{"id":"f62c3a30-d624-402f-98b4-58336e32f10a","arxiv_id":"2412.04047","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Bridge-type nonconvex sparse estimators can be optimized pathwise with accelerated proximal gradient and PALM algorithms, with convergence to critical points and pointwise path consistency under basin-of-attraction assumptions.","lead":"The paper applies two existing nonconvex optimization algorithms to compute the full solution path of bridge-type penalized estimators, which use ℓ^q penalties with q between 0 and 1 for sparse model selection. It proves convergence to critical points, derives the largest penalty that zeroes all coefficients, and tests the methods on sparse regression and on sparse diffusion processes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The q<1 path-consistency theorem assumes a close start in the basin of the bridge estimator, but the pathwise algorithm never verifies or guarantees this, so Theorem 1(ii) is circular for the actual path.","rationale":"The reader identifies the same weakest assumption: the close-start condition in Theorem 1(ii) is assumed rather than guaranteed. I agree that this is the most load-bearing concern. The paper's main computational contribution, convergence to critical points via Algorithm 1 and Algorithm 2, is reasonably supported by existing nonconvex optimization theory. However, the advertised path-consistency result for bridge penalties with q<1 depends on the algorithm being initialized in the basin of attraction of the bridge-type estimator. Since the actual pathwise implementation starts at zero and warm-starts from previous lambda values, there is no reason to believe this condition holds, and the theorem does not offer a verification procedure. This makes the q<1 path-consistency claim circular in its current form. The secondary issue with the lambda_max definition in (5.1) reinforces the need for revision: for q<1 the limiting subdifferential of |theta|^q at zero is the whole real line, so 'null vector is stationary' is not the right defining condition for lambda_max; the formula actually computes a fixed-point threshold. These issues are addressable, either by proving a basin-of-attraction condition for the warm-start protocol or by weakening the claim to convergence to critical points only, so the existing CONDITIONAL verdict remains appropriate rather than moving to ACCEPT or REJECT.","tokens_in":21558,"tokens_out":8086,"duration_ms":86564,"concrete_test":"Fix the diffusion simulation of Section 7.2 with n=1000 and q=1/2, and choose a lambda just below the lambda_max from Proposition 1. Run Algorithm 2 from three initializations: (a) the zero vector, which is the prescribed start at the largest lambda; (b) the warm-started solution from the adjacent larger lambda as in Remark 5.1; and (c) several random starts in Theta together with the true theta0. Record the critical point and objective value at convergence for each start. If starts (a) or (b) land on a different critical point than the best or true-initialized runs, the close-start hypothesis in Theorem 1(ii) is not satisfied by the pathwise protocol, and the q<1 path-consistency claim needs an additional basin-of-attraction guarantee. If all starts coincide at every lambda, the concern is mitigated empirically.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central algorithmic convergence claim (4.9) is plausible: both Algorithm 1, via the KL/semi-algebraic framework, and Algorithm 2, as a PALM method, can be expected to produce sequences converging to critical points. The load-bearing weakness is Theorem 1(ii). For 0<q_i<1, the theorem promises pointwise path consistency of the estimates produced by Algorithm 2 only 'if the starting point is quite close to the bridge-type estimate' (Section 4.2, property 2, and Section 5, Theorem 1(ii), citing Theorem 2.12 in [3]). But the pathwise protocol in Remark 5.1 starts at the largest lambda, where the initial point is the zero vector, and then warm-starts from the previous lambda-solution. Nothing in the paper checks that such initializations lie in the basin of attraction of the bridge-type estimator theta-hat_n, which is the very object the path is supposed to recover. The proof of Theorem 1(ii) says only that 'the argument is analogous, where (local) convergence to a global minimum is guaranteed under the same conditions discussed for PALM Algorithm'; this does not supply a condition that the algorithm can verify or that holds for the prescribed starts. Thus the theorem conditionally assumes the conclusion: if the algorithm is already close enough to theta-hat_n, the path is consistent. Without a practical basin-of-attraction check or an initialization rule that provably lands in it, the q<1 path-consistency claim is unsupported. A secondary definitional issue compounds this: (5.1) calls lambda_max the smallest lambda for which the null vector is stationary, but for q<1 the lq penalty has partial subdifferential R at zero, so stationarity of the null vector is trivial for every lambda; the quantity actually computed is an algorithmic fixed-point threshold. Lemma 1 and the indicator in Theorem 1 rely on this computed threshold, so the definition needs repair.