{"id":"cb314105-2b4c-406b-b650-49e67f30d34b","arxiv_id":"1908.02493","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The HPE projects pinned Euler characteristic curves onto Hermite-weighted EC densities to estimate Lipschitz-Killing curvatures without assuming stationarity, with a bootstrap extension for non-Gaussian fields.","lead":"This paper introduces a projection-based estimator for the Lipschitz-Killing curvatures that govern expected Euler characteristic curves of nonstationary Gaussian random fields. A bootstrap variant extends it to non-Gaussian data, enabling smooth EEC estimates and significance thresholds in brain imaging and cosmology.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3(ii) proof requires fourth derivatives of the empirical correlation, but (R2) only supplies derivatives up to order 2.","rationale":"I read the paper in good faith. The HPE for Gaussian fields is well supported: Theorem 2 gives unbiasedness and finite variance under (G1)-(G4), and the simulations confirm the main behavior. The fragile part is the bootstrap extension in Theorem 3. The reader identified the assumption that the Gaussian multiplier field inherits (G1)-(G4a) and the reliance on (R2); that is correct and important. My stress-test adds a more specific internal gap: even granting the multiplier field's regularity, the proof of Theorem 3(ii) silently requires convergence of fourth-order mixed derivatives of the sample correlation, while (R2) only provides second-order convergence. This is not an outside-consensus disagreement; it is a missing step in the stated proof. It affects the central claim because bHPE consistency is the paper's main vehicle for non-Gaussian and non-iid data, including the 3D fMRI application. The issue is likely repairable by strengthening (R2) to include derivatives up to order 4 and by stating explicit sufficient conditions for (G4a), so I do not recommend rejection or a changed verdict; the conditional verdict already given is appropriate. My concrete test would settle whether the missing fourth-derivative convergence is genuinely needed or can be avoided by a more careful continuity argument.","tokens_in":937,"tokens_out":1042,"duration_ms":133457,"concrete_test":"Independently re-derive the convergence statement in Supplementary B.4 using only (R2): list every partial derivative of rhat^{(N)} that enters the metric and curvature terms of L(rhat^{(N)}) via the formula in Adler and Taylor (2009, Thm 12.4.2). If the list contains any derivative of total order 3 or 4, construct a smooth sequence rhat^{(N)} with rhat^{(N)}(s,s)=1 whose first and second derivatives converge uniformly to those of r but whose fourth derivatives do not converge, and compute the LKC functional; if the LKCs fail to converge, the theorem requires an extra assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the bHPE consistently estimates LKCs of the limiting field is not established under the stated assumptions. In the proof of Theorem 3 (Supplementary B.4), the authors write that the Riemannian curvature tensor depends on E[partial_d partial_d' R_g^{(N)} partial_d'' partial_d''' R_g^{(N)}], which equals a fourth-order mixed partial derivative of the sample correlation rhat^{(N)}, and assert that this converges by (R2). But (R2) assumes convergence of rhat^{(N)} and its partial derivatives only up to order 2. Since the metric components are second derivatives of r and curvature is a second derivative of the metric, the LKC map L(r) is not a continuous function of only second derivatives of r in general. Thus the convergence L(rhat^{(N)}) -> L(r) is a further assumption, not a consequence of (R2). The theorem also simply assumes that the Gaussian multiplier field satisfies (G1)-(G4a) almost surely; no sufficient conditions on the residuals are given, so even the unbiasedness step for the bootstrap is conditional on an unverified regularity property. These gaps are fixable by strengthening (R2) to fourth-order uniform convergence and adding explicit conditions for (G1)-(G4a), but as written the bHPE consistency theorem is incomplete.