{"id":"ea9076a7-f35f-4833-b0be-e8a8e6fa14a7","arxiv_id":"1908.10426","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"The iteration-stability bound claimed for the explicit-implicit Euler discretization of the Euler-Lagrange acceleration ODE is not proven and is false for some parameter values permitted by the theorem.","lead":"The paper analyzes Euler discretizations of an optimization acceleration ODE and claims a bound on how long the naive explicit-implicit method stays stable. The main theorem's proof contains a logical error and the bound fails for some allowed parameter values.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.1(b) is unsupported: Appendix A.4 treats a necessary condition as sufficient, and a parameter set satisfying the hypotheses contradicts the bound.","rationale":"The reader's weakest-assumption identification is precise: the proof of Theorem 3.1(b) in Appendix A.4 replaces the exact stability inequality y + sqrt(y^2 - 4x) <= 4 with the weaker condition 2*sqrt(x) < 4, relying on a one-way implication. This is indeed the load-bearing step, because the entire theorem's sufficient condition depends on that inversion. My own re-derivation of the eigenvalue formulas confirms that the condition k < (4/(C A p^2 epsilon))^(1/(p-2)) does not imply x < 4 once the exact expression x = (p/k) * C p epsilon (k+1)^(p-1) A is used; the proof's replacement of (k+1)^(p-1) by k^(p-1) only compounds the gap. The explicit counterexample with p=3, L=10, delta=0.099, C=10.3, k=4 falls inside the claimed bound yet yields a spectral radius of about 4.31, so the theorem is false as stated. The implicit and explicit Euler analyses are not implicated by this concern. Since the central claim is unsound and the paper's main contribution is that bound, the appropriate verdict remains REJECT; my stress test does not change the reader's conclusion.","tokens_in":9806,"tokens_out":7945,"duration_ms":72025,"concrete_test":"Evaluate the exact spectral radius of M_k in Eq. (3.3) for p=3, L=A=10, delta=0.099, epsilon=delta^3, C=10.3, k=4, using the eigenvalue formulas (A.1)-(A.2). If the spectral radius exceeds 1, as direct computation gives approximately 4.31, then the theorem's bound k < (4/(CLp^2 epsilon))^(1/(p-2)) = 4.45 does not guarantee stability. Confirm by running the two-step recurrence (3.1) for twenty iterations and observing divergence from the minimizer.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central load-bearing claim is Theorem 3.1(b), which asserts stable behavior for p>2 whenever k < (4/(CLp^2 epsilon))^(1/(p-2)). The proof in Appendix A.4 derives the exact stability condition y + sqrt(y^2 - 4x) <= 4, with x = a_k b_k and y = x + a_k = x + p/k. It then argues: 'By the Triangle Inequality, y + s2 <= 4 implies s1 < 4. Thus, we have that stability or convergence is achieved when s1 < 4.' This is a logical inversion: the implication runs only from the exact condition to s1 < 4, not conversely. Consequently, the derived bound is necessary, not sufficient, for the condition a_k b_k < 4 in which the proof places its faith. Because y = x + a_k exceeds x, there are regimes where x < 4 yet y + sqrt(y^2 - 4x) > 4, so divergence occurs before the bound is reached. The error is not merely cosmetic: taking p=3, L=A=10, delta=0.099, epsilon=delta^3, C=10.3, and k=4 satisfies every stated hypothesis and gives k < (4/(CLp^2 epsilon))^(1/(p-2)) = 4.45, but the iteration matrix M_4 in Eq. (3.3) has eigenvalues approximately -0.058 and -4.31, so the spectral radius exceeds 1 and the iterates diverge. This directly contradicts Theorem 3.1(b). The abstract and theorem also disagree over the presence of C in the bound, and the proof's replacement of (k+1)^(p-1) by k^(p-1) further weakens the claimed sufficient condition.