{"id":"b3145a84-c198-44c9-9af0-96d3e1abe472","arxiv_id":"1908.07211","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An anchored forward-backward-forward iteration is shown to converge strongly to the minimal-norm solution of pseudo-monotone variational inequalities, with an adaptive variant and traffic-network tests.","lead":"This paper adds an anchoring step to Tseng's forward-backward-forward method so that iterates converge strongly to the smallest-norm solution of pseudo-monotone variational inequalities in Hilbert spaces, and tests the scheme on dynamic user equilibrium in traffic networks. Strong convergence matters because it gives stable numerical behavior in infinite-dimensional equilibrium problems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 3.6 applies pseudo-monotonicity to a point y + ε_j F(z_Nj)/||F(z_Nj)||² that is not shown to lie in X; under Assumption 2 (pseudo-monotone only on X) the step before (3.18) is unjustified, so Theorem 3.1 is proved only under a stronger unstated hypothesis.","rationale":"The reader's verdict is CONDITIONAL, and its weakest assumption is exactly the point I find most load-bearing: Lemma 3.6 invokes pseudo-monotonicity at a point that need not belong to X. I agree with that assessment. The surrounding proof has a coherent recursive structure: Lemma 3.3 gives the Fejér-type inequality, Lemma 3.4 gives boundedness, Lemma 3.5 gives the error recursion handled by Xu's lemma, and Lemma 3.6 is the only place that converts weak cluster points into solutions. If Lemma 3.6 is not repaired, neither case of Theorem 3.1 closes. I do not see a second independent route to X* in the manuscript. The adaptive theorem is only sketched, but since it is stated to follow analogously, it inherits the problem. I would not reject the paper: the gap is a hypothesis-domain mismatch, not an algebraic error. If Assumption 2 is strengthened to pseudo-monotonicity on all of H (or on a bounded neighborhood of X), the epsilon-trick in Lemma 3.6 goes through, because the perturbed point is eventually inside that neighborhood and Lipschitz continuity supplies the strong convergence of F(y + ε_j u_j) needed in the final limit. So the appropriate status remains CONDITIONAL, and since the reader already issued that verdict, I recommend no change. The numerical experiments are not directly affected by this theoretical gap, but the claim that Algorithm 3 converges under pseudo-monotonicity on Λ would need the strengthened property or a new proof.","tokens_in":16570,"tokens_out":17873,"duration_ms":188392,"concrete_test":"Run the following analytical check. Set H = R², X = {x ∈ R² : x₁ ≥ 0}, F(x) = (x₁ − 1, x₂); F is L-Lipschitz, sequentially weak-to-weak continuous, and monotone, hence pseudo-monotone on X. In Lemma 3.6 take y = 0 and z_Nj = 0. Then F(z_Nj) = (−1, 0), so y + ε_j F(z_Nj)/||F(z_Nj)||² = (−ε_j, 0) ∉ X. This shows the pseudo-monotonicity step before (3.18) cannot be applied under Assumption 2 for an admissible instance. As a second part of the check, attempt to replace the perturbed point by P_X(y + ε_j u_j) and control the projection error to re-derive (3.18); if the nonnegativity sign cannot be preserved, the lemma is only provable under pseudo-monotonicity on a neighborhood of X or on all of H.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 3.6 is the only step that converts the gap condition ||x_k − z_k|| → 0 into the conclusion that a weak cluster point lies in X*. Its final part rests on the line before (3.18): since ⟨F(z_Nj), y + ε_j u_Nj − z_Nj⟩ ≥ 0 and 'F is pseudo-monotone', the authors assert ⟨F(y + ε_j u_Nj), y + ε_j u_Nj − z_Nj⟩ ≥ 0. But Assumption 2 defines pseudo-monotonicity only for pairs in X, and the perturbed point is not shown to lie in X. For a general closed convex X it is outside: take H = R², X = {x₁ ≥ 0}, y = 0, z_Nj = 0, and F(x) = (x₁ − 1, x₂). This F is Lipschitz, sequentially weak-to-weak continuous, and monotone, hence pseudo-monotone on X, yet the perturbed point is (−ε_j, 0) ∉ X. Thus (3.18) is not justified in the stated generality. This gap is load-bearing: both cases of the proof of Theorem 3.1 invoke Lemma 3.6, and Theorem 3.2 is said to follow analogously, so it inherits the problem. The natural repair is to strengthen Assumption 2 to pseudo-monotonicity on a neighborhood of X or on all of H, and to verify that condition in the DUE application, or to supply a valid argument keeping the perturbed point inside X.