{"id":"5b4edb16-0b93-4d4f-ad41-ea3b87177c87","arxiv_id":"2509.08155","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A three-manuscript thesis: nonlinear MI-based screening for neuroimaging, faster accelerated-gradient hyperparameters for nonconvex sparse learning, and a theoretically derived but unvalidated qGaussian linear mixed model.","lead":"Kai Yang's McGill PhD thesis bundles three statistical tools for high-dimensional biomedical data: a mutual-information screening package (fastHDMI) for nonlinear associations, tuned accelerated-gradient settings for nonconvex SCAD/MCP penalties, and a qGaussian linear mixed-effects model for correlated observations. Only the first two are backed by simulations; the third is presented as theory without experiments.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"M3's qGaussian mixed-effects model is internally inconsistent: Property 3 says sums of independent qGaussians are not qGaussian, so the y = Xβ + Zb + ε generative model cannot have the qGaussian likelihood (5.20).","rationale":"The reader's weakest-assumption analysis identifies exactly the same load-bearing concern: Manuscript 3's qGaussian mixed-effects model depends on the closure of the qGaussian family under linear combinations, which the manuscript itself refutes in Property 3. This is not a peripheral issue. If the marginal is not qGaussian, then the likelihood (5.20) describes no well-defined generative model of the form y = Xβ + Zb + ε, and the statistical interpretation of the estimator is ambiguous: it is either a quasi-likelihood whose target is unknown, or a marginal qGaussian model that has silently abandoned the mixed-effects story. Remark 8 further shows that the contribution over existing robust LMMs is reduced to a reparameterization of the t-distribution. Manuscripts 1 and 2 are less affected. Manuscript 1's consistency argument for plug-in MI estimation is questionable, but the central screening claim is supported by controlled simulations and real-data comparisons; the consistency argument is not load-bearing for those empirical claims. Manuscript 2's theoretical and simulation evidence is the strongest chapter, with explicit hyperparameters and reproducible code. Therefore, the overall conditional verdict remains appropriate: Manuscripts 1 and 2 are credible, while Manuscript 3 requires a substantive redefinition of the model and repositioning relative to the t-distribution LMM literature. The reader's verdict should not change, so I mark UNCHANGED.","tokens_in":56159,"tokens_out":4635,"duration_ms":49953,"concrete_test":"Analytically derive the marginal of y = Xβ + Zb + ε under the generative assumptions of Section 5.5, with b and ε independent qGaussian and q>1. It suffices to examine the scalar submodel S = b + ε, with b ~ qGaussian(q, σ_b^2) and ε ~ qGaussian(q, σ_e^2). Compute the characteristic function of S: it is the product of the characteristic functions of two Student-t (qGaussian) variables. Since the product of two such characteristic functions is not the characteristic function of any qGaussian (except in the Gaussian q→1 limit), the density of S cannot be of the form (5.20) for any (q', Σ'). This would settle the concern by showing the stated likelihood is not the true marginal of the stated model. A complementary simulation check: generate b and ε, estimate the density of S, and compare fits of (5.20) versus a general Student-t; a systematic lack of fit confirms the closure failure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of Manuscript 3 is a qGaussian linear mixed-effects model: observations y = Xβ + Zb + ε are assigned the qGaussian density (5.20), with a covariance inherited from random effects and error, and inference proceeds by maximizing this likelihood. The load-bearing condition is that the marginal distribution of y is qGaussian with the covariance appearing in (5.20). But Section 5.3's Property 3 states that independent qGaussian random vectors are not closed under linear combination: H1X1 + H2X2 is not qGaussian. In a mixed-effects model, y is exactly such a linear combination of the random effect b and the error ε. Therefore, unless a very special, unstated condition holds (e.g., one component is degenerate, or the model is defined only marginally rather than through a random-effects generative process), the density in (5.20) is not the true marginal likelihood of the stated model. If the model is instead interpreted as a working/quasi-likelihood, the target of estimation is not the true data-generating distribution, and the 'mixed-effects' interpretation must be abandoned. Remark 8 compounds the problem by conceding that for q>1, (5.20) is just a reparameterized multivariate t distribution, so the claimed robustness novelty over existing t-distribution LMMs is largely a reparameterization. This internal incoherence is the most load-bearing weakness: it affects the definition of the model, the interpretation of all estimates, and the coverage properties of any prediction intervals, not merely an implementation detail.