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two pathwise algorithms for computing adaptive bridge-type estimators with multiple nonconvex penalties of the form sum of weighted l_q norms, q in (0,1]. Algorithm 1 is a monotone accelerated proximal gradient method that uses the exact adaptive thresholding operator T^q for the q-norm proximal map; Algorithm 2 is a blockwise proximal alternating linearized minimization (PALM) method. The authors claim convergence to critical points for both algorithms, derive a formula for the largest penalty lambda_max, state a path-consistency theorem (Theorem 1), and demonstrate the methods on GLM regression and on discretely observed ergodic diffusion processes. The computational machinery is mostly imported from existing nonconvex optimization theory, while the statistical oracle properties are imported from the authors' earlier paper [16].","tokens_in":21851,"tokens_out":15203,"duration_ms":145490,"significance":"If fully established, the paper would be a useful contribution because it computes the bridge path directly from the nonconvex, nonsmooth objective rather than through a convex relaxation, and it handles mixed-rate asymptotics via blockwise penalties. The thresholding formulas and the application to diffusion processes are concrete strengths, and the convergence-to-critical-point claims for the two algorithms are plausible from the cited KL/semi-algebraic framework. The main new statistical claim, however, is path consistency for 0<q_i<1 in Theorem 1(ii), and that claim is not supported as stated because it assumes an unverified basin-of-attraction condition. The LASSO case (Theorem 1(i)) is more defensible, but it also relies on oracle properties imported from [16] rather than being proved here.","major_comments":[{"comment":"The path-consistency result for 0<q_i<1 is conditional on the starting point being 'quite close' to the bridge-type estimate, with Theorem 2.12 of [3] cited for local convergence. The pathwise protocol in Remark 5.1 initializes at the zero vector at lambda_max and then warm-starts from the previous lambda-solution; the manuscript contains no argument that such starts lie in the basin of attraction of the global bridge estimator for large n, nor any verifiable condition that a user can check. As stated, Theorem 1(ii) assumes essentially that the algorithm already converges to the estimator whose consistency is being claimed, making the q<1 path-consistency claim circular for the actual pathwise outputs.","section":"Section 4.2, property 2; Section 5, Theorem 1(ii); Remark 5.1"},{"comment":"The proof of part (ii) consists of the sentence 'In the bridge case the argument is analogous, where (local) convergence to a global minimum is guaranteed under the same conditions discussed for PALM Algorithm.' This does not establish the theorem. For q<1, Algorithm 2 is only guaranteed to converge to a critical point (Theorem 1 in [8]), and local convergence to the global minimizer is exactly the basin-of-attraction condition flagged above. The proof needs to show that, under the prescribed pathwise initialization, the output of Algorithm 2 satisfies A_n^{-1}(hat_theta_n(lambda)-theta_0) 1{lambda_max^{(n)}>lambda}=O_p(1); no such argument is supplied.","section":"Section 5, proof of Theorem 1(ii)"},{"comment":"The extrapolation step in Algorithm 1 uses c_{t-1} and zeta^t, but the initialization only sets c_1=1 and theta^1=theta^0=eta_0; c_0 and zeta^1 are never defined, so for t=1 the quantities c_{t-1}/c_t and zeta^t are undefined. The update should be rewritten in the standard FISTA/monotone-APG form of [29], or explicit initialization of c_0 and zeta^0 should be provided. As written, Algorithm 1 is not reproducible from the pseudocode.","section":"Section 4.1, Eq. (4.4)"},{"comment":"The paper defines lambda_max using the phrase 'stationary point' but then characterizes it by the fixed-point condition T^q_{lambda s w}(0)=0. For 0<q<1 the thresholding operator is discontinuous, and the two notions are not equivalent: the fixed-point condition is sufficient for the algorithm to output zero but does not, by itself, imply the subdifferential stationarity condition 0 in partial G_n(0;q) used in (4.9). Since Lemma 1 and Theorem 1 use lambda_max^{(n)} to define the non-null part of the path, this distinction needs to be clarified