{"id":"cba9e64a-bbbe-4364-b74f-47a8a2c549dd","arxiv_id":"2506.09451","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A doubly dynamic safe screening rule for Group SLOPE that discards inactive feature groups during training, reducing computation without changing the final solution.","lead":"The paper proposes a safe screening rule for Group SLOPE that removes groups guaranteed to have zero coefficients during training, aiming to speed up proximal gradient solvers. It claims the first such rule for this model, with theoretical safety guarantees and experiments showing 2.5x to 14x speedups.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The screening rule is not implementable as specified: Algorithms 1 and 2 never construct the feasible dual iterate θ_k that Eq. (22) and Corollary 1 require, so the claimed safe integration with APGD/SPGD is unsupported.","rationale":"The reader's verdict of REJECT is justified, but the most load-bearing gap is not exactly the one identified as the weakest assumption. The QR decoupling concern is real: the statement that U_i has |I_i| orthogonal columns is false when a group has more variables than samples, and the paper never states the needed full-column-rank condition. However, with a rank-r QR the objective equivalence can be repaired for group effects, so that issue is a serious but potentially fixable presentation gap. The more direct blocker is that the screening rule itself cannot be executed: neither Algorithm 1 nor Algorithm 2 provides a dual feasible iterate or a duality gap, yet Eq. (22) and Corollary 1 require both. This means the central claim that the rule 'can be seamlessly integrated into existing solvers' is unsupported as written. The paper's empirical results are plausible and the idea is worthwhile, but the theoretical guarantee and the algorithmic specification need substantial revision before the claim can be accepted. I therefore concur with rejection, while noting that the fix may be straightforward if the authors supply a valid dual-update or projection procedure.","tokens_in":13197,"tokens_out":24405,"duration_ms":298379,"concrete_test":"Implement Algorithm 1 exactly as pseudocoded on a small Group SLOPE instance (e.g., n = 20, m = 10 groups, each of size 5) and record all quantities available at line 3; observe that no θ or duality gap exists. Then test the natural candidate θ = y − X̂b_k: check whether it satisfies the dual feasibility constraints defining Δ. If it does not, plugging it into Eq. (22) can produce an unsafe screen, and a correct implementation would require a projection onto Δ that the paper never specifies. Compare the screening decisions of such a corrected implementation against a brute-force KKT verification to determine whether the omission is material.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim depends on evaluating the screening condition (22), which uses a dual iterate θ and the duality gap G(b, θ). Corollary 1 bounds ||θ − θ*|| by sqrt(2G(b, θ)) only when θ is feasible for the dual constraint set Δ. However, Algorithms 1 and 2 contain no dual variable: their inputs and loops update only primal iterates b_k, ˆb_k, and t_k, and the step 'Apply the safe screening from (22)' has no θ_k or G(b_k, θ_k) to plug in. The assertion in Section 2.3 that θ is 'directly available during each iteration of standard proximal gradient methods' is not true for the vanilla APGD/SPGD algorithms cited in Section 3.1; the natural residual y − X̂b_k is generally not feasible for Δ. Consequently, Property 1 and Theorem 1 do not connect to the proposed algorithms as written. A secondary issue is the QR decoupling: the paper states U_i has |I_i| orthogonal columns, which is impossible when a group contains more variables than samples; a rank-r QR would be needed, and the correspondence between zero transformed groups and zero original coefficients only holds under full column rank.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a safe screening rule for Group SLOPE, a sparse regression model whose non-separable group penalty has previously blocked screening-based acceleration. The authors introduce a transformation intended to decouple the group-effect penalty, derive a dual problem and a screening test based on the duality gap, and claim that the resulting rule can be integrated into APGD and SPGD solvers while preserving exact solutions. They provide a finite-time screening theorem and report large speedups on four datasets. The core claims are that the screening rule is safe, that it provably discards every inactive group in finitely many iterations, and that it is seamlessly implementable in existing first-order