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes stability of Euler discretizations of the Euler-Lagrange ODE (1.2) used to model accelerated optimization methods. The central result is Theorem 3.1, which claims that for a strongly convex L-smooth quadratic objective, step size δ < 1/L, ε = δ^p, and C < 1/(εL), the explicit-implicit Euler discretization is stable for p = 2 and, for p > 2, exhibits stable behavior while k < (4/(CLp^2 ε))^{1/(p-2)}. The paper also analyzes the implicit Euler discretization (claimed converging) and the explicit Euler discretization (claimed diverging), with numerical experiments in Section 5.","tokens_in":10090,"tokens_out":11966,"duration_ms":106569,"significance":"The question addressed is of genuine interest: explaining and bounding the divergence of the naive discretization of the Wibisono et al. ODE could clarify when such discretizations are usable. The paper is clearly written, and the reduction to a scalar eigenvalue problem via eigendecomposition (Proposition 2.1) is standard and reasonable. However, the central stability bound is false as stated. I provide below an explicit parameter set satisfying every hypothesis of Theorem 3.1 for which the claimed bound holds while the iteration matrix has spectral radius greater than one. The flaw is not a minor technicality; it is a logical inversion in Appendix A.4 that invalidates Theorem 3.1(b). Because that theorem is the main contribution, the manuscript is not publishable in current form.","major_comments":[{"comment":"Equation (A.4) gives the divergence condition y + sqrt(y^2 - 4x) > 4, where x = a_k b_k and y = x + a_k. The text then states that by the triangle inequality, y + s2 <= 4 implies s1 = 2 sqrt(x) < 4, and concludes that 'stability or convergence is achieved when s1 < 4.' This is a one-way implication: y + s2 <= 4 does imply s1 < 4, but s1 < 4 does not imply y + s2 <= 4 because y = x + a_k > x. The exact stability condition is not s1 < 4; it is y + sqrt(y^2 - 4x) <= 4. Thus the bound in Theorem 3.1(b) is derived from a necessary condition treated as sufficient. A concrete counterexample: take p = 3, A = L = 10, δ = 0.099, ε = δ^3, C = 5.73, and k = 4. All hypotheses of Theorem 3.1 are satisfied (δ < 1/L, k > p, C < 1/(εL)), and the claimed bound gives k < (4/(CLp^2 ε))^{1/(p-2)} = 8, so k = 4 is within the bound. For these parameters, a_k b_k ≈ 3.125 < 4, but y + sqrt(y^2 - 4x) ≈ 5.46 > 4; the matrix M_4 in Eq. (3.3) has eigenvalues approximately -0.14 and -1.73, so the spectral radius exceeds 1 and the iterates diverge. This directly contradicts Theorem 3.1(b).","section":"Appendix A.4, Eq. (A.4)"},{"comment":"The asymptotic analysis for p > 2 is not legitimate because the matrices M_k in Eq. (3.3) contain entries proportional to (k+1)^{p-1} and p/k; for p > 2 these entries do not converge to a finite matrix. The characteristic equation λ^2 + λ(c - 2) + 1 with c = lim_{k→∞} C p^2 ε k^{p-2} A therefore relies on a divergent limit, since c → ∞ when p > 2. This part of the proof should be revised or removed before any end-behavior claim for p > 2 can be made.","section":"Section A.2"},{"comment":"The abstract states the stability bound as k < (4/(L p^2 δ^p))^{1/(p-2)}, without the constant C, while Theorem 3.1(b) states k < (4/(C L p^2 ε))^{1/(p-2)} with ε = δ^p. Since C is a free parameter subject only to C < 1/(εL), these are different claims and the discrepancy can be arbitrarily large. The statement of the main result must be internally consistent before the theorem can be evaluated.","section":"Abstract and Theorem 3.1(b)"}],"minor_comments":[{"comment":"The notation 'p−2√' for the (p-2)-th root is confusing; it should be written as (·)^{1/(p-2)}.","section":"Section 1.4, table"},{"comment":"The definition of stability requires |x_k - x*| = |x_{k+1} - x*| for all sufficiently large k, but the proof of Proposition 2.2(b) only shows that the eigencomponent along an eigenvalue of magnitude 1 stays constant in magnitude. Complex conjugate eigenvalues on the unit circle rotate the vector, so the stated equality of consecutive absolute errors is not established.","section":"Section 2, Definition 2.1 and Proposition 2.2"},{"comment":"The tabulated bound 44,445 cannot be reproduced from the formulas in the abstract or