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a forward-backward-forward algorithm for solving variational inequalities VI(X,F) in real Hilbert spaces under Lipschitz continuity, sequential weak-to-weak continuity, and pseudo-monotonicity of F. The update adds an anchoring extrapolation x_{k+1}=(1-alpha_k-beta_k)x_k+beta_k r_k to the classical Tseng iteration, and the paper claims strong convergence to the minimal-norm solution p=P_{X*}(0). An adaptive step-size variant is also claimed. The final section applies the algorithm to dynamic user equilibrium in traffic networks and reports numerical comparisons on Nguyen and Sioux Falls instances.","tokens_in":16895,"tokens_out":19978,"duration_ms":182176,"significance":"The intended contribution is timely: strong convergence with a single projection per iteration under pseudo-monotonicity would improve over weakly convergent Tseng methods and over strongly convergent methods requiring extra projections. The non-adaptive part of the proof is largely self-contained and uses standard tools (Xu's lemma, quasi-Fejer inequalities, two-case Mange argument), and the numerical study is useful empirical evidence. However, the central lemma that identifies weak cluster points as solutions is not proved under the stated assumptions: it applies pseudo-monotonicity at points outside X and passes to the limit in an inner product of two weakly convergent sequences. The adaptive theorem is only sketched and its displayed estimate is not sufficient. These gaps are load-bearing; the main theorems do not yet follow. The DUE application also does not verify the assumptions.","major_comments":[{"comment":"The proof of Lemma 3.6 applies the pseudo-monotonicity implication to the pair (y+epsilon_j u_Nj, z_Nj) immediately before Eq. (3.18). Assumption 2 defines pseudo-monotonicity only for pairs in X, and the point y+epsilon_j u_Nj is never shown to belong to X. This is not a technicality: for X={x in R^2 : x_1 >= 0}, y=0, z_Nj=0, and F(x)=(x_1-1, x_2), the map F is Lipschitz, sequentially weak-to-weak continuous, and monotone (hence pseudo-monotone on X), yet y+epsilon_j u_Nj = (-epsilon_j, 0) is outside X. Therefore the inequality (3.18) is not justified. Since Lemma 3.6 is invoked in both cases of the proof of Theorem 3.1 and the proof of Theorem 3.2 is said to follow analogously, Theorems 3.1 and 3.2 are not established under the stated assumptions. The proof can be repaired by assuming pseudo-monotonicity on a neighborhood of X or on all of H, or by providing a valid argument that keeps the perturbed point inside X.","section":"§3.2, Lemma 3.6, Eq. (3.18)"},{"comment":"Even if (3.18) were valid, the passage to the limit to conclude 0 <= <F(y), y-x_hat> is not justified: the first factor F(y+epsilon_j u_Nj) converges weakly to F(y), and the second factor y+epsilon_j u_Nj - z_Nj converges weakly to y-x_hat, but weak convergence of both factors does not imply convergence of their inner product to the inner product of the limits. A strong convergence of at least one factor (or a different proof structure) is required. Consequently the conclusion x_hat in X* is not obtained from the displayed argument.","section":"§3.2, Lemma 3.6, limit after Eq. (3.18)"},{"comment":"The proof of the adaptive variant is a one-line sketch. The displayed bound ||r_k-x*||^2 <= ||x_k-x*||^2 - (1 - gamma_k^2 rho^2 / gamma_{k+1}^2) ||x_k-z_k||^2 does not establish the analogue of Lemma 3.3, because gamma_{k+1} <= gamma_k implies gamma_k^2 rho^2 / gamma_{k+1}^2 can exceed 1, making the alleged contraction coefficient negative. The adaptive rule (3.2) does not prevent this. Thus the assertion that 'the rest of the proofs follows analogously' is not sufficient; Theorem 3.2 requires a complete