REVIEW 4 major objections 5 minor 2 references
A single convex estimator classifies each covariate effect in longitudinal models as zero, constant, or time-varying while estimating the nonzero curves smoothly.
Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →
TV-Select uses a combined group-sparsity and smoothness penalty to classify longitudinal covariate effects as zero, constant, or time-varying while estimating the varying curves.
T0 review reviewed 2026-08-03 challenge →
load-bearing objection The core idea is worth engaging, but the paper is not ready: the algorithm is not shown to solve the criterion the theory analyzes, and the main simulation table contains suspicious identical entries. the 4 major comments →
Group-Sparse Smoothing for Longitudinal Models with Time-Varying Coefficients
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
Core claim
The paper claims that classifying covariate effects in longitudinal varying-coefficient models reduces to a group-sparsity problem: after decomposing each coefficient function as β_k(t)=μ_k+g_k(t) with the deviation g_k centered to have zero integral, a covariate has a time-varying effect exactly when its spline coefficient block θ_k is nonzero. TV-Select minimizes a convex objective that combines a group penalty on each θ_k to zero out constant effects with a roughness penalty θ_k^T Ω θ_k to keep the estimated curves smooth. The paper proves near-optimal L2 estimation rates for the fitted curves, exact recovery of the set of time-varying predictors under irrepresentability and beta-min cond
What carries the argument
The central device is the decomposition β_k(t)=μ_k+g_k(t) with the constraint ∫_0^1 g_k(t)dt=0, together with a centered spline expansion of g_k. This makes 'time-varying versus constant' equivalent to 'nonzero versus zero deviation block,' so a group-sparsity penalty on the block selects the varying set, while a quadratic roughness penalty on the same coefficients controls curvature. The two penalties form a convex objective that is solved by cyclic block coordinate descent with a ridge-then-soft-threshold step.
Load-bearing premise
The exact-recovery guarantee rests on an incoherence condition that requires inactive deviation blocks to stay almost orthogonal to active ones after controlling for the constant effects; this condition is unverifiable from data and must hold for the partition to be recovered consistently.
What would settle it
Simulate a longitudinal dataset where one covariate has a genuine time-varying effect and several other covariates are generated as linear combinations of its centered spline basis, making the incoherence quantity in Theorem 2 exceed 1−η; if TV-Select still recovers the varying set, the condition is not necessary, and if it fails, the theory's reliance on the condition is confirmed.
If this is right
- Analysts can fit one model and read off which effects are constant and which vary, instead of running separate hypothesis tests.
- Once the time-varying set is identified, the estimated constant effects achieve oracle efficiency, making inference as sharp as if the structure were known in advance.
- Combining group sparsity with roughness yields smoother and more stable curves than either mechanism alone, improving prediction in the paper's simulations.
- Choosing the number of spline basis functions as n^{1/(2r+1)} gives the usual nonparametric error rate up to a log p factor.
Where Pith is reading between the lines
- The incoherence condition required for exact recovery is unlikely to be verifiable from data; in practice, sensitivity checks on tuning parameters could show whether the selected varying set is stable.
- The algorithm's ridge-then-threshold update is not proven to reach the global minimum of the convex objective, so the theory applies to the oracle minimizer and a gap may remain for the computed estimator.
- The decomposition could extend to non-Gaussian longitudinal outcomes or to adaptive weights that down-weight large deviations, which might improve finite-sample selection.
- The theory's prediction could be tested directly by constructing designs where the incoherence quantity exceeds 1−η and checking whether recovery failures match the theorem's conditions.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes TV-Select, a doubly penalized estimator for longitudinal varying-coefficient models. Each coefficient function is decomposed as a constant mean plus a centered B-spline deviation, and estimation minimizes convex least squares with a group-Lasso penalty on each deviation block and a roughness penalty on its curvature. The authors claim that a single convex criterion can recover the partition into zero, constant, and time-varying effects, and they state theoretical results: near-optimal prediction/estimation rates (Theorem 1), exact recovery of the time-varying set under incoherence and beta-min conditions (Theorem 2 and Corollary 1), and oracle asymptotic normality for constant effects after correct structure recovery (Theorem 3). Computation is performed by a block coordinate descent algorithm, and simulations plus a Sleep-EDF application are used to support the method.