Theorem 3.1 without specifying the value of C; with the abstract's formula the strict inequality gives 44,444 as the largest integer k for the stated parameters.","section":"Section 5.1, Figure 2"}],"recommendation":"reject","confidential_remarks":"The counterexample in the report is checkable by direct computation; I verified the spectral radius numerically. The paper appears to be an early-stage report, and the main theorem is false as stated, so I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, the punchline is that Theorem 3.1(b) is false as stated. In Appendix A.4 the authors derive the exact stability condition y + sqrt(y^2 - 4x) <= 4, then argue that by the triangle inequality this implies s1 < 4, and conclude stability when s1 < 4. That is a one-way implication: s1 < 4 is necessary, not sufficient. A concrete parameter set satisfying all the theorem's hypotheses (p=3, L=10, delta=0.099, C=10.3, k=4) gives an iteration matrix with spectral radius above 1, so the claimed bound does not guarantee stable behavior. The abstract and the theorem also disagree over whether C belongs in the bound, and the proof's replacement of (k+1)^(p-1) by k^(p-1) is another step in the wrong direction. These are not cosmetic issues; they invert the logic of the central claim.\n\nWhat is genuinely new and useful: the paper addresses a real open puzzle—Wibisono et al. observed divergence of the naive explicit-implicit discretization but gave no quantitative handle on when it happens. The matrix formulation for all three Euler schemes is clean, and the implicit Euler analysis (Theorem 4.1) is correct and standard. The one-dimensional reduction via diagonalization is fine. The citation pattern is normal and gives prior work proper credit.\n\nThe numerical experiments are presented as confirmation of the bound, but since the bound is not a sufficient condition, the figures cannot validate it. They do show the expected qualitative behavior for the implicit and explicit methods, which is consistent with the correct parts of the paper.\n\nWho gets value from this? Someone working on continuous-time analyses of accelerated methods might find the implicit/explicit material worth a footnote, but the explicit-implicit result should not be cited or relied on. If this lands on your desk, I would send it back for major revision: either derive a genuine sufficient condition from the exact spectral inequality, or clearly reframe the bound as a necessary condition. The paper deserves referee time because the question is real and part of the analysis is sound, but the main theorem as written cannot stand.","headline":"The central stability bound is unsupported—the proof in A.4 uses a necessary condition as if it were sufficient—and the paper needs major revision, though the implicit/explicit Euler analysis has some value.","tokens_in":10642,"tokens_out":3826,"would_cite":false,"duration_ms":40869,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65K05","90C25"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a cheap explicit-implicit Euler discretization of the accelerated-optimization Euler-Lagrange ODE remains stable up to an explicit, parameter-dependent iteration bound.","keywords":["Euler-Lagrange ODE","accelerated optimization","explicit-implicit Euler","discretization stability","strongly convex quadratic","stability function","iteration bound","continuous-time optimization"],"falsifier":"Take a one-dimensional quadratic with $p=3$, $L=10$, $\\delta=0.01$, choose $C<1/(\\epsilon L)$, and evaluate the exact eigenvalues of $M_k$ at $k$ around $44,445$. If the exact stability inequality $y+\\sqrt{y^2-4x}\\le 4$ fails at or below the claimed bound (equivalently, $R(M_k)>1$), the theorem's guarantee is false; a direct run of the update equations would settle the question empirically.","tokens_in":9516,"feed_emoji":"⏳","tokens_out":11124,"duration_ms":107652,"temperature":0.7,"pith_summary":"This paper asks why the natural explicit-implicit Euler discretization of a continuous-time accelerated