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This thesis is organized as three manuscripts. Manuscript 1 presents fastHDMI, a Python package for mutual-information-based variable screening using FFTKDE, kNN, binning, and Pearson correlation; it claims that FFTKDE-based MI is best for continuous nonlinear outcomes, binning-based MI is best for certain nonlinear binary outcomes, and that the method is computationally practical on neuroimaging-scale ABIDE data. Manuscript 2 adapts the Ghadimi--Lan nonconvex accelerated gradient method to SCAD/MCP-penalized linear and logistic models, proposes a hyperparameter choice based on minimizing a complexity upper bound, proves an O(1/N) rate, and reports faster convergence and better signal recovery than ncvreg. Manuscript 3 re-derives the multivariate qGaussian density from Tsallis entropy maximization and proposes a qGaussian linear mixed-effects model with a proximal Hager--Zhang conjugate gradient algorithm. The thesis claims to deliver robust and efficient tools for high-dimensional screening, sparse estimation, and correlated-data modeling.","tokens_in":56370,"tokens_out":5172,"duration_ms":64070,"significance":"Each manuscript has standalone value. Manuscript 1 provides a working software tool and a rare neuroimaging-scale comparison of MI estimators; the simulations preserve the real ABIDE covariate structure, and the code is publicly available. Manuscript 2 is already published in Statistics and Computing, ships reproducible code, and contributes a practical hyperparameter recipe, an O(1/N) convergence rate, and a broad simulation study. Manuscript 3 contains a detailed algebraic derivation of the multivariate qGaussian density and an original framework for proximal conjugate gradient methods. However, the central qGaussian mixed-effects model is internally inconsistent with Property 3 of Section 5.3, and the hyperparameter optimality in Manuscript 2 is proven only in the convex case. These issues substantially temper the generality of the claimed contributions.","major_comments":[{"comment":"The central model is not internally consistent. Section 5.3 Property 3 states that independent qGaussian random vectors are not closed under linear combination, and Property 2 gives closure only under left multiplication by a fixed matrix. In the mixed-effects formulation y = Xβ + Zb + ε, the random effect b and error ε enter additively, so y is a linear combination of independent qGaussian vectors. Consequently the marginal density of y is generically not the qGaussian density (5.20), and maximizing (5.20) is not maximum likelihood for the stated generative model. The chapter must either provide a special condition under which the sum remains qGaussian, or abandon the random-effects generative interpretation and define (5.20) as a working/marginal likelihood. As written, the estimation, prediction, and any statistical guarantees attached to the 'mixed-effects' interpretation are not jus","section":"Manuscript 3, Sections 5.3 and 5.5"},{"comment":"The optimality of the proposed hyperparameters is established only for h ≡ 0 in Theorem 1, while the application is to SCAD/MCP composite objectives with nonconvex smooth h given by (4.3)--(4.4). Conditions (4.8)--(4.9) guarantee convergence for these h, but minimizing the complexity bound in the convex specialization does not by itself minimize the nonconvex bound. The manuscript should state this limitation explicitly and either prove an analogue for nonzero h or characterize the nonconvex use as a heuristic supported by the reported simulations. As written, the theoretical 'optimal damping sequence' claim is convex-only, while the abstract's 'considerably faster' claim is empirical.","section":"Manuscript 2, Theorem 1 and Section 