covariance matrices in Theorem 3 are hard to parse because of the block notation and the fact that Sigma_Qj is deferred to an omitted appendix. A clean partition of the parameter vector and a full definition of Sigma_Qj in the main text or a visible supplement would help.","section":"Theorem 3 display"},{"comment":"The statement that 'a CP_j is randomly selected over 1<=j<=p' is not reproducible. Reporting the distribution of coverage probabilities across j, or at least the range and median, would be more informative.","section":"Section 3.2, Table 1"},{"comment":"The factor dimension in the real-data example is selected by eigenvalue ratios without formal post-selection justification. A reference to Bai and Ng (2002) or Lam and Yao (2012) for the ratio-based rule would strengthen the presentation.","section":"Section 3.3"}],"recommendation":"major_revision","confidential_remarks":"The main gap is the unverified alignment condition for the random-partition projection matrix; this is not merely a theoretical subtlety because the simulation and real-data results are presented as evidence for the theorems. The omission of appendices from the arXiv version makes the manuscript difficult to referee. The paper is otherwise clearly written and the central modeling idea is valuable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a solid new model for high-dimensional multivariate spatial data, but there is a real gap between the theoretical assumptions on the projection matrix and what the simulations actually do. Worth refereeing, but the authors need to close that gap.\n\nWhat's new: the FSAR model—componentwise SAR with a factor structure on the error vector—plus a three-step estimator: CMLE for initial residuals, diversified projections for factors, then a SCAD-BIC for each component. Theorem 4, uniform selection consistency over all p components, is a real theoretical contribution. The parallelizable structure is attractive. Simulations match the claimed rates, and the real-data example is sensible and interpretable.\n\nSoft spots, in order of importance.\n\n1. Condition (C9) and the data-dependent M. The theorems take M as pre-specified. In Section 2.3 they define M that way, but in simulations and the real data they construct M via a random partition on 10% of the data. Nothing in the text proves this data-dependent M satisfies (C9)(2). The stress-test note is correct: for an independent random projection with bounded entries, H = M^T B*/p has entries O_p(p^{-1/2}), so eigenvalues of H^T H are O_p(1/p), not the required lambda_min >> 1/p. You need M aligned with the loading space. If the random-partition method uses PCA from the first subsample, the alignment likely holds under strong factors, but that needs to be stated and proved. This is a genuine gap, though probably fixable.\n\n2. Proofs are not in the arXiv version. Appendices A–C are referenced for Theorems 1–5 and for equations (2.7), (A.23), etc., but the PDF ends before the appendices. A referee cannot verify the asymptotic expansion or the definitions of A_j and D_j in (C10). That is a practical obstacle for review.\n\n3. Minor: conditions (C10) (row-sparsity of Sigma_omega) and (C11) (fixed risk separation) are strong but explicit. The BIC penalty factor (log n)(log pq)^{2/alpha}/n is plausible and standard.\n\nCitation pattern is fine: they build on Fan–Liao, Lee, Zhu et al., and credit them properly.\n\nWho this is for: spatial econometrics and high-dimensional factor modeling. It deserves a serious referee. I would send it out, but ask for a lemma or remark connecting the random-partition M to (C9), and for the appendices to be made available. Also check in simulation whether the 10% subsample PCA always yields an M satisfying the eigenvalue condition—sensitivity to weak factors should be examined.","headline":"A genuinely useful new FSAR model for high-dimensional spatial data, but the gap between the theory (fixed M satisfying an eigenvalue condition) and the practice (data-dependent random-partition M) needs bridging before the main theorems can be taken at face value.","tokens_in":21610,"tokens_out":3734,"would_cite":true,"duration_ms":46196,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M30","62H25","62J07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper's central claim is that a factor-augmented spatial autoregression (FSAR) can estimate hidden common shocks from high-dimensional responses and select the true covariates for every response component simultaneously with probability","keywords":["factor-augmented spatial autoregression","multivariate spatial data","latent factor model","diversified projections","SCAD variable selection","uniform selection consistency","high-dimensional response","network data"],"falsifier":"Simulate data with the true loading matrix B* nearly orthogonal to a fixed projection M (so H = M^T B*/p has a singular value near zero), then run the full three-step procedure; the theorem predicts the FMLE and SCAD selection should lose consistency, while any setup that still yields good estimates would contradict the role of Condition (C9)(2).","tokens_in":20765,"feed_emoji":"📊","tokens_out":7381,"duration_ms":86718,"temperature":0.7,"pith_summary":"This paper is trying to establish that multivariate spatial autoregression remains feasible when the number of responses p diverges with the sample size, a regime where existing multivariate spatial models become computationally prohibitive. Its proposal, the FSAR model, imposes a fixed-dimensional factor structure on the p-dimensional error vector, cutting the parameter count from O(p^2) to O(p(d+q)). The main theoretical claims are that a diversified-projection step recovers the latent factors from componentwise residuals, that the factor-augmented likelihood estimator is √n-consistent and asymptotically normal, and that a SCAD penalty with a BIC-type tuning rule selects the true covariate set for every response component simultaneously with probability tending to one. A reader should care because the method turns an apparently intractable high-dimensional spatial estimation problem into p parallel low-dimensional regressions while keeping uniform statistical guarantees.","feed_headline":"Spatial regressions scale to thousands of outcomes via latent factors","feed_subtitle":"A projection step recovers hidden common shocks, enabling per-outcome fitting with uniform selection guarantees.","key_machinery":"The central object is the projection matrix M together with the rotated factor matrix H = M^T B^*/p. The diversified projections estimator bZ_i = M^T bε_i / p estimates Z_i H^T; Condition (C9)(2), which requires rank(H)=d and comparable eigenvalues for H^T H, is what