{"id":"4dfce4a6-fd99-4607-a907-bc7ab12bd5fb","arxiv_id":"1908.03633","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A short technical report claims Chen-Teboulle is the proximal point algorithm and derives a larger step-size bound, but the equivalence is contradicted by the paper's equations.","lead":"This note argues that the Chen-Teboulle primal-dual algorithm is a proximal point algorithm with a specially chosen metric, and uses that view to allow a larger step size. The claimed equivalence is not demonstrated and appears inconsistent with the paper's own update equations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed PPA equivalence is real but the text omits a required dual-variable shift; paper needs revision, not rejection.","rationale":"The central claim is that Chen-Teboulle is a proximal point method in the metric (12), which transfers the PPA step bound (17). The weakest assumption is the equivalence in Section 2.1. A direct substitution shows the naive identification fails, so the paper's proof as written is invalid. However, the equivalence is salvageable by a one-step shift in the dual variable: the PPA variable y_k must be the CT predictor p_{k+1}, not the CT dual variable y_k. With that shift, the x,z updates coincide exactly and the y-update (15) becomes the identity p_{k+2}=p_{k+1}+lambda(2Ax_{k+1}-Ax_k-2z_{k+1}+z_k), which follows from (5)-(8). The V positive definite calculation leading to (17) is independent and correct. Thus the appropriate disposition is conditional acceptance: the main result appears true, but the manuscript must state the missing variable redefinition and correct the indexing claim, or the equivalence remains unproven. This agrees partially with the reader: the identified weak step is genuine, but the counterexample is a symptom of a missing transformation rather than evidence that the improved step bound is false.","tokens_in":4048,"tokens_out":19402,"duration_ms":179470,"concrete_test":"Verify the variable-shift equivalence explicitly: for arbitrary f,g in Gamma_0 and bounded linear A, set v_k=p_{k+1}=y_k^{CT}+lambda(Ax_k-z_k) and write the PPA iteration (13)-(15) with y_k replaced by v_k. Derive that x_{k+1} and z_{k+1} equal the Chen-Teboulle updates (6)-(7), and that v_{k+1}=p_{k+2}=y_{k+1}^{CT}+lambda(Ax_{k+1}-z_{k+1}). Also rerun the reader's f=g=0, A=I, lambda=1/2 test using v_0=p_1; the resulting x,z,v sequence should coincide with the Chen-Teboulle sequence, resolving the counterexample.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing step is the Section 2.1 assertion that (5)-(8) and (13)-(15) agree after 're-organizing the steps.' Read naively, they do not: for f=g=0, A=I, lambda=1/2, a single Chen-Teboulle step from (x,z,y)=(1,0,0) gives (0.75,0.25,0.25), while (13)-(15) give (1,0,0.5). This matches the reader's objection and shows the displayed equations are not the same map on the same variables. The theorem can still be rescued: define v_k=p_{k+1}=y_k+lambda(Ax_k-z_k). Then (13) becomes x_{k+1}=prox_{lambda f}(x_k-lambda A^* v_k), matching (6), and (14) matches (7); the recurrence (15) gives v_{k+1}=v_k+lambda(2Ax_{k+1}-Ax_k-2z_{k+1}+z_k)=p_{k+2}, so the PPA dual variable is the Chen-Teboulle predictor, not its y_k. The paper never states this transformation; its sentence that p^{k+1} equals y_{k+1} in (15) is off by one index. As written, the transfer of PPA convergence and bound (17) to Chen-Teboulle is therefore not proven, even though the underlying identification is correct.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims that the Chen-Teboulle algorithm (equations (5)-(8)) is exactly the proximal point algorithm applied to the monotone inclusion (4) with a specially chosen metric V defined in (12). On the basis of this equivalence, the author derives the step-size condition lambda <= 1/sqrt(||A||^2+1) in (17), which is weaker than the original condition lambda < 1/(2L) in (9). The paper is a short technical report; its main contribution is the identification with a variable-metric proximal point method and the resulting improved step-size bound.","tokens_in":4338,"tokens_out":5496,"duration_ms":53576,"significance":"If the central equivalence were correctly established, the paper would provide a clean and conceptually useful derivation of the Chen-Teboulle algorithm and a genuinely less restrictive step-size condition, inheriting proximal point