optimization method eventually diverges even on strongly convex quadratics. It argues the naive scheme is not simply broken: for p=2 the update is stable in the limit, and for p>2 it provably stays stable for $k < (\\frac{4}{CLp^2 \\epsilon})^{1/(p-2)}$ iterations when $\\epsilon=\\delta^p$, $\\delta<1/L$, and $C<1/(\\epsilon L)$. That matters because this discretization is far easier to implement than the third-sequence 'rate-matching' discretization proposed elsewhere, so a usable stability window makes the cheap scheme a practical alternative. The same spectral analysis is applied to the implicit and explicit Euler methods, giving convergence for the implicit method and divergence for the explicit method.","feed_headline":"Accelerated-optimization scheme is stable only up to a provable bound","feed_subtitle":"The bound tells users how many iterations a cheap explicit-implicit Euler update can run before divergence.","key_machinery":"The central object is the stability function $R(M_k)=|\\lambda_{\\max}(M_k)|$ and its limit $R(M_\\infty)=\\lim_{k\\to\\infty} R(M_k)$, which classifies end behavior: $R<1$ converges, $R=1$ is stable, and $R>1$ diverges. For the explicit-implicit method the iteration matrix is $$M_k=\\begin{pmatrix} 1-a_k & a_k \\\\ -b_k+a_k b_k & 1-a_k b_k \\end{pmatrix}$$ with $a_k=p/k$ and $b_k=C p \\epsilon (k+1)^{p-1} A$. The proof computes its eigenvalues, reduces the divergence threshold to $y+\\sqrt{y^2-4x}\\le 4$ with $x=a_k b_k$ and $y=x+a_k$, and then converts the threshold into the iteration bound $a_k b_k<4$. A prior diagonalization step justifies doing the analysis in one dimension for a general symmetric positive definite $A$, and the bound is stated using the largest eigenvalue, the smoothness constant $L$.","core_discovery":"The central claim is that, for a strongly convex $L$-smooth quadratic $f$, the explicit-implicit Euler discretization of $$\\ddot{X}_t + \\frac{p+1}{t}\\dot{X}_t + $Cp^{2}$ $t^{{p-2}}$\\nabla f(X_t)=0$$ has a computable epoch of reliable behavior. For $p=2$ the spectral radius of the iteration matrix tends to $1$, so the iterates oscillate around the minimizer without diverging; for $p>2$, stability holds while $k<\\left(\\frac{4}{C L p^2 \\epsilon}\\right)^{1/(p-2)}$, with $\\epsilon=\\delta^p$, $\\delta<1/L$, and $C<1/(\\epsilon L)$. The authors derive this by writing the update as $[x_{k+1};\\,z_{k+1}]=M_k[x_k;\\,z_k]$, computing the eigenvalues of $M_k$, and identifying where the largest eigenvalue crosses magnitude $1$. They conclude from the same calculation that implicit Euler converges for all $p$ and explicit Euler diverges for all $p$.","pith_inferences":["The paper does not prove its bound is sharp; solving the exact threshold $y+\\sqrt{y^2-4x}=4$ numerically would reveal how much conservatism the triangle-inequality simplification adds and whether the true divergence time is longer.","For non-quadratic smooth objectives, linearizing $\\nabla f$ at the current iterate turns the same one-dimensional spectral calculation into a local, iteration-dependent stability estimate, an extension the paper only mentions.","The numerical observation that an explicit fourth-order Runge-Kutta discretization survives longer before diverging suggests the same stability-function analysis could bound Runge-Kutta variants.","If the implicit Euler scheme's apparent acceleration holds in general, it offers a backward-stable route to realizing the $O(1/t^p)$ rate in discrete time; the paper reports the numerics but does not prove the rate."],"forward_implications":["For $p>2$, a practitioner running the naive explicit-implicit discretization with $\\delta<1/L$ is guaranteed stable behavior through roughly $(4/(C L p^2 \\delta^p))^{1/(p-2)}$ iterations, after which divergence is possible.","For $p=2$ the naive discretization is claimed not to diverge at all in the strongly convex quadratic setting, matching the stable end behavior of the underlying ODE.","The implicit Euler discretization converges for every $p$ and appears numerically to