Significance. If the main claims are correct, the paper offers a useful and conceptually clean unification of structure identification and functional estimation in longitudinal varying-coefficient models: one convex objective selects which effects vary over time and which are constant, while also controlling smoothness. The theoretical program is well aligned with existing group-Lasso and nonparametric-smoothing literature, and the simulation design is broad, covering correlation, heavy tails, heteroscedasticity, time-varying covariates, and weak signals. The paper is generally clearly written and states standard assumptions. However, the central guarantee currently has a serious algorithm–theory disconnect, and there are proof-level issues in the stated theory that must be resolved before the results can be taken as established.
major comments (4)
- [§3.1, Eq. (10)–(11)] Algorithm 1's update is not shown to solve the convex criterion (3), while Theorems 1–3 are stated for a global minimizer. For a fixed block, the ridge solve gives θ~k = (G_k + 2λ2Ω)^{-1}v_k and the update is θ_k = S(θ~k, λ1). A necessary KKT condition for θ_k to minimize the block subproblem is v_k = (G_k + 2λ2Ω)θ_k + λ1 u_k with u_k ∈ ∂||θ_k||_2. The soft-threshold identity implies v_k = (G_k + 2λ2Ω)θ_k + λ1(G_k + 2λ2Ω)u'_k for some u'_k ∈ ∂||θ_k||_2. Unless G_k + 2λ2Ω is a scalar multiple of the identity, (G_k + 2λ2Ω)u'_k is generally not in the subdifferential. Thus fixed points of Algorithm 1 need not satisfy the KKT conditions of (3), and the optional damping mentioned in §3.1 is not a convergence guarantee. The numerical results therefore may not describe the estimator analyzed in Theorems 1–3. This gap is load-bearing and needs either a convergence proof for Algorithm 1 to a mini
- [§4.2, Theorem 2, condition 1] The incoherence condition as printed is vacuous. It defines P_{X,ZS} as the projection onto the column space of [1n, X, ZS]. Then (I − P_{X,ZS})Z_S = 0 identically, so max_{k∉S} ||(1/n) Z_k^T (I − P_{X,ZS})Z_S (Z_S^T Z_S)^{-1}||_op = 0 ≤ 1−η for any η>0. The condition therefore imposes no restriction and cannot govern selection consistency. The proof in the appendix appears to use the expression as though it controlled leakage from active to inactive groups, which is inconsistent with the definition. The projector should presumably be onto [1n, X] only, or the condition should be written with (I − P_{[1n,X]})Z_S. As stated, Theorem 2 does not contain the substantive irrepresentability assumption claimed in the introduction.
- [Proof of Theorem 1 (Appendix, after Eq. (15))] The proof drops the λ2Δridge term by asserting Δridge ≤ 0 because “the ridge term is minimized at θhat.” This is not correct: Ω is positive semidefinite, so θhat^T Ω θhat ≥ 0, and Δridge = θ*_0^T Ω θ*_0 − θhat^T Ω θhat can be positive. For a nonzero true deviation g0k, θ*_0k^T Ω θ*_0k is typically positive and does not shrink to zero as q grows. Under Assumption 5, which permits fixed λ2 with λ2||Ω||_op = O(1), this introduces an unaccounted positive term of order O(1) in the basic inequality, which would destroy the claimed vanishing rates. Either the proof must bound this roughness-bias contribution, or λ2 must be required to tend to zero at a rate compatible with the stated error bounds. As written, Theorem 1 is not established for the stated generality.
- [§5.6, Table 1] In every scenario, the VC-Ridge and Group-Lasso rows report identical MSEµ, MSEµ,act, and MSPE values, e.g., Scenario A (100,5,100): 0.3603, 0.3707, 41.5937; Scenario E (200,8,300): 0.0064, 0.0212, 6.0830. Since Group-Lasso is defined with λ2 = 0 and VC-Ridge with λ1 = 0, exact equality across these metrics is not plausible unless the implementations coincide or the table was copied incorrectly. The text uses these columns to argue that sparsity and roughness penalties each contribute; the identical entries undermine that comparison and the tables need to be corrected or recomputed.
minor comments (5)
- [§3.1] Heading typo: “Update of a for time-varying block update” should be “Update of time-varying block update.”
- [§4.1, Assumption 4 vs. Abstract] The abstract says “full-column-rank unpenalized design,” but Assumption 4 is a restricted eigenvalue/compatibility condition, not full column rank. Please align the abstract with the actual assumption.
- [§5.6, text after Table 1] The text says TV-Select attains the smallest MSEµ,act “in nearly all settings,” but in Scenario E with (N,ni,p)=(100,5,100), VC-Ridge has MSEµ,act 0.0361 versus TV-Select 0.0406. Please qualify or resolve the discrepancy.
- [§6.1] Typo: “Roghtness” should be “Roughness.” Also, the real-data section reports 20 records from 10 subjects but does not state the number of folds in the subject-wise cross-validation or the block count; adding these details would improve reproducibility.
- [Figures 5–8] Figure 5 has empty captions, and Figures 6–8 are referenced but their content is not described in sufficient detail; the plots should be made self-contained or supplemented with captions explaining what is displayed.