proof.","section":"§3.2, Proof of Theorem 3.2"},{"comment":"The DUE application does not verify that the effective delay operator Psi satisfies the hypotheses of Theorems 3.1-3.2, namely Lipschitz continuity, sequential weak-to-weak continuity, pseudo-monotonicity on the feasible set (or on a neighborhood), and nonempty solution set. Section 4.2 states only that Algorithm 3 is equivalent to Algorithm 1 'if' Psi is Lipschitz continuous and pseudomonotone, and the numerics are run without establishing these conditions. As a result, the convergence guarantee does not formally apply to the DUE computations reported.","section":"§4.2-4.3"}],"minor_comments":[{"comment":"The abstract contains two typos: 'device' should be 'devise' and 'pseudomonote' should be 'pseudomonotone'.","section":"Abstract"},{"comment":"The statement of Theorem 3.2 says the sequence is generated by Algorithm 1; it should refer to Algorithm 2.","section":"Theorem 3.2 statement"},{"comment":"The quantity y_{tau(k)} is used in the Case 2 argument but never defined; the surrounding formulas indicate z_{tau(k)} is intended.","section":"Proof of Theorem 3.1, Case 2"},{"comment":"The sentence claiming the limit point is 'not smaller than {gamma_0, rho/L}' should read min{gamma_0, rho/L}.","section":"§3.1, after Eq. (3.2)"},{"comment":"The assertion that 'for each j >= 1, F(z_Nj) != 0' is not justified; if F(z_Nj)=0 then z_Nj is already in X* and this case should be handled separately before defining u_Nj.","section":"Lemma 3.6"},{"comment":"The figure numbering is inconsistent: Figure 2 appears with different captions, and the caption referring to 'four test networks' includes networks not discussed in the numerical comparisons.","section":"§4.3, figures"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem and the algorithmic idea is attractive, but the proof of Lemma 3.6 needs substantial repair before the central claims are established. If the authors fix the pseudo-monotonicity domain issue and the limit argument, and provide a full proof for the adaptive variant, the paper would be a solid contribution. The numerical experiments are useful but need either verification of assumptions for DUE or an explicit framing as heuristic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the core idea: anchor Tseng's forward-backward-forward iteration with a Halpern-type term to force strong convergence to the minimal-norm solution of a pseudo-monotone variational inequality, without hyperplane projections. That combination is new in the cited literature, and the algorithmic design is sensible. The main estimates in Lemmas 3.3–3.5 are standard and correctly derived. The proof is a genuine derivation, not a circular argument; the anchoring drift is what selects the minimal-norm solution, not some fitted constant.\n\nThe soft spot is real and it is load-bearing. In Lemma 3.6, after constructing u_Nj = F(z_Nj)/||F(z_Nj)||², the proof applies pseudo-monotonicity to the pair (z_Nj, y + ε_j u_Nj). Assumption 2 only defines pseudo-monotonicity for points in X, and the perturbed point is never shown to lie in X. For a general closed convex X it can lie outside—take H=R², X={x₁≥0}, y=0, z_Nj=0, and F(x)=(x₁−1, x₂); then the perturbed point is (−ε_j, 0) ∉ X. So the line before (3.18) is unjustified as stated. This matters because Lemma 3.6 is the step that converts ||x_k−z_k||→0 into the cluster point being a solution. Both cases of Theorem 3.1 rely on it, and Theorem 3.2 inherits the problem. The natural repair is to strengthen Assumption 2 to pseudo-monotonicity on a neighborhood of X or on all of H, and then verify that condition in the DUE application, or else to supply a valid argument that keeps the perturbed point inside X.\n\nOther soft spots are minor by comparison. Theorem 3.2's proof is only a sketch (\"left to the reader\"), which is a real omission for an adaptive variant. The numerical comparison is weakened by per-instance tuning of step-size and momentum parameters, no code or data release, and an apparent template artifact—a \"Section 5.1\" about a fixed-point algorithm that does not fit this paper's numbering. That suggests carelessness rather than a flaw in the mathematics.