4.3.2"},{"comment":"Remark 8 concedes that for q > 1, the density (5.20) is exactly a reparameterized multivariate t density. Combined with the closure failure described above, the claimed robustness contribution of a 'qGaussian mixed-effects model' reduces, in substance, to a t-likelihood with a particular covariance parameterization. The chapter should compare explicitly with existing t-distribution LMMs and identify what is new beyond the reparameterization and the proximal Hager--Zhang algorithm; otherwise the novelty claim is overstated.","section":"Manuscript 3, Remark 8 and Section 5.5"}],"minor_comments":[{"comment":"The legend labels are inconsistent: Figure 3.1 uses 'MI based on kNN' while Figure 3.2 uses 'MI based on sklearn'. Please use consistent method names.","section":"Manuscript 1, Figures 3.1 and 3.2"},{"comment":"The input specification 'ρ ∈ (0, ρ^{-1})' appears to be a typo; from Lemmas 11--12 it should be 'ρ ∈ (0, L_{∇g}^{-1})'.","section":"Manuscript 3, Algorithm 3"},{"comment":"The consistency claim for the MI estimators is made by invoking the continuous mapping theorem, but the required functional continuity and uniform convergence conditions are not stated. This is too informal for a theoretical claim; either provide a precise statement with conditions or cite a source.","section":"Manuscript 1, Section 3.2"},{"comment":"The notation 'qtrain' is used in the optimization subsection but is not explicitly defined in relation to the parameter q in (5.20). Clarify whether qtrain is the shape parameter q or a training-set-specific estimate.","section":"Manuscript 3, Section 5.5"},{"comment":"There are several typographical slips, e.g., 'prepossessed' for 'preprocessed' in Manuscript 1, and 'Baghwat' inconsistently spelled in the author list. A careful proofread is needed.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main gate for this thesis is Manuscript 3: its stated qGaussian linear mixed-effects model is internally inconsistent, and the revision needs to either supply a closure condition or reframe the model as a marginal/working likelihood. Manuscript 2 is in better shape but should also correct the convex-only scope of the theoretical optimality claim. If the authors can make these changes, the remaining contributions are publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nThe thesis is a mixed bag. The second manuscript is the real contribution: explicit hyperparameter formulas (4.13)-(4.15) for the Ghadimi-Lan nonconvex accelerated gradient method and an O(log k) bound for the optimal damping sequence (Corollary 4) that I haven't seen in the cited literature. The simulations support faster convergence and competitive signal recovery. That part deserves a serious referee and is citable.\n\nManuscript 1 is a useful Python package (fastHDMI) but the framing has problems. The \"inaugural application\" claim is contradicted by the paper's own citations of MI-based feature selection in EEG/fMRI, and the consistency argument in Section 3.2—using the continuous mapping theorem on a \"continuous\" mutual information functional—is mathematically untenable as written. The package itself and the ABIDE case studies seem credible; the novelty claims need to be scaled back.\n\nManuscript 3 has a load-bearing flaw. Section 5.3's Property 3 states that independent qGaussian vectors are not closed under linear combination. In a mixed-effects model y = Xβ + Zb + ε, the observation is exactly such a linear combination, so the qGaussian density (5.20) cannot be the true marginal likelihood of the stated generative model. The paper's own Remark 8 concedes that for q>1 this is just a reparameterized multivariate t distribution, so the novelty over existing robust/t LMMs is overstated. This isn't a modest gap; it affects the definition of the estimand and any interval estimates. Either the mixed-effects story goes or the likelihood does.\n\nNet: one solid manuscript, one credible tool with repairable claims, and one chapter that needs a substantive rethink. I'd still send the thesis to peer review rather than desk reject—a good referee can separate the parts that survive. For a reading group, the M3 inconsistency is worth discussing, but I wouldn't build on it as is.