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Hermite projection estimators (HPE) for the Lipschitz-Killing curvatures (LKCs) of nonstationary smooth Gaussian random fields, based on projecting the observed Euler characteristic (EC) curve onto the EC densities in a weighted L2 space. A multiplier bootstrap variant (bHPE) is introduced to estimate the LKCs of the Gaussian limiting field of non-Gaussian or non-iid data. The paper proves unbiasedness, finite variance, consistency, and CLTs for the HPE and the resulting plug-in EEC estimator, derives asymptotic confidence bands, and analyzes the estimated detection threshold. The methods are evaluated on simulated 2D fields and applied to CMB and fMRI data. The main results are Theorems 1-5, with proofs in the appendix and code provided online.","tokens_in":24696,"tokens_out":6689,"duration_ms":74159,"significance":"If the results hold, this is a valuable contribution to random field theory and its applications in neuroimaging and cosmology. It offers a computationally simple, theoretically grounded alternative to stationarity-based LKC estimators, and the multiplier bootstrap extension to non-Gaussian limiting fields is novel and practically important. The paper ships reproducible code, includes extensive simulations comparing against existing methods, and gives concrete, falsifiable predictions (e.g., unbiasedness of the HPE under (G1)-(G4a), consistency of the bHPE under (R1)-(R2)). The proofs of Theorems 1, 2, 4, and 5 are largely rigorous, and the simulation results convincingly support the finite-sample claims. However, the proof of the central bHPE consistency theorem (Theorem 3(ii)) has a gap, and the theorem relies on an unverified regularity assumption about the Gaussian multiplier field, so the full consistency claim is not yet established.","major_comments":[{"comment":"The proof of the convergence L(ˆr^{(N)}) → L(r) in Theorem 3(ii) requires convergence of the fourth-order mixed partial derivatives of the empirical correlation, because the Riemannian curvature tensor is expressed in B.4 in terms of E[∂d∂d' R_g^{(N)} ∂d''∂d''' R_g^{(N)}] = ∂^4_{s,s'} ˆr^{(N)}(s,s')|_{s=s'}. However, assumption (R2) only assumes uniform almost sure convergence of ˆr^{(N)} and its partial derivatives up to order 2. The sentence in B.4 asserting that this fourth-order quantity 'converges by (R2)' is therefore not justified. Since the LKC map is not, in general, a continuous function of only the second derivatives of the correlation, the convergence L(ˆr^{(N)}) → L(r) is an additional assumption rather than a consequence of (R2). This is a load-bearing gap for the main bHPE consistency claim. The authors should strengthen (R2) to include uniform convergence of derivatives up to order 4 (with appropriate domination) or prove continuity of the LKC functional under a weaker metric that only involves second derivatives.","section":"Section 3, Theorem 3(ii); Supplementary B.4"},{"comment":"The theorem assumes that, for almost all residuals R1,...,RN and all sufficiently large N, the Gaussian multiplier field R_g^{(N)} satisfies (G1)-(G4a). No sufficient conditions on the observed fields or residuals are provided under which this assumption holds. Since (G4) is itself admitted in Section 2 to be an active research topic, the bHPE consistency result is conditional on an unverified regularity property. The paper should either supply checkable conditions under which the GMF inherits (G1)-(G4a) from the residuals, or explicitly state Theorem 3 as a conditional result. As written, the abstract's claim that the bHPE 'allows estimation of LKCs ... of the limiting field' is stronger than what the theorem establishes.","section":"Section 3, Theorem 3"},{"comment":"The proof of Corollary 1 infers superpolynomial decay of Cov[χf(u), χf(v)] from the finiteness of the double integrals in (10) for all d,d'. Finiteness of all polynomial moments of an integrable function does not by itself imply pointwise decay faster than every polynomial unless additional regularity (e.g., monotonicity or continuity of the covariance kernel in its tails) is assumed. This does not affect the main theorems, but as written the corollary's proof is incomplete. The authors should either prove the decay under explicit additional conditions or reformulate the corollary as a statement about polynomial moments of the covariance kernel.","section":"Section 2, Corollary 1; Supplementary B.3"}],"minor_comments":[{"comment":"The phrase 'for almost all N > N′' is awkward because N is the fixed sample size in the theorem. It would be clearer to state 'for all N ≥ N0 almost surely' or 'for all sufficiently large N, almost surely over the residuals.'","section":"Section 3, Theorem 3 statement"},{"comment":"In the statement 'the covariances σdd′ are finite for all d,d′ = 1,...,N', the upper limit N appears to be a typo and should be D, the dimension of the manifold.","section":"Section 2, Theorem 2(ii)"},{"comment":"Assumption (R2) says that ˆr^{(N)}(s,s') and its partial derivatives