Circularity Check
No circular reasoning: Theorems follow from stated assumptions; the only notable gap (algorithm-theory link) is a correctness concern, not circularity.
full rationale
I walked the derivation chain from the criterion (3) through Theorems 1–3 and Corollary 1. The proofs are conventional: Theorem 1 uses Assumptions 1–5, Lemma 1 on spline approximation, and Lemma 2 on noise–design correlation to bound estimation error; Theorem 2 adds an incoherence/irrepresentability condition and a beta-min condition and uses a primal–dual witness argument; Corollary 1 uses a threshold of order sqrt(log p / n) justified by a standard uniform fluctuation bound; Theorem 3 is a FWL-based oracle CLT after selection. None of these steps assumes the conclusion. The structural partition (Szero, Sconst, Svary) is defined from the decomposition beta_k(t)=mu_k+g_k(t) with centered g_k, and the estimator's selection rule is the group-Lasso block sparsity of theta_k; that is a modeling choice, not a fitted-input-called-prediction. Tuning parameters lambda1, lambda2 are selected by EBIC in simulations and the theoretical lambda1 scale is derived from the noise bound, not fitted to the results. The real-data analysis is predictive and does not back-fit constants. There are no author self-citations used as load-bearing evidence, no uniqueness theorem imported from the authors, and no ansatz smuggled in via citation. The manuscript itself contains a real algorithmic caveat: Algorithm 1's 'smoothing + selection' update is not proven to minimize (3), so the theorems for the global minimizer may not describe the computed estimator; the paper only offers an optional damped update to ensure monotone decrease. That is a statistical/computational correctness risk, not circularity, because the claimed results are about the convex minimizer and the assumptions do not include the desired recovery. Consistently with the reader's assessment, no circular step was identified; score 0.
Axiom & Free-Parameter Ledger
free parameters (5)
- λ1 (group penalty) =
data-dependent via EBIC grid
- λ2 (roughness penalty) =
data-dependent via EBIC grid
- q (number of B-spline basis functions) =
q ∈ {8,10,12} increasing with sample size
- τN (classification threshold) =
τN = sqrt(log p / n)
- γ in EBIC =
0.5
axioms (6)
- domain assumption Errors are sub-Gaussian and independent after de-meaning (working independence)
- domain assumption Covariates are sub-Gaussian and design satisfies a restricted eigenvalue / compatibility condition
- domain assumption True deviation functions are r > 2 Hölder smooth and spline approximation error decays as q^{-r}
- domain assumption Incoherence (irrepresentability) and beta-min conditions hold
- domain assumption p + s_v q = o(n) and log p = o(n)
- standard math Centered B-spline basis has L2 norm equivalent to Euclidean norm
Cite this review
Pith. "Pith review of Group-Sparse Smoothing for Longitudinal Models with Time-Varying Coefficients." pith.science (2026). https://pith.science/paper/LTAQK3WG
@misc{pith2026260307656,
author = {Pith},
title = {Pith review of: Group-Sparse Smoothing for Longitudinal Models with Time-Varying Coefficients},
year = {2026},
howpublished = {\url{https://pith.science/paper/LTAQK3WG}},
note = {Machine review of arXiv:2603.07656}
}
read the original abstract
Longitudinal associations may vary over time, yet allowing every regression effect to be dynamic can inflate estimation variance and obscure interpretable structure. We develop time-varying-effect selection (TV-Select), a group-sparse smoothing framework that classifies covariate effects as zero, constant, or time varying. Each coefficient is decomposed into a constant mean and a centered temporal deviation represented by a full-rank, L2-normalized effective spline basis. A group penalty identifies varying components, while a roughness penalty controls their curvature. The resulting convex criterion is solved by cyclic block proximal-gradient updates and followed by smooth refitting. Under a full-column-rank unpenalized design and an effective model dimension that is small relative to the total number of observations, we establish prediction and parameter rates, blockwise function-estimation bounds, and exact recovery of the varying set under irrepresentability and beta-min conditions. A stable classification refit further separates zero from constant effects. For fixed-dimensional contrasts, we construct an oracle-equivalent one-step estimator with cluster-robust asymptotic normality and consistent sandwich variance estimation. Simulations demonstrate that TV-Select combines low false-positive rates with accurate function estimation and competitive prediction across a range of longitudinal settings. An application to Sleep-EDF data produces smooth and parsimonious temporal effect estimates with essentially unchanged held-out predictive performance.
Figures
Reference graph
Works this paper leans on
-
[1]
and Verhasselt, A
Antoniadis, A., Gijbels, I. and Verhasselt, A. (2012). Variable selection in varying-coefficient models using P-splines. Journal of Computational and Graphical Statistics 21: 638–661. Bai, J. and Ng, S. (2002). Determining the number of factors in approximate factor models. Econometrica 70: 191–221. Bai, R., Boland, M. R. and Chen, Y. (2023). Scalable high-...
2012
-
[2705]
and Zeger, S
Liang, K.-Y. and Zeger, S. L. (1986). Longitudinal data analysis using generalized linear models. Biometrika 73: 13–22. Lin, Y. and Zhang, H. H. (2006). Component selection and smoothing in multivariate nonpara- metric regression. The Annals of Statistics 34: 2272–2297. Lounici, K., Pontil, M., Geer, S. van de and Tsybakov, A. B. (2011). Oracle inequaliti...
1986
This paper was first reviewed by deepseek-v4-flash on August 3, 2026.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.