solvers.","tokens_in":13470,"tokens_out":6212,"duration_ms":63302,"significance":"If the claims were correct, this would be the first safe screening rule for Group SLOPE and would meaningfully extend the line of gap-based safe screening methods to a model with ordered, non-separable group penalties. The paper also attempts a doubly dynamic rule that tightens both sides of the screening test over iterations. However, the central derivation contains a fundamental error in the dual reformulation, the proposed algorithms do not construct the dual quantities required by the screening test, and the paper's own experimental design violates the rank assumption behind the decoupling transformation. These issues invalidate the main theoretical and algorithmic claims as stated, so the contribution is not currently usable.","major_comments":[{"comment":"The decoupling step assumes that each group submatrix X_Ii admits a QR factorization X_Ii = U_i R_i with U_i having |I_i| orthogonal columns and R_i being full-row rank. This requires rank(X_Ii) = |I_i| and |I_i| <= n, which is not stated in the paper and fails for collinear columns or for groups larger than the sample size. The equivalence between problems (1) and (5) is therefore not established in general. This is not merely a technical gap: Section 4.1 says that each feature group is formed by duplicating each feature, so any group with |I_i| > 1 consists of identical columns and has rank 1, making the transformation in Eq. (2) invalid for the paper's own experimental setup.","section":"Section 2.1, Eqs. (2)-(5)"},{"comment":"The dual derivation replaces the ordered group penalty J_lambda(||b||_I) with the separable term sum_i lambda_i ||b_{I[i]}||_2. These are not equal: J_lambda sorts the group norms and applies the lambda_i to the sorted values, whereas the separable sum assigns each lambda_i to a fixed group. Consequently, the minimization over b in Eq. (9) is not the correct infimal convolution for the Group SLOPE penalty, and the dual feasible set Delta together with the optimality conditions (10)-(14) do not follow from the primal problem (6). The screening condition (14) is therefore derived from an incorrect dual problem.","section":"Section 2.2, Eqs. (7)-(9)"},{"comment":"Algorithm 1 and Algorithm 2 evaluate the screening test (22) with the current active-set size |A|, i.e., with threshold lambda_{|A_k|}. However, the proof of Theorem 1 establishes screening with respect to the fixed threshold lambda_{|A*|} of the final active set. The paper asserts that lambda_{|A_k|} rises to lambda_{|A*|} as the active set shrinks, but no proof is given that the running threshold eventually reaches the final threshold. The step in Section 2.4 that assigns the m - m_k zero-valued groups 'any permutation' of the smallest lambda values is also not justified by the sorted structure of the penalty. Thus the finite-time screening guarantee for the algorithm actually implemented does not follow from Theorem 1.","section":"Section 2.4 and Theorem 1"},{"comment":"Corollary 1 bounds ||theta - theta*|| by sqrt(2G(b, theta)) only for a dual-feasible point theta, and the screening test (22) requires a well-defined theta and duality gap G(b, theta). However, Algorithms 1 and 2 maintain only primal iterates b_k, hat-b_k, and t_k; they never construct a dual-feasible theta_k. The statement in Section 2.3 that theta is 'directly available during each iteration of standard proximal gradient methods' is not correct for the APGD and SPGD methods cited in Section 3.1, and the natural residual y - X hat-b_k is generally not feasible for Delta. Consequently, the screening rule as specified cannot be evaluated inside these algorithms, and Property 1 and Theorem 1 are not connected to the proposed procedures.","section":"Section 3.1 and Section 2.3"}],"minor_comments":[{"comment":"Equation (19) states ||theta - theta*|| <= sqrt(D(theta*) - D(theta)), but the preceding inequality (16) gives ||theta - theta*||^2 <= 2(D(theta*) - D(theta)), so the missing factor sqrt(2) should appear already in Eq. (19).","section":"Section 2.3, Corollary 1 proof"},{"comment":"The notation ||XIi||_2 is used for the spectral norm of the matrix X_Ii, while ||X_Ii^T theta||_2 is an Euclidean norm; the two norms should be distinguished explicitly.","section":"Section 2.1, Eq. (15)"},{"comment":"The symbol beta is used both for the primal coefficient vector and as the optimization variable in the inner product beta^T X^T theta inside the dual derivation, which is confusing and should be renamed.","section":"Section 2.2, Eqs. (7)-(9)"},{"comment":"The sentence 'The remaining m - m_k groups can then be assigned any