\n\nThis paper is for readers working on strong convergence for pseudo-monotone VIs and on DUE algorithms. The idea is worth engaging; the gap is addressable. I would send it to a serious referee, and if I were reviewing I would ask for the assumption to be strengthened and the adaptive proof completed. As it stands, the main theorem is not established in the stated generality, but the paper deserves referee time rather than desk rejection.","headline":"Anchored Tseng forward-backward-forward for pseudo-monotone VIs is a clean and genuinely new combination, but Lemma 3.6 has a load-bearing gap—pseudo-monotonicity is applied at points outside X—so the main theorem needs a stronger assumption or a repaired proof.","tokens_in":17470,"tokens_out":2617,"would_cite":false,"duration_ms":24337,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["47J20","47J25","49J40","65K15","90C33"],"pacs":[],"model":"deepseek-v4-flash","headline":"For pseudo-monotone variational inequalities in Hilbert spaces, a cheaply modified forward-backward-forward algorithm converges strongly to the minimal-norm solution, and an adaptive step-size variant does so without knowledge of the…","keywords":["variational inequalities","pseudo-monotone maps","forward-backward-forward method","strong convergence","minimal-norm solution","adaptive step-size","dynamic user equilibrium","traffic networks"],"falsifier":"Search for a bounded, Lipschitz, weak-to-weak continuous map F that is pseudo-monotone on X but fails pseudo-monotonicity at every point outside X, with X* nonempty; if running Algorithm 1 with parameters satisfying (3.1) yields a trajectory whose weak cluster point is not in X*, or whose ||x_k - z_k|| does not vanish, then the omitted neighborhood condition is essential and Theorem 3.1 as stated is false. A quicker check: compute the points y + epsilon_j F(z_{N_j})/||F(z_{N_j})||^2 from the proof and test whether the implication defining pseudo-monotonicity holds there; if it fails, the proof step (3.18) has no justification.","tokens_in":16324,"feed_emoji":"🎯","tokens_out":11933,"duration_ms":103883,"temperature":0.7,"pith_summary":"This paper claims that adding one convex-combination step to the forward-backward-forward algorithm turns weak convergence into strong convergence for variational inequalities defined by a pseudo-monotone map on a Hilbert space. The modification is computationally cheap: the iteration still uses only one projection per step, and the same guarantee holds for an adaptive step-size version that does not need to know the Lipschitz constant. A sympathetic reader would care because strong convergence is stable under discretization in infinite-dimensional settings, while pseudo-monotonicity is the realistic assumption for dynamic user equilibrium in traffic networks, where strong monotonicity is known to fail. The paper shows the method working on two standard traffic networks, reporting that it matches or beats the projected-gradient baseline while providing stronger guarantees.","feed_headline":"A cheap modification forces strong convergence to minimal-norm solution","feed_subtitle":"No extra projections, no need to know the Lipschitz constant, and tests on traffic networks match the best solvers.","key_machinery":"The central object is the augmented forward-backward-forward iteration. In the base step, z_k is the projection of x_k - gamma F(x_k) onto X, and r_k = z_k + gamma(F(x_k)-F(z_k)) is the forward-backward-forward correction; the new point x_{k+1} is then the convex combination (1-alpha_k-beta_k)x_k + beta_k r_k. Relative to the target p, this update carries a small pull term -alpha_k p, and the conditions on alpha_k and beta_k make that pull vanish slowly enough to force strong convergence while preserving the weak-convergence structure inherited from the base method. The proof chains three ingredients: a basic recursion (Lemma 3.3) that contracts ||r_k - x*|| by the gap (1-(gamma L)^2)||x_k - z_k||^2, a quasi-Fejér inequality (Lemma 3.5) that