\n\nBest.","headline":"Second manuscript is a real contribution; third has a load-bearing closure failure; first is useful software with weak framing.","tokens_in":57048,"tokens_out":2358,"would_cite":true,"duration_ms":26903,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H20","62J07","90C26","62F35","62-08"],"pacs":[],"model":"deepseek-v4-flash","headline":"Three advances: FFT mutual-information screening, an O(1/k) accelerated-gradient rule for SCAD/MCP, and a qGaussian mixed-effects model.","keywords":["high-dimensional data","variable screening","mutual information","nonconvex sparse learning","accelerated gradient","qGaussian distribution","linear mixed-effects model","Tsallis entropy"],"falsifier":"Generate data from a genuine qGaussian linear mixed model: draw the random effects and noise separately from qGaussian distributions, form their sum, and compare the empirical marginal density of that sum with density (5.20) at the fitted q and Sigma; a material mismatch, which Property 3 predicts, would show the proposed likelihood is not the marginal of that generative model.","tokens_in":55864,"feed_emoji":"🧠","tokens_out":12420,"duration_ms":139751,"temperature":0.7,"pith_summary":"Across three linked manuscripts, this thesis tries to establish that three standard bottlenecks in high-dimensional biostatistics—nonlinear variable screening, optimization of nonconvex penalties, and robust likelihoods for correlated observations—each have a computationally grounded solution. The first is variable screening built on mutual information estimated through fast Fourier transform kernel density estimation; simulations on a large neuroimaging dataset show it catches quadratic and other nonlinear associations that Pearson correlation misses, at practical running times. The second is a nonconvex accelerated gradient method with explicit hyperparameters chosen by minimizing the algorithm's complexity upper bound, giving an O(1/k) convergence rate and, in simulations, better signal recovery than the coordinate-descent baseline. The third is a qGaussian linear mixed-effects model derived from Tsallis entropy maximization, paired with a proximal conjugate-gradient solver. If these claims hold, analysts can screen nonlinearly, fit sparse models, and remain heavy-tail aware in one workflow.","feed_headline":"Mutual information screening scales to 150,000 brain features","feed_subtitle":"Nonlinear screening beats Pearson, then a fast sparse solver and a heavy-tailed mixed model take over.","key_machinery":"Three mechanisms carry the argument. The first is the FFT-based kernel density estimator: replacing direct kernel density estimation with Silverman's fast Fourier convolution turns mutual-information estimation into an O(n log n) operation, which is what makes screening 150,000 features practical. The second is the complexity-bound hyperparameter rule for the nonconvex accelerated gradient method: the recursion alpha_{k+1} = 2/(1 + sqrt(1 + 4/alpha_k^2)) with delta1 = omega_k = 2/(3L_Psi) minimizes the algorithm's upper bound and keeps the worst-case rate O(1/k). The third is the qGaussian density (5.20), the unique maximizer of Tsallis entropy under q-moment constraints; it supplies a bell-","core_discovery":"On the paper's own terms, the discoveries are: (i) mutual information computed from FFT-accelerated kernel density estimates is the leading screening statistic for continuous outcomes with nonlinear dependence, while binning-based mutual information leads for binary outcomes whose probability transform is nonlinear; on the preprocessed ABIDE fMRI data, screening roughly 150,000 features completes in practical time and the selected variables improve downstream prediction. (ii) For SCAD- and MCP-penalized linear and logistic models, the accelerated gradient algorithm becomes practically usable when its hyperparameters are set to delta1 = omega_k = omega = 2/(3L_Psi), delta_{k+1} = omega/alpha_","pith_inferences":["Beyond the paper: the qGaussian mixed-effects model is most defensible when read as a quasi-likelihood or as a reparameterized t-distribution model, because the paper's own property that independent qGaussian vectors are not closed under linear combination means a qGaussian random effect plus qGaussian error is not generically qGaussian.","Beyond the paper: the complexity-bound hyperparameter rule likely transfers to other nonconvex composite objectives, including penalized generalized linear mixed models, since the derivation depends on smoothness constants rather than on the specific forms of SCAD and MCP.","Beyond the paper: the