permutation of lambda_{m_k+1}, ..., lambda_m' is vague; a formal statement about which groups receive which lambda values would clarify why the threshold lambda_{|A|} is valid at intermediate iterations.","section":"Section 2.4"},{"comment":"Properties 1 and 2 are stated without proof; if they are intended to follow from Theorem 1, this should be stated explicitly, otherwise they need their own arguments.","section":"Section 3, Properties 1 and 2"}],"recommendation":"reject","confidential_remarks":"The paper's core construction appears unsalvageable within the submitted scope: the dual derivation is invalid, the algorithms lack a required dual iterate, and the experiments construct rank-deficient groups that violate the paper's own decoupling assumption. I would expect a substantially rewritten manuscript with a correct dual derivation and a genuinely implementable screening test before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First safe screening rule for Group SLOPE is a genuine extension, but the paper as written has two load-bearing gaps that need patching: the algorithms never construct a feasible dual point to plug into the screening test, and the finite-screening proof uses the true active-set threshold instead of the running one. The QR decoupling, meanwhile, silently assumes full column rank per group.\n\nThe new part is real: the group-effect norm makes the screening non-trivial, and the doubly dynamic lower bound (rising lambda_{|A|}) is a nice addition over your earlier Group OWL rule. The speedups (3-14x batch, 2.5-8x stochastic) are plausible and match what gap-based screening usually achieves. Credit where due: the connection to [4] is explicit, so there's no circularity.\n\nThe soft spots are in proportion:\n\n- Eq. (2) states U_i has |I_i| orthogonal columns. That requires X_{I_i} to have full column rank, i.e., |I_i| <= n. In the high-dimensional regime the paper targets, groups can be larger than n. The equivalence between zero groups in the transformed and original problems only holds under full rank. Needs a stated assumption or a rank-revealing treatment.\n- Algorithms 1 and 2 say \"apply the safe screening from (22)\" but never define theta_k. The screening test needs a feasible dual iterate and a duality gap. Vanilla residual y - X b_k is generally not in the dual constraint set. Standard gap screening constructs one by scaling the residual. As written, the algorithms are not implementable.\n- Theorem 1 shows LHS eventually falls below lambda_{|A*|}, but the algorithm uses lambda_{|A_k|} with |A_k| >= |A*|, so the running threshold is lower. The finite-screening guarantee doesn't follow for the actual algorithm. This is the main theoretical gap.\n- Experiments: no code, no error bars; runtime averaged over 5 trials. Minor given the paper's clarity otherwise.\n\nOverall, this is a fixable paper. The core idea is sound; the presentation misses scaffolding that exists in the gap-safe-screening literature, and the proof needs a monotonic argument to bridge the threshold gap. I'd like to see a serious referee engage with it, because the extension is new and the experiments show enough promise.","headline":"A legitimate extension of safe screening to Group SLOPE, but the algorithms omit the dual iterate and the finite-screening proof doesn't match the running threshold; fixable, worth a serious look.","tokens_in":13964,"tokens_out":4797,"would_cite":false,"duration_ms":48813,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62J07","90C25","65K05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper introduces the first safe screening rule for Group SLOPE, provably discarding inactive feature groups during training so batch and stochastic solvers return the same solution much faster.","keywords":["safe screening rules","Group SLOPE","group feature selection","duality gap","proximal gradient descent","stochastic optimization","high-dimensional regression","ordered weighted l1 norm"],"falsifier":"Construct a design where one group has more features than samples or contains linearly dependent columns, solve the original Group SLOPE problem and the transformed problem in Eq. (5) on the same data, and compare their solutions; any difference in the objective value or in which groups have zero coefficients would show the equivalence and the safety guarantee do not hold in that setting.","tokens_in":13001,"feed_emoji":"⚡","tokens_out":9392,"duration_ms":86027,"temperature":0.7,"pith_summary":"Group SLOPE selects groups of predictors adaptively, but its block-nonseparable penalty makes existing solvers slow and memory-hungry in high dimensions. This paper claims to give the first