converts that gap into a perturbed recursion for ||x_{k+1} - p||^2, and a cluster-point lemma (Lemma 3.6) that uses pseudo-monotonicity plus weak-to-weak continuity to identify weak limits with vanishing gap as solutions. In the adaptive variant, the same contraction holds with (1 - $gamma_k^{2}$ $rho^{2}$ / gamma_{k+1}^2) in place of (1-(gamma L)^2).","core_discovery":"Under Lipschitz continuity, sequential weak-to-weak continuity, and pseudo-monotonicity of F, with a nonempty closed convex solution set X*, Algorithm 1 produces a sequence (x_k) that converges strongly to p = argmin{||z|| : z in X*}, the minimal-norm solution of VI(X,F). The update is x_{k+1} = (1-alpha_k-beta_k)x_k + beta_k r_k, with z_k = P_X(x_k - gamma F(x_k)) and r_k = z_k + gamma(F(x_k)-F(z_k)), and the parameter sequences satisfy alpha_k -> 0, sum alpha_k = infinity, and beta_k bounded away from 0 and from 1-alpha_k. Theorem 3.2 shows the same strong convergence when the constant step-size is replaced by the recursive rule (3.2), which shrinks gamma_k according to the ratio ||z_k - x_k||/||F(z_k)-F(x_k)||, so no global Lipschitz constant is needed. The proof establishes a perturbed Fejér-type recursion for ||x_k - p||^2 whose perturbation vanishes suitably, then shows that every weak cluster point of the sequence must be a solution, and concludes strong convergence via a classical lemma on perturbed contractions.","pith_inferences":["The same 'pull toward the origin' augmentation could be grafted onto other single-projection algorithms (for example, extragradient variants or proximal-point methods) to force strong convergence for pseudo-monotone maps, as long as a quasi-Fejér relation with a vanishing gap can be established.","A testable refinement would be to replace the unverified neighborhood pseudo-monotonicity with the weaker and checkable condition that F is pseudo-monotone on the convex hull of X and the iterates; if that suffices, the theorem could be proved without extending the assumption off X.","In the traffic application, the delay operator is only evaluated at finitely many discretization points through the dynamic network loading subroutine; the strong-convergence guarantee would carry over to the discretized algorithm only if the approximate operators are uniformly pseudo-monotone on the relevant region, which is an empirical question the paper does not settle.","Converging to the minimal-norm solution gives a principled selection rule among multiple dynamic user equilibria—the one with least total departure-flow energy—which could matter for policy evaluation beyond the paper's examples."],"forward_implications":["For any pseudo-monotone variational inequality satisfying the paper's assumptions, the minimal-norm solution can now be computed with one projection per iteration and guaranteed strong convergence in the Hilbert-space norm.","The adaptive step-size rule removes the need to estimate a global Lipschitz constant, so the method can be applied to black-box operators such as the effective-delay operator in dynamic traffic assignment.","Because strong convergence is stable under discretization, the method's iterates from any finite-dimensional approximation will not drift away from the true solution set.","The experiments indicate the method is at least competitive with the projected-gradient baseline on the Nguyen and Sioux Falls networks, while carrying stronger theoretical guarantees.","Pseudo-monotonicity is strictly weaker than monotonicity, so the method covers equilibrium problems such as dynamic user equilibrium where strong monotonicity provably fails."],"supporting_citations":[{"why":"Supplies the base forward-backward-forward splitting iteration that the paper augments.","marker":"[32]"},{"why":"Provides the perturbed-contraction lemma (Lemma 2.3) used to conclude strong convergence from the recursion.","marker":"[35]"},{"why":"Supplies the projection characterization (Lemma 2.1) used in the fundamental recursion of Lemma 3.3.","marker":"[13]"},{"why":"Provides the proof technique for the non-monotone case (the tau(k) argument in Case 2 