convergence guarantees. The Schur complement computation leading to (17) is straightforward and self-contained, with no fitted parameters. However, the load-bearing claim that equations (13)-(15) recover Chen-Teboulle is asserted without a proof and, as written, is false because the dual variable is misidentified. The underlying identification is likely repairable by shifting the dual variable, but the present manuscript does not contain the required statement or proof.","major_comments":[{"comment":"The paper asserts that choosing tau_x = tau_y = tau_z = 1/lambda and re-organizing the steps recovers the Chen-Teboulle algorithm, but the displayed updates are not the same map on the same variables. For f = g = 0, A = I, and lambda = 1/2, starting from (x0,z0,y0) = (1,0,0), the Chen-Teboulle updates (5)-(8) give (x1,z1,y1) = (0.75, 0.25, 0.25), while the proximal updates (13)-(15) give (1, 0, 0.5). This numerical discrepancy shows that the equivalence as stated is false, and since the step-size bound (17) and the proximal point convergence results are transferred to Chen-Teboulle through this equivalence, the paper's central claim is not proven as written.","section":"Section 2.1, equations (13)-(15) and the claim after (15)"},{"comment":"The statement that 'the p^{k+1} in (5) is the same as the y_{k+1} in (15)' is off by one index. The correct identification appears to be v_k = p_{k+1} = y_k + lambda(A x_k - z_k): with this shift, equation (6) matches (13), equation (7) matches (14), and equation (15) yields v_{k+1} = p_{k+2}. The manuscript needs to introduce this shifted dual variable explicitly and prove the equivalence in detail; until this is done, the improved bound (17) has not been shown to apply to Chen-Teboulle.","section":"Section 2.1, sentence comparing p^{k+1} and y_{k+1}"}],"minor_comments":[{"comment":"The condition for V to be positive definite requires the strict inequality lambda < 1/sqrt(||A||^2+1), because at equality the Schur complement is only positive semidefinite. The paper writes 'lambda <=' in (17); if the endpoint is meant to be included, a separate argument or a non-strict variant of the convergence result is needed.","section":"Equation (17)"},{"comment":"The 'matrix-multiplication' notation in (10) is described as 'A . (x,y,z)', but the block vector in the surrounding text is (x,z,y); please make the ordering consistent.","section":"Equation (10) and surrounding text"},{"comment":"The notation 'zero(A)' is introduced but not used later; the non-emptiness assumption is stated once and then the text proceeds. Consider defining the set of zeros once and referring back to it when needed.","section":"Section 1.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a short technical report whose central claim is currently unproven and, as stated, false. However, the apparent off-by-one identification is plausibly repairable by introducing the shifted dual variable v_k = p_{k+1}, which suggests that the result may be correct. I therefore recommend major revision rather than rejection: the authors should provide a rigorous proof of the equivalence with the correct variable shift and correct the strictness of the step-size bound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi —\n\nThe short version: this note has a real insight, but the central equivalence is not established as written. I think the gap is fixable and the result is probably correct, so it deserves a serious referee—not a desk reject.\n\nWhat is actually new: the paper reads Chen-Teboulle as the proximal point algorithm under the scaled norm (12), and derives the step-size condition λ ≤ 1/√(‖A‖²+1), which is looser than the original λ < 1/(2L) from [CT94]. That bound is a genuine improvement for a method the author honestly describes as “not widely used.” The scaled-norm technique is credited to [HY12], so the novelty is the application to Chen-Teboulle and the sharper condition. The algebra up to the Schur complement test is clean and self-contained; no fitted parameters, no overclaiming about numerical behavior.