and proved.","section":"Section 5, Eq. (5.1) and opening sentence"}],"minor_comments":[{"comment":"The constant is written c_{q_i}^{-2}, whereas Proposition 1 and Eq. (5.2) require c_{q_i}^{-(2-q_i)}; please correct the exponent.","section":"Lemma 1, display after Eq. (5.9)"},{"comment":"The text and the Figure 6 caption refer to 'Algorithms 1 and 3' and 'PALM (Alg. 3)', but only Algorithms 1 and 2 are defined in the paper; the references should be to Algorithm 2.","section":"Section 7.2 and Figure 6"},{"comment":"The number of Monte Carlo replications used to estimate MSE and P0 is not reported; please state the number of simulated datasets.","section":"Section 7.2, Table 2"},{"comment":"There are several typos and informal phrases: 'adpative' in the abstract, 'shed in light' in Section 3, and 'yi in R^n' in Section 6.2 where y in R^n is intended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the algorithmic part may be salvageable, but the central statistical claim for q<1 needs substantial reworking or a clear limitation statement. The heavy reliance on the authors' own prior paper [16] for oracle properties and on external nonconvex optimization theory [29], [8] should be made explicit in the intro and in the proofs. Given the unverified basin-of-attraction assumption, I cannot recommend acceptance in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this is a plausible engineering-style paper, not a finished theorem paper. The genuinely new pieces are applying monotone APG and PALM to adaptive multi-penalty bridge estimators, a formula for λmax (Proposition 1), and a path-consistency claim (Theorem 1). The algorithmic convergence claims are imported from [29] and [8] and are plausible; the thresholding operators are standard, and the diffusion example is a reasonable application. If the algorithms work as claimed, practitioners get an exact pathwise tool for a class of nonconvex penalties that is usually handled by convex relaxation. That is useful.\n\nWhat the paper does well: it identifies a real computational gap, states the algorithms cleanly, and does not overclaim the oracle properties—those are imported from the authors' prior work [16]. The GLM and SDE simulations are at least indicative.\n\nSoft spots, in order. First, Theorem 1(ii) is circular for q<1. It promises path consistency only if the algorithm is started “quite close” to the bridge estimate. But the pathwise protocol (Remark 5.1) starts at λmax with zero and then warm-starts from the previous λ-solution. Nothing checks or guarantees the close-start condition; the proof says “the argument is analogous,” which is not a proof. This is the load-bearing weakness.\n\nSecond, the definition of λmax in (5.1) is not sound for q<1. It defines λmax as the smallest λ for which 0 is stationary, but stationarity of the lq penalty at zero is not the same as being a fixed point of the thresholding update. The quantity actually computed is an algorithmic threshold. Lemma 1 and Theorem 1's indicator use this quantity, so the definition needs repair.\n\nThird, the assumptions of [29] and [8] are “easy to check” but not checked. A short verification for the bridge objective would be cheap and would make the paper more convincing. Fourth, the simulations have no error bars, gains over LLA/LASSO are modest, and no code or data are released. The text also has many typos.\n\nNone of this is fatal. The algorithmic core is probably right, and the problems are fixable. But as written, the q<1 path-consistency theorem is conditional on the very object the path is supposed to compute.\n\nI would accept this for peer review: a serious referee can push for the fixes. It is not a desk reject. I would not, however, cite the path-consistency theorem in its current form.","headline":"Computationally plausible and worth a referee, but the q<1 path-consistency theorem assumes the very closeness it needs to prove.","tokens_in":22487,"tokens_out":4738,"would_cite":false,"duration_ms":46995,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62J07","62F12","65K10","90C26","62M05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The full solution path of a nonconvex multi-penalty bridge estimator can be computed directly by proximal thresholding updates, without convex relaxation, and the resulting path estimates are pointwise consistent.","keywords":["bridge-type estimators","adaptive thresholding operators","nonconvex optimization","path consistency","proximal maps","regularized estimators","Lq penalties","stochastic differential equations"],"falsifier":"Run Algorithm 2 on the diffusion model of Section 7.2 from a starting point far from the quasi-likelihood estimate, at a fixed $\\lambda$, and record whether the iterates converge to the same critical point as when the