mutual-information screening ranking could be piped directly into the qGaussian sparse model to form one end-to-end pipeline; the thesis describes the two tools as sequential stages but does not benchmark them as a single combined procedure."],"forward_implications":["Mutual-information screening can be used as a first pass on neuroimaging-scale data, preserving nonlinear signals that Pearson correlation would miss before downstream penalized models are fit.","The hyperparameter rule (4.13)-(4.15) gives practitioners a default recipe for nonconvex accelerated gradient, with a worst-case O(1/k) guarantee, so SCAD/MCP sparse estimation no longer requires per-problem tuning.","The proximal conjugate-gradient framework extends a class of numerical methods for finding equilibria of Lipschitz flows to composite nonsmooth objectives, so any globally convergent smooth solver with a line search can be given a proximal version.","The qGaussian likelihood broadens linear mixed-effects modeling to heavy-tailed data; in simulations it recovers signals where the coordinate-descent baseline fails to converge within practical time limits."],"supporting_citations":[{"why":"supplies the preprocessed ABIDE dataset that anchors Manuscript 1's simulations and case studies.","marker":"[Cameron et al., 2013]"},{"why":"introduces the FFT-based kernel density estimation that makes mutual-information screening computationally tractable at scale.","marker":"[Silverman, 1982]"},{"why":"provides the bin-count rule used by the binning mutual-information estimator.","marker":"[Birgé and Rozenholc, 2006]"},{"why":"introduces the SCAD penalty and oracle property that Manuscript 2's optimizer targets.","marker":"[Fan and Li, 2001]"},{"why":"introduces the MCP penalty, the second target problem for the accelerated gradient method.","marker":"[Zhang et al., 2010]"},{"why":"defines the coordinate-descent baseline against which signal recovery and runtime are compared.","marker":"[Breheny and Huang, 2011]"},{"why":"provides the nonconvex accelerated gradient framework whose complexity upper bound Manuscript 2 minimizes.","marker":"[Ghadimi and Lan, 2015]"},{"why":"defines the Tsallis entropy whose maximization yields the qGaussian density.","marker":"[Tsallis, 1988]"},{"why":"supplies uniqueness and duality properties of qGaussian vectors used in the density re-derivation.","marker":"[Vignat et al., 2004]"},{"why":"provides the conjugate gradient method that Manuscript 3 turns into a proximal algorithm.","marker":"[Hager and Zhang, 2005]"}],"fun_headline_variants":["Mutual info screening scales to 150K brain features","Nonlinear screening beats Pearson on 150K brain features","Fast nonlinear screening, sparse solvers, and Tsallis robustness","MI screening, nonconvex penalties, heavy-tailed models for high-dim data","Accelerated sparse solvers and robust mixed models for big data"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise of the third contribution is that a qGaussian random effect plus qGaussian error can be represented by the qGaussian likelihood (5.20); the paper's own closure property—independent qGaussian vectors are not qGaussian under linear combination—means this premise holds only in special cases.","fun_headline_variants_meta":{"raw":{"variants":["Mutual info screening scales to 150K brain features","Nonlinear screening beats Pearson on 150K brain features","Fast nonlinear screening, sparse solvers, and Tsallis robustness","MI screening, nonconvex penalties, heavy-tailed models for high-dim data","Accelerated sparse solvers and robust mixed models for big data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00077,"raw_usage":{"total_tokens":3260,"prompt_tokens":770,"completion_tokens":2490,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":2402}},"tokens_in":514,"tokens_out":2490,"duration_ms":18297,"temperature":1.0,"reasoning_tokens":2402,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T21:11:49.691371+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate data from a genuine qGaussian linear mixed model: draw the random effects and noise separately from qGaussian distributions, form their sum, and compare the empirical marginal density of that sum with density (5.20) at the fitted q and Sigma; a material mismatch, which Property 3 predicts, would show the proposed likelihood is not the marginal of that generative model.","supporting_citations":[],"review_version":1}