\n\nWhere it softens: Section 2.1 says that after “re-organizing the steps” the PPA scheme (13)–(15) is Chen-Teboulle. That is the load-bearing step, and the paper does not prove it. The reader checked a trivial case—f=g=0, A=I, λ=1/2—and gets different iterates from (5)–(8) and (13)–(15). I rechecked; the reader is right. The displayed equations are not the same map on the same variables. The stress-test note shows how to fix it: the PPA dual variable should be the Chen-Teboulle predictor p_{k+1} = y_k + λ(Ax_k − z_k), not y_k. With that shift, (13)–(15) do correspond to (5)–(8). But the paper never states this transformation; its claim that p^{k+1} equals y_{k+1} is off by one index. So the transfer of the PPA convergence and the bound (17) to Chen-Teboulle is not proven as written.\n\nThis is a serious but narrow flaw, not a sign of incoherence. The paper is honest about prior work and the derivation is reproducible. The fix is a few paragraphs. The lack of numerical experiments is fine for a theoretical note; the contribution is conceptual.\n\nWho gets value: anyone working on primal-dual splitting methods, especially the lineage from Chen-Teboulle through Chambolle-Pock, Condat-Vu, and related methods. The improved step-size bound is practically interesting if the equivalence holds.\n\nRecommendation: send to review. The editor should not desk-reject; the idea is sound and the flaw is repairable. I would ask for a major revision that reworks Section 2.1 with an explicit change of variables and a correct index. If that lands, the paper would be a useful short contribution.","headline":"A real conceptual link with a repairable proof gap; the step-size bound is likely correct but the equivalence needs an explicit variable shift.","tokens_in":4821,"tokens_out":3509,"would_cite":false,"duration_ms":34687,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C25","47H05","65K05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The Chen-Teboulle algorithm is the proximal point algorithm carried out in a weighted norm, and this identification permits a larger step size.","keywords":["Chen-Teboulle algorithm","proximal point algorithm","primal-dual splitting","weighted norm","step-size bound","monotone operator","convex optimization","resolvent"],"falsifier":"Take $f=0$, $g=0$, $A=I$, $\\lambda=1/2$, start both algorithms at the same triple $(x_0,z_0,y_0)$, and compare the first iterates of equations (5)–(8) with those of (13)–(15). The paper's identification predicts identical iterates; a direct computation shows they differ, so this example either falsifies the claimed equality or forces a precise restatement of the 're-organizing the steps' equivalence.","tokens_in":3833,"feed_emoji":"⚖️","tokens_out":9280,"duration_ms":90917,"temperature":0.7,"pith_summary":"The paper argues that the Chen-Teboulle splitting method for minimizing $f(x)+g(Ax)$ is not a standalone predictor-corrector scheme but a special case of the proximal point algorithm. The special case is obtained by choosing a weighted norm whose metric $V$ is engineered to decouple the $x$ and $z$ updates of the monotone inclusion. If the identification is correct, Chen-Teboulle inherits the full convergence theory of proximal point methods, and its step-size constraint improves from $\\lambda<1/(2L)$ to $\\lambda\\le 1/\\sqrt{\\|A\\|^2+1}$. A sympathetic reader would care because the same weighted-norm viewpoint turns one algorithm into a template for constructing and analyzing primal-dual splitting methods.","feed_headline":"Chen-Teboulle is the proximal point algorithm in a weighted norm","feed_subtitle":"This identification widens the allowable step size, so the splitting method can run with fewer iterations.","key_machinery":"The load-bearing object is the metric operator $V$ of equation (12): a positive-definite block matrix with diagonal blocks $\\tau_x I$, $\\tau_z I$, $\\tau_y I$ and off-diagonal blocks chosen so that $V+\\mathcal A$ has zero upper-triangular blocks. That cancellation makes the resolvent block-separable, so the $x$ and $z$ updates decouple and the $y$ update is a simple affine correction, which is precisely the predictor-corrector structure of Chen-Teboulle. The same $V$ carries the convergence argument: applying the Schur complement test to $V\\succ 0$ produces the improved step-size condition.","core_discovery":"The paper's central claim is an identity between two iteration rules. The saddle-point conditions for (1) are written as a monotone inclusion $0\\in\\mathcal A(x,z,y)$ with $\\mathcal A(x,z,y)=(\\partial f(x)+A^*y,\\ \\partial g(z)-y,\\ z-Ax)$. Choosing the Hermitian positive-definite metric $V$ in equation (12) cancels the upper-triangular coupling of $\\mathcal A$, so the