safe screening rule for the model: during optimization, groups whose coefficients are provably zero are detected and removed, so the solver works only on the active groups. The rule is dynamic in two directions, using the duality gap to tighten an upper bound and an iterative active-set refinement to raise a lower bound, which lets it plug into batch and stochastic proximal gradient solvers without changing their outputs. If correct, Group SLOPE becomes far more practical on large sparse datasets, with reported speedups of 3--14$\\times$ in batch settings and 2.5--8$\\times$ in stochastic settings, plus lower memory use.","feed_headline":"Safe screening speeds Group SLOPE up to 14x","feed_subtitle":"Provably inactive feature groups are discarded during training, preserving the exact solution while cutting compute and memory.","key_machinery":"The central object is the decoupling transformation $X_{I_i} = U_i R_i$, which converts the block-nonseparable group-effect penalty $J_\\lambda(W\\llbracket X\\beta\\rrbracket_I)$ into $J_\\lambda(\\llbracket b\\rrbracket_I)$ on transformed variables, together with the screening inequality $\\|X_{I_i}^T\\theta\\|_2 + \\|X_{I_i}\\|_2\\sqrt{2G(b,\\theta)} < \\lambda_{|A|}$, which guarantees $b^*_{I_i}=0$. The left-hand side tightens as the duality gap $G(b,\\theta)$ shrinks along the optimization trajectory; the right-hand side rises as the active set is iteratively refined and the smallest unused $\\lambda$ weights are assigned to screened groups. This pair of moving bounds is what makes the rule doubly dynamic, and it carries the proof of safeness and finite-time screening.","core_discovery":"The paper's central claim is that Group SLOPE can be equivalently rewritten by factorizing each group's design submatrix, $X_{I_i} = U_i R_i$, so the penalty acts on decoupled variables $\\eta_{I_i} = R_i \\beta_{I_i}$ with group norms $\\|\\eta_{I_i}\\|_2$. In this form, the dual optimality condition yields a screening test: if $\\|X_{I_i}^T\\theta^*\\|_2 < \\lambda_{|A^*|}$, then group $i$ is inactive, meaning $b^*_{I_i} = 0$. Since $\\theta^*$ and $|A^*|$ are unknown during training, the paper derives computable bounds: an upper bound on $\\|X_{I_i}^T\\theta^*\\|_2$ from the duality gap using strong concavity of the dual, and a lower bound on $\\lambda_{|A^*|}$ by iteratively shrinking the active set and assigning the smallest $\\lambda$ values to screened groups. This doubly dynamic screening region expands as training proceeds, and Theorem 1 asserts every truly inactive group is screened out in finitely many iterations. The paper concludes the rule is safe for any iterative solver that maintains primal--dual iterates, and it demonstrates the resulting speedups on four benchmark datasets without accuracy loss.","pith_inferences":["The equivalence step requires each $X_{I_i}$ to have full column rank, so a natural extension is a rank-revealing factorization that handles rank-deficient or oversized groups, at the cost of redefining groups or accepting a different penalty.","In genuine mini-batch stochastic training, the duality gap entering the screening test would typically be estimated from batches; a practical extension is to evaluate the gap only periodically or on a full pass, since the strong-concavity bound is stated for the true dual objective.","The iterative assignment of the smallest unused $\\lambda$ values to screened groups is an ordering heuristic justified by active-set stabilization; a testable extension is to compare alternative orderings on problems where the active set does not shrink monotonically.","Because the construction uses only convexity and the ordered-weight structure, the same doubly dynamic screening idea could plausibly be adapted to other ordered weighted $\\ell^1$ group models, though the paper does not claim this."],"forward_implications":["Batch APGD solvers for Group SLOPE can run 3--14$\\times$ faster on $n \\ll d$ datasets, and stochastic SPGD solvers 2.5--8$\\times$ faster on large-scale datasets, while returning the original Group SLOPE solution.","Memory usage drops because only active groups are stored and updated after screening, and in sparse high-dimensional settings the active dimension $d_k$ is much smaller than $d$.","If the solver already computes the duality gap for its stopping criterion, the screening test adds only $O(d_k)$ overhead per iteration.","Because Group Lasso and SLOPE are special cases of Group SLOPE, the same rule supplies safe screening for those models as well.","Theorem 1 implies the active set stabilizes after finitely many iterations, so subsequent training runs on a much smaller problem with the same