up to order 2 converge uniformly, but the proof of Theorem 3(ii) uses fourth-order mixed partial derivatives. The notation should clarify whether derivatives are taken with respect to both arguments s and s', and the assumption should be stated in a form that makes the required order explicit.","section":"Section 3, assumption (R2)"},{"comment":"Theorem 4 is stated 'under the assumptions (G1)-(G4)', but part (i) only requires that the arbitrary estimator ˆL^{(N)} be consistent and part (ii) only requires the CLT for ˆL^{(N)}. The hypotheses should be stated directly in terms of the estimator ˆL^{(N)} rather than referencing (G1)-(G4), which are conditions on the underlying field and not on the estimator.","section":"Section 4, Theorem 4"}],"recommendation":"major_revision","confidential_remarks":"This is a promising paper with a useful methodology and strong empirical support. The main issue is the gap in the proof of Theorem 3(ii) and the unverified assumption about the Gaussian multiplier field satisfying (G1)-(G4a). Both are fixable by strengthening the assumptions and adding sufficient conditions, so I recommend major revision rather than rejection. The authors should also be asked to tighten the proof of Corollary 1 and to clarify the notation around (R2). The code release and the breadth of simulations and applications are definite strengths."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The HPE is a genuinely new estimation principle for LKCs: it exploits the orthogonality of the EC densities in a weighted L2 space to get a simple closed-form estimator, and it gives a clean theoretical treatment—unbiasedness, finite variance, CLT under explicit conditions. The connection to Adler et al.'s LKC regression as a continuous weighted version is nicely drawn, and the estimator is computationally simple. The bootstrap variant bHPE is a good idea for non-Gaussian fields, and the simulations (including the sphere and fMRI) suggest it behaves well in practice.\n\nThe HPE proofs are mostly solid. The proof of Corollary 1 is terse and the inference from finite integrals of all Hermite polynomial orders to rapid decay of the covariance is plausible but not fully spelled out; that's a minor issue.\n\nThe real problem is Theorem 3(ii). The proof asserts that the Riemannian curvature tensor converges via (R2), but curvature involves fourth derivatives of the correlation function, while (R2) only supplies uniform convergence of rhat and derivatives up to order 2. So L(rhat) -> L(r) is not a consequence of the stated assumptions. You need to strengthen (R2) to derivatives up to order four, or prove continuity of the LKC map in a weaker topology. On top of that, the assumption that the Gaussian multiplier field satisfies (G1)-(G4a) almost surely is put in by fiat; no sufficient conditions on the residuals are offered. These are fixable, but as written the bHPE consistency theorem is incomplete.\n\nThe bigger idea remains sound. The HPE deserves to be known, and the bHPE is likely fixable with modest changes. I'd send this to a serious referee. A referee should ask for the stronger smoothness assumption and for some checkable conditions on residuals that guarantee the GMF inherits (G1)-(G4a). The paper is aimed at people doing thresholding on nonstationary random fields, especially in neuroimaging and cosmology, and they will get real value from it even with the gap.","headline":"Solid new estimator for LKCs via Hermite projection, with a bootstrap extension that is promising but has a real proof gap in the consistency theorem that needs an extra assumption.","tokens_in":25165,"tokens_out":3311,"would_cite":true,"duration_ms":34243,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M40","60G60","62G09"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper shows that a Hermite projection of pinned Euler characteristic curves unbiasedly estimates the Lipschitz-Killing curvatures of nonstationary smooth Gaussian fields, and that a multiplier bootstrap extends this to the limiting…","keywords":["random field theory","Lipschitz-Killing curvatures","expected Euler characteristic curve","Gaussian related fields","Hermite projection estimator","multiplier bootstrap","nonstationary Gaussian fields","excursion sets"],"falsifier":"Simulate data from a non-Gaussian field with a known Gaussian limit but a spatial correlation whose empirical second derivatives converge non-uniformly, such as spatially varying smoothing; run the bHPE for increasing sample sizes and check whether the estimates approach the true limiting LKCs. If