resolvent $(V+\\mathcal A)^{-1}$ separates into independent proximal steps in $x$ and $z$ plus an affine update in $y$. The proximal point iteration $\\mathbf{x}_{k+1}=(V+\\mathcal A)^{-1}V\\mathbf{x}_k$ then reproduces the Chen-Teboulle updates after a re-ordering of the steps. Under that identification, the condition $V\\succ 0$ yields the refined step-size bound $\\lambda\\le 1/\\sqrt{\\|A\\|^2+1}$, replacing the original $\\lambda<1/(2L)$.","pith_inferences":["The hinge of the argument is the 're-organizing the steps' step in Section 2.1; my own spot check on $f=g=0$, $A=I$, $\\lambda=1/2$ gives different first iterates for (13)–(15) versus (5)–(8), so the equivalence is a substantive claim to verify rather than a notational relabeling.","If the identity holds, the Schur-complement condition in (16) is a general recipe: other primal-dual splittings correspond to other choices of the off-diagonal block $B$, and each would inherit a step-size bound of the form $\\tau_y I\\succ\\tau_x^{-1}AA^*+\\tau_z^{-1}BB^*$.","A practical test of the larger bound would be to run Chen-Teboulle with $\\lambda$ between $1/(2L)$ and $1/\\sqrt{\\|A\\|^2+1}$ on a small quadratic program; the old theory predicts possible divergence and the new theory predicts convergence, so the experiment separates the two claims."],"forward_implications":["For $\\|A\\|\\ge 1$, the admissible step size grows from roughly $1/(2\\|A\\|)$ to roughly $1/\\|A\\|$, so the method can take larger steps without leaving the convergence regime.","Because the algorithm is an instance of proximal point, it inherits weak convergence guarantees and allowance for inexact evaluations of the proximal steps.","Distinct weights $\\tau_x\\neq\\tau_z\\neq\\tau_y$ are permitted; the paper highlights $\\tau_x=\\|AA^*\\|$, $\\tau_z=1$, $\\tau_y<1/2$ as a usable parameter choice.","The same weighted-metric construction extends the analysis to more general compositions such as $f(E_1x+b_1)+g(E_2x+b_2)$, as the paper states follows from the viewpoint."],"supporting_citations":[{"why":"Introduces the Chen-Teboulle algorithm and the original conservative step-size condition $\\lambda<1/(2L)$ that the paper refines.","marker":"[CT94]"},{"why":"Supplies the product-space weighted-norm viewpoint that the paper adapts to re-derive Chen-Teboulle as a proximal point method.","marker":"[HY12]"},{"why":"Provides the monotone operator, resolvent, and proximal point theory used to identify the algorithm and transfer convergence guarantees.","marker":"[BC11]"},{"why":"Supplies the Fenchel-Rockafellar duality framework in which the primal-dual problem (1)-(2) is posed.","marker":"[Roc70]"}],"fun_headline_variants":["Chen-Teboulle is proximal point in disguise","Step-size bound sharpened via proximal point link","Proximal point identity widens Chen-Teboulle step","Chen-Teboulle equals proximal point, better step","Splitting method reveals proximal point nature"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the proximal point updates (13)–(15) are, after re-ordering the steps, exactly the Chen-Teboulle updates (5)–(8); if those two iteration rules are not the same, the improved step-size bound does not transfer to Chen-Teboulle.","fun_headline_variants_meta":{"raw":{"variants":["Chen-Teboulle is proximal point in disguise","Step-size bound sharpened via proximal point link","Proximal point identity widens Chen-Teboulle step","Chen-Teboulle equals proximal point, better step","Splitting method reveals proximal point nature"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000152,"raw_usage":{"total_tokens":1109,"prompt_tokens":754,"completion_tokens":355,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":370,"completion_tokens_details":{"reasoning_tokens":278}},"tokens_in":370,"tokens_out":355,"duration_ms":4291,"temperature":1.0,"reasoning_tokens":278,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:08:49.764775+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $f=0$, $g=0$, $A=I$, $\\lambda=1/2$, start both algorithms at the same triple $(x_0,z_0,y_0)$, and compare the first iterates of equations (5)–(8) with those of (13)–(15). The paper's identification predicts identical iterates; a direct computation shows they differ, so this example either falsifies the claimed equality or forces a precise restatement of the 're-organizing the steps' equivalence.","supporting_citations":[],"review_version":1}