final solution."],"supporting_citations":[{"why":"Defines Group SLOPE and supplies the accelerated proximal gradient solver that the screening rule is designed to accelerate.","marker":"[10]"},{"why":"Introduces duality-gap-based dynamic safe screening for Lasso, the technique the paper adapts to the Group SLOPE penalty.","marker":"[13]"},{"why":"Provides the strong-concavity result for the dual that yields the bound $\\|\\theta-\\theta^*\\|_2 \\le \\sqrt{2G(b,\\theta)}$.","marker":"[22]"},{"why":"Establishes dynamic safe screening for Group Lasso, the group-structured precedent that this work extends to non-separable penalties.","marker":"[9]"},{"why":"Supplies the proximal stochastic gradient method with progressive variance reduction used as the SPGD baseline and vehicle for the stochastic screening experiments.","marker":"[39]"},{"why":"Provides the variance-reduction estimator that underlies the stochastic gradient updates in the SPGD algorithm.","marker":"[21]"},{"why":"Introduces the general idea of safe feature elimination in sparse supervised learning on which the screening guarantee rests.","marker":"[25]"}],"fun_headline_variants":["Safe screening rules speed Group SLOPE 14x","Provably safe screening prunes Group SLOPE zero groups","Group SLOPE screening cuts compute without accuracy loss","Screening rule slashes Group SLOPE memory and time","Safe screening discards inactive groups in Group SLOPE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument's load-bearing premise is that each feature group's design submatrix can be factored into as many orthogonal columns as the group has variables, which requires full column rank and group size no larger than the sample size; the paper does not state or verify this, and if it fails, the transformed problem is not equivalent to the original Group SLOPE problem.","fun_headline_variants_meta":{"raw":{"variants":["Safe screening rules speed Group SLOPE 14x","Provably safe screening prunes Group SLOPE zero groups","Group SLOPE screening cuts compute without accuracy loss","Screening rule slashes Group SLOPE memory and time","Safe screening discards inactive groups in Group SLOPE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000319,"raw_usage":{"total_tokens":1817,"prompt_tokens":980,"completion_tokens":837,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":760}},"tokens_in":596,"tokens_out":837,"duration_ms":8043,"temperature":1.0,"reasoning_tokens":760,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:48:48.430359+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a design where one group has more features than samples or contains linearly dependent columns, solve the original Group SLOPE problem and the transformed problem in Eq. (5) on the same data, and compare their solutions; any difference in the objective value or in which groups have zero coefficients would show the equivalence and the safety guarantee do not hold in that setting.","supporting_citations":[{"cited_title":"Journal of the American Statistical Associa tion 114(525), 419–433 (2019)","cited_arxiv_id":null,"evidence_quote":"Defines Group SLOPE and supplies the accelerated proximal gradient solver that the screening rule is designed to accelerate."},{"cited_title":"In: International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Introduces duality-gap-based dynamic safe screening for Lasso, the technique the paper adapts to the Group SLOPE penalty."},{"cited_title":"In: International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Provides the strong-concavity result for the dual that yields the bound $\\|\\theta-\\theta^*\\|_2 \\le \\sqrt{2G(b,\\theta)}$."},{"cited_title":"IEEE T ransactions on Signal Processing 63(19), 5121–5132 (2015)","cited_arxiv_id":null,"evidence_quote":"Establishes dynamic safe screening for Group Lasso, the group-structured precedent that this work extends to non-separable penalties."},{"cited_title":"SIAM Journal on Optimization 24(4), 2057–2075 (2014)","cited_arxiv_id":null,"evidence_quote":"Supplies the proximal stochastic gradient method with progressive variance reduction used as the SPGD baseline and vehicle for the stochastic screening experiments."},{"cited_title":"In: Advances in neural information processing s ystems","cited_arxiv_id":null,"evidence_quote":"Provides the variance-reduction estimator that underlies the stochastic gradient updates in the SPGD algorithm."},{"cited_title":"Pacific Journal of Optimization 8, 667–698 (2012)","cited_arxiv_id":null,"evidence_quote":"Introduces the general idea of safe feature elimination in sparse supervised learning on which the screening guarantee rests."}],"review_version":1}