the estimates fail to approach despite marginal correlations looking correct, the uniform-convergence condition (R2) is doing the load-bearing work.","tokens_in":24245,"feed_emoji":"📐","tokens_out":9834,"duration_ms":97960,"temperature":0.7,"pith_summary":"The expected Euler characteristic curve of a smooth Gaussian field is a finite linear combination of known density functions, with coefficients given by the Lipschitz-Killing curvatures of the domain under the metric induced by the field. The paper proves that those coefficients can be estimated by projecting the observed, pinned Euler characteristic curve onto the density functions using a weighted Hermite inner product. Under smoothness and finite-moment conditions on the field's critical points, the resulting Hermite projection estimator is unbiased, has finite variance, and averages over repeated observations to a consistent, asymptotically normal estimate. A multiplier bootstrap version applies the same projection to Gaussian multiplier fields built from standardized residuals, consistently estimating the curvatures of the Gaussian limiting field of non-Gaussian or non-iid data. This matters because LKC estimation no longer needs stationarity or isotropy, so EEC-based thresholding can be used on arbitrary curved domains and on data such as brain images and cosmic microwave background maps.","feed_headline":"One projection gives unbiased field-topology estimates","feed_subtitle":"A Hermite-weighted integral of Euler curves recovers curvature coefficients without assuming stationarity.","key_machinery":"The workhorse is the Hermite projector on a weighted $L^2$ space with inner product $\\langle g,h\\rangle=\\int g(u)h(u)e^{u^2/2}\\,du$: the EC densities $\\rho_d(u)=(2\\pi)^{-(d+1)/2}H_{d-1}(u)e^{-u^2/2}$ are orthogonal in this inner product, so each LKC is the projection coefficient of the pinned EEC curve onto $\\rho_d$. The estimator evaluates this indefinite integral as a finite sum over the field's critical values, because the EC curve is constant between critical levels. The bootstrap extension constructs a Gaussian multiplier field $R_g^{(N)}=\\sum_{n=1}^N g_n R_n$ from standardized residuals $R_n$, which conditional on the data is a mean-zero Gaussian field with covariance equal to the empirical correlation; applying the projector to this field and letting the number of multiplier samples grow yields the bHPE.","core_discovery":"The central claim is that the Lipschitz-Killing curvatures are recoverable from a single observed Euler characteristic curve by linear projection: after pinning the curve by subtracting the known baseline $L_0\\Phi^+$, the $d$-th LKC is estimated by $\\hat L_d = (2\\pi)^{d/2}/(d-1)! \\int_{-\\infty}^{\\infty} H_{d-1}(u)\\,\\chi_f^\\circ(u)\\,du$. For Gaussian fields satisfying assumptions (G1)-(G4), this estimate is unbiased and has finite variance, and its sample average is consistent and asymptotically normal. For non-Gaussian fields satisfying a functional CLT, the bootstrap HPE applies the same projector to Gaussian multiplier fields generated from standardized residuals; the estimator equals the LKCs of the Gaussian field with the empirical correlation, and converges to the LKCs of the limiting field as the sample size grows. Plugging either LKC estimate into the Gaussian kinematic formula gives a smooth estimator of the EEC curve that inherits a functional CLT and yields confidence bands and consistent threshold estimates.","pith_inferences":["The same projection structure should extend to Gaussian-related fields covered by the Gaussian kinematic formula, such as chi-squared and t fields, by replacing the EC densities with the corresponding Gaussian-related densities; the paper notes the extension for plug-in EEC estimators but does not develop the HPE there.","The bHPE's consistency leans on uniform convergence of the empirical correlation and its derivatives, so in short samples the estimator is effectively estimating the LKCs of the empirical correlation; a practical diagnostic would be to compare bHPE outputs under different residual constructions on the same data.","The critical-value representation of the HPE suggests an online or streaming implementation: an algorithm that passes through the field once and records critical values and topology changes yields the estimator without storing the entire field, connecting naturally to streaming Euler-curve algorithms.","The CMB comparison points to a reusable diagnostic: comparing observed LKC estimates to simulation means can flag physical model mismatch, and the