algorithm starts near the estimate; if the final points differ, or if the normalized path error $A_n^{-1}(\\hat{\\theta}_n(\\lambda)-\\theta_0)$ fails to stay bounded across $n$ for such starts, then the basin assumption required by Theorem 1(ii) is not satisfied in that setting.","tokens_in":21302,"feed_emoji":"🎯","tokens_out":12346,"duration_ms":98907,"temperature":0.7,"pith_summary":"Bridge-type estimators minimize a loss plus a sum of $\\ell^q$ penalties with $q\\in(0,1]$; for $q<1$ the objective is nonconvex and nondifferentiable, which is why standard practice replaces the penalty by a convex surrogate. This paper claims that the exact solution path of such estimators can nevertheless be computed, by iterating adaptive thresholding operators that are proximal maps of the $\\ell^q$ penalty. Two algorithms are proposed: a monitored accelerated proximal gradient method and a blockwise proximal alternating (PALM) scheme; both are shown to converge, for each $\\lambda$ and sample size, to a critical point of the bridge objective. The paper further claims path consistency: after normalization by the estimator's convergence rates, the computed path at any fixed $\\lambda$ is bounded in probability, with oracle selection properties under additional assumptions. A reader should care because exact computation, if correct, lets practitioners use bridge estimators' oracle behavior directly rather than through convex relaxations.","feed_headline":"Compute the full path of bridge estimators exactly","feed_subtitle":"Two proximal-gradient algorithms recover the nonconvex Lq path, keeping sparse selection and oracle consistency.","key_machinery":"The load-bearing object is the adaptive thresholding operator $T^q_{\\lambda w}(z) = (T^q_{\\lambda w_1}(z_1),\\dots,T^q_{\\lambda w_p}(z_p))^\\top$, defined as the proximal map of the weighted $\\ell^q$ penalty, whose one-dimensional form generalizes soft thresholding (for $q=1$) and hard thresholding (as $q\\to 0$). Every iteration of both algorithms is a gradient step followed by this operator, so the nonconvex penalty is used exactly, not relaxed. The convergence arguments rely on the monitored acceleration scheme for nonconvex proximal gradient methods and on the Kurdyka--Lojasiewicz/semi-algebraic framework for PALM, which together ensure that accumulation points are critical points; the path-consistency argument additionally uses the formula for $\\lambda_{\\max}$, the smallest penalty at which the null vector is stationary.","core_discovery":"The central discovery is that the bridge-type estimation problem (3.5), despite its nonconvexity, admits a pathwise optimization scheme whose updates are componentwise applications of the adaptive $q$-thresholding operator $T^q_{\\lambda w}$, the proximal map of the weighted $\\ell^q$ penalty. Algorithm 1, a monitored accelerated proximal gradient method, and Algorithm 2, a blockwise proximal alternating linearized minimization method, both generate bounded sequences that converge to a critical point with $0 \\in \\partial G_n(\\theta^*; q)$. Theorem 1 states that, under the mixed-rates assumptions (3.1), (5.6), and (5.7), the path estimates satisfy $A_n^{-1}(\\hat{\\theta}_n(\\lambda) - \\theta_0)1_{\\{\\lambda_{\\max}(n)>\\lambda\\}} = O_p(1)$ for the LASSO case for either algorithm, and for the bridge case $q<1$ for Algorithm 2 provided it is started sufficiently close to the bridge-type estimate; under further assumptions the oracle properties of selection and asymptotic normality transfer to the path.","pith_inferences":["One extension the paper leaves implicit: warm-starting each $\\lambda$ from the previous solution, recommended in the paper for speed, may in practice place the algorithm inside the basin required by Theorem 1(ii); this could be tested by comparing warm starts against random starts on the same problem.","The discontinuity of the path at the threshold points suggests a data-driven rule for choosing $\\lambda$, such as selecting the largest $\\lambda$ before a variable drops out, which the paper does not explore.","The same proximal-thresholding machinery applies to other concave penalties with known proximal maps, such as SCAD or MCP, offering a route to exact pathwise computation for those estimators as well.","A practical consequence of the basin assumption is that path consistency for $q<1$ is only guaranteed for favorable initializations; checking it empirically with multiple random starts would clarify how often the condition holds in real data."],"forward_implications":["For every $\\lambda$ and sample size $n$, the path output by either algorithm converges to a critical point of the bridge objective, so the computed path is a genuine stationary-point path rather than a relaxed approximation.","Because the thresholding