of Theorem 3.1).","marker":"[25]"},{"why":"The closest predecessor proving strong convergence for Tseng-type methods in the monotone setting, which the paper extends to pseudo-monotone VIs.","marker":"[12]"},{"why":"The alternative strongly convergent Tseng-type method using hyperplane projections, against which the paper offers a simpler scheme.","marker":"[31]"},{"why":"Gives the variational inequality formulation of dynamic network user equilibrium used in Section 4.","marker":"[10]"},{"why":"Provides the test networks, the dynamic network loading subroutine, and the projected-gradient baseline used in the numerical experiments.","marker":"[17]"}],"fun_headline_variants":["Cheap tweak forces strong convergence in pseudo-monotone VIs","Adaptive step-size yields strong convergence, no Lipschitz constant","Strong convergence to minimal-norm solution via simple update","Forward-backward-forward upgrade: strong convergence, adaptive","Traffic network equilibria: strongly convergent, adaptive solver"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof's load-bearing premise is that F is pseudo-monotone not only on the feasible set X but on a neighborhood of X or all of H, because Lemma 3.6 applies the pseudo-monotonicity definition to the auxiliary point y + epsilon_j F(z_{N_j})/||F(z_{N_j})||^2, which need not lie in X.","fun_headline_variants_meta":{"raw":{"variants":["Cheap tweak forces strong convergence in pseudo-monotone VIs","Adaptive step-size yields strong convergence, no Lipschitz constant","Strong convergence to minimal-norm solution via simple update","Forward-backward-forward upgrade: strong convergence, adaptive","Traffic network equilibria: strongly convergent, adaptive solver"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000296,"raw_usage":{"total_tokens":1731,"prompt_tokens":969,"completion_tokens":762,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":678}},"tokens_in":585,"tokens_out":762,"duration_ms":7653,"temperature":1.0,"reasoning_tokens":678,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:25:37.136331+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Search for a bounded, Lipschitz, weak-to-weak continuous map F that is pseudo-monotone on X but fails pseudo-monotonicity at every point outside X, with X* nonempty; if running Algorithm 1 with parameters satisfying (3.1) yields a trajectory whose weak cluster point is not in X*, or whose ||x_k - z_k|| does not vanish, then the omitted neighborhood condition is essential and Theorem 3.1 as stated is false. A quicker check: compute the points y + epsilon_j F(z_{N_j})/||F(z_{N_j})||^2 from the proof and test whether the implication defining pseudo-monotonicity holds there; if it fails, the proof step (3.18) has no justification.","supporting_citations":[{"cited_title":"Iterative algorithms for nonlinear operators","cited_arxiv_id":null,"evidence_quote":"Provides the perturbed-contraction lemma (Lemma 2.3) used to conclude strong convergence from the recursion."},{"cited_title":"Goebel and S","cited_arxiv_id":null,"evidence_quote":"Supplies the projection characterization (Lemma 2.1) used in the fundamental recursion of Lemma 3.3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the proof technique for the non-monotone case (the tau(k) argument in Case 2 of Theorem 3.1)."},{"cited_title":"A strong convergence theorem for tseng’s extragradient method for solving variational inequality prob- lems","cited_arxiv_id":null,"evidence_quote":"The alternative strongly convergent Tseng-type method using hyperplane projections, against which the paper offers a simpler scheme."},{"cited_title":"Friesz, David Bernstein, Tony E","cited_arxiv_id":null,"evidence_quote":"Gives the variational inequality formulation of dynamic network user equilibrium used in Section 4."},{"cited_title":"Computing dynamic user equilibria on large- scale networks with software implementation","cited_arxiv_id":null,"evidence_quote":"Provides the test networks, the dynamic network loading subroutine, and the projected-gradient baseline used in the numerical experiments."}],"review_version":1}