guarantees the projection preserves the factor signal. The second load-bearing mechanism is the componentwise SCAD likelihood with the BIC penalty (log n)(log(pq))^{2/α}/n, which drives the uniform selection consistency. These two pieces convert one high-dimensional factor-plus-spatial problem into p small penalized regressions.","core_discovery":"Treating each response dimension as its own spatial autoregression and placing a shared factor model on the errors yields the FSAR model. The paper shows the latent factors can be learned by diversified projections: projecting the estimated residual matrix onto a well-chosen matrix M makes M^T ε_i/p concentrate around a rotated version of the latent factor Z_i. Once the estimated factors are inserted as covariates, each component is fit by maximum likelihood with a SCAD penalty. The decisive result is uniform selection consistency: under λ_n → 0 and √n λ_n / (log(pq))^{1/α} → ∞, the selected model equals the true model for every j = 1,...,p with probability going to one. The FMLE is √n-consi","pith_inferences":["The factor-recovery step is agnostic about the sparse estimator used afterward, so swapping SCAD for another componentwise penalty should preserve the core argument and may allow post-selection inference.","Since Condition (C9)(2) is not verified by the data, a practical stability check is to repeat the analysis over several projection matrices and compare estimated factors and selected models; disagreement warns that the projection may be nearly orthogonal to the loadings.","The (log pq)^{2/α} factor in the BIC penalty implies that when p and q grow at different rates, the effective signal requirement tightens; applied users should monitor selection stability instead of trusting all p models simultaneously.","For small or slowly growing p, the consistency guarantee disappears; in that regime a dynamic factor or full multivariate spatial model is a safer alternative."],"forward_implications":["Full multivariate spatial autoregression requires O(p^2) parameters; under FSAR the count drops to O(p(d+q)) and each response component can be fitted in parallel.","Factor-estimation error is not ignored: it appears in the FMLE's asymptotic covariance, so standard errors remain valid even though the factors are estimated rather than observed.","As long as the response dimension p is large relative to √n, latent common shocks can be separated from spatial spillover effects in estimation.","The SCAD-plus-BIC procedure recovers all p true covariate sets simultaneously, so model selection no longer has to be checked one outcome at a time."],"supporting_citations":[{"why":"Supplies the diversified projections estimator that recovers latent factors from high-dimensional residuals.","marker":"Fan and Liao (2022)"},{"why":"Provides the quasi-maximum likelihood theory for the componentwise SAR estimators that seed the residual-based factor step.","marker":"Lee (2004)"},{"why":"Introduces the SCAD penalty whose oracle and selection properties the paper's uniform selection theorem extends.","marker":"Fan and Li (2001)"},{"why":"Establishes the BIC-style tuning-parameter selection for SCAD that the paper modifies for diverging p and q.","marker":"Wang et al. (2007)"},{"why":"Defines the multivariate spatial autoregression model whose computational burden the FSAR model is designed to avoid.","marker":"Zhu et al. (2020)"},{"why":"Provides the approximate factor model framework for the error covariance that motivates imposing a factor structure on the error vector.","marker":"Fan et al. (2008)"}],"fun_headline_variants":["Latent factors unlock high-dimensional spatial autoregression","Projection step recovers hidden shocks for spatial models","FSAR: spatial regression that handles thousands of outcomes","Spatial regression, now with latent factors for big data","A factor-augmented SAR model scales to high dimensions"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"Everything rests on the analyst's choice of projection matrix M: if M is nearly orthogonal to the true factor loadings, the factor estimates never converge to the latent signals, and the subsequent likelihood and selection results collapse.","fun_headline_variants_meta":{"raw":{"variants":["Latent factors unlock high-dimensional spatial autoregression","Projection step recovers hidden shocks for spatial models","FSAR: spatial regression that handles thousands of outcomes","Spatial regression, now with latent factors for big data","A factor-augmented SAR model scales to high dimensions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000855,"raw_usage":{"total_tokens":3570,"prompt_tokens":782,"completion_tokens":2788,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":2710}},"tokens_in":526,"tokens_out":2788,"duration_ms":24784,"temperature":1.0,"reasoning_tokens":2710,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:14:54.951595+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate data with the true loading matrix B* nearly orthogonal to a fixed projection M (so H = M^T B*/p has a singular value near zero), then run the full three-step procedure; the theorem predicts the FMLE and SCAD selection should lose consistency, while any setup that still yields good estimates would contradict the role of Condition (C9)(2).","supporting_citations":[{"cited_title":"(2004), Asymptotic distributions of quasi-maximum likelihood estimators for spatial autoregressive models, Econometrica, 72, 1899--1925","cited_arxiv_id":null,"evidence_quote":"Provides the quasi-maximum likelihood theory for the componentwise SAR estimators that seed the residual-based factor step."},{"cited_title":"and Li, R","cited_arxiv_id":null,"evidence_quote":"Introduces the SCAD penalty whose oracle and selection properties the paper's uniform selection theorem extends."},{"cited_title":"(2007), Tuning parameter selectors for the smoothly clipped absolute deviation method, Biometrika, 94, 553--568","cited_arxiv_id":null,"evidence_quote":"Establishes the BIC-style tuning-parameter selection for SCAD that the paper modifies for diverging p and q."},{"cited_title":"(2020), Multivariate spatial autoregressive model for large scale social networks, Journal of Econometrics, 215, 591--606","cited_arxiv_id":null,"evidence_quote":"Defines the multivariate spatial autoregression model whose computational burden the FSAR model is designed to avoid."},{"cited_title":"(2008), High dimensional covariance matrix estimation using a factor model, Journal of Econometrics, 147, 186--197","cited_arxiv_id":null,"evidence_quote":"Provides the approximate factor model framework for the error covariance that motivates imposing a factor structure on the error vector."}],"review_version":1}