same logic could be applied to other cosmological maps to probe non-Gaussianity in the early universe."],"forward_implications":["LKC estimation no longer requires stationarity or isotropy: any domain where the Euler characteristic of excursion sets can be computed, including curved surfaces like the sphere or the cortical surface, becomes usable.","The smooth HPE of the EEC curve is an orthogonal projection of the nonparametric average and has lower variance; the paper's functional CLT gives both pointwise and simultaneous confidence bands for the true EEC curve.","The threshold solving $\\widehat{\\mathrm{EEC}}(u)=\\alpha$ is a consistent estimator of the excursion threshold and is asymptotically Gaussian, so significance levels for fMRI activation maps and the expected number of false clusters can be reported with standard errors.","For non-Gaussian observations, the bHPE estimates the LKCs of the Gaussian limiting field without knowing the field's mean or variance, extending the method to general linear models and standardized residual fields.","In the cosmic microwave background example, the observed field's second LKC lies about 2.93 standard deviations from the mean of the simulations, giving a quantitative topological comparison between observed and simulated sky maps."],"supporting_citations":[{"why":"Supplies the Gaussian kinematic formula, the definition of LKCs as intrinsic volumes, and the Morse-theoretic facts that make EC curves piecewise constant between critical values.","marker":"[Adler and Taylor, 2009]"},{"why":"Establishes validity of the EEC heuristic for high thresholds, motivating the use of the EEC curve for supremum and FWER inference.","marker":"[Taylor et al., 2005]"},{"why":"Introduces LKC regression by fitting observed EC curves to EC densities; the HPE is a continuous projection version of that regression.","marker":"[Adler et al., 2017]"},{"why":"Provides the warping-to-isotropy estimator, the main nonstationary competitor against which the HPE and bHPE are compared in simulations.","marker":"[Taylor and Worsley, 2007]"},{"why":"Supplies the stationary-isotropic roughness estimator used as a baseline in simulations and in neuroimaging.","marker":"[Kiebel et al., 1999]"},{"why":"Pioneered EEC-based FWER thresholding in neuroimaging, the inference setting the paper targets.","marker":"[Worsley et al., 1996]"},{"why":"Gives the condition for (G3) in terms of C3 sample paths and the multiplier-t construction used for simultaneous confidence bands.","marker":"[Telschow and Schwartzman, 2019]"},{"why":"Provides the full focal plane simulations used to estimate EEC curves of simulated cosmic microwave background fields on the sphere.","marker":"[Ade et al., 2016]"}],"fun_headline_variants":["One Euler curve recovers field curvatures","Topology without stationarity: consistent estimates","Bootstrap improves nonstationary field topology","Hermite-weighted integral estimates LKCs","Pinned curve yields efficient topology estimator"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The bootstrap estimator rests on the empirical correlation of the standardized residuals, and its derivatives up to order two, converging uniformly to the true correlation; if that convergence fails, the constructed Gaussian multiplier field does not have the right geometry and the LKC estimates no longer converge to the limiting field's LKCs.","fun_headline_variants_meta":{"raw":{"variants":["One Euler curve recovers field curvatures","Topology without stationarity: consistent estimates","Bootstrap improves nonstationary field topology","Hermite-weighted integral estimates LKCs","Pinned curve yields efficient topology estimator"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000304,"raw_usage":{"total_tokens":1762,"prompt_tokens":974,"completion_tokens":788,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":723}},"tokens_in":590,"tokens_out":788,"duration_ms":8610,"temperature":1.0,"reasoning_tokens":723,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:41:54.302409+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate data from a non-Gaussian field with a known Gaussian limit but a spatial correlation whose empirical second derivatives converge non-uniformly, such as spatially varying smoothing; run the bHPE for increasing sample sizes and check whether the estimates approach the true limiting LKCs. If the estimates fail to approach despite marginal correlations looking correct, the uniform-convergence condition (R2) is doing the load-bearing work.","supporting_citations":[],"review_version":1}