operator is evaluated componentwise, the algorithms scale to high-dimensional parameter vectors and to multiple penalty groups with different exponents $q_i$ and different convergence rates.","When $q=1$, the path estimates are pointwise consistent for either algorithm; when $q<1$, the same consistency holds for the PALM algorithm under a basin-of-attraction condition on the starting point.","In the diffusion-process application, the bridge path outperforms the LASSO path in selection probability and is less sensitive to suboptimal tuning, and the blockwise scheme reduces iteration counts by roughly a factor of three relative to coordinate descent.","Under the additional assumptions imported from the authors' earlier work, the path estimates inherit the oracle properties of variable selection and asymptotic normality."],"supporting_citations":[{"why":"Supplies the monitored accelerated proximal gradient scheme used as Algorithm 1, including convergence to a critical point.","marker":"[29]"},{"why":"Supplies the PALM blockwise proximal alternating algorithm used as Algorithm 2 and its convergence for semi-algebraic objectives.","marker":"[8]"},{"why":"Provides Theorem 2.12, the basin-of-attraction condition that Theorem 1(ii) invokes to guarantee convergence to the bridge-type estimate.","marker":"[3]"},{"why":"Gives the thresholding representation of the Lq proximal map used as the componentwise update in both algorithms.","marker":"[30]"},{"why":"Establishes the oracle properties and mixed-rates asymptotic setting for the bridge estimator, which Theorem 1 extends to the computed path.","marker":"[16]"},{"why":"Provides the numerical evaluation of the proximity operator of the Lp norm for 0<p<1, needed to compute the thresholding updates in practice.","marker":"[12]"}],"fun_headline_variants":["Pathwise proximal methods for bridge-type estimators","Efficient path computation for nonconvex bridge penalties","Proximal algorithms trace the full Lq path","Adaptive thresholding enables bridge path optimization","Accelerated gradient and alternating descent for bridge paths"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For $q<1$, the proof that the computed path is consistent assumes the algorithm is started very close to the bridge-type estimator it is supposed to recover, and the paper does not provide a check or guarantee for that starting condition.","fun_headline_variants_meta":{"raw":{"variants":["Pathwise proximal methods for bridge-type estimators","Efficient path computation for nonconvex bridge penalties","Proximal algorithms trace the full Lq path","Adaptive thresholding enables bridge path optimization","Accelerated gradient and alternating descent for bridge paths"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000313,"raw_usage":{"total_tokens":1811,"prompt_tokens":1008,"completion_tokens":803,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":731}},"tokens_in":624,"tokens_out":803,"duration_ms":7411,"temperature":1.0,"reasoning_tokens":731,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:49:24.948630+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 2 on the diffusion model of Section 7.2 from a starting point far from the quasi-likelihood estimate, at a fixed $\\lambda$, and record whether the iterates converge to the same critical point as when the algorithm starts near the estimate; if the final points differ, or if the normalized path error $A_n^{-1}(\\hat{\\theta}_n(\\lambda)-\\theta_0)$ fails to stay bounded across $n$ for such starts, then the basin assumption required by Theorem 1(ii) is not satisfied in that setting.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the monitored accelerated proximal gradient scheme used as Algorithm 1, including convergence to a critical point."},{"cited_title":"Bolte, S","cited_arxiv_id":null,"evidence_quote":"Supplies the PALM blockwise proximal alternating algorithm used as Algorithm 2 and its convergence for semi-algebraic objectives."},{"cited_title":"Attouch, J","cited_arxiv_id":null,"evidence_quote":"Provides Theorem 2.12, the basin-of-attraction condition that Theorem 1(ii) invokes to guarantee convergence to the bridge-type estimate."},{"cited_title":"Li and Z","cited_arxiv_id":null,"evidence_quote":"Gives the thresholding representation of the Lq proximal map used as the componentwise update in both algorithms."},{"cited_title":"De Gregorio and S","cited_arxiv_id":null,"evidence_quote":"Establishes the oracle properties and mixed-rates asymptotic setting for the bridge estimator, which Theorem 1 extends to the computed path."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the numerical evaluation of the proximity operator of the Lp norm for 0<p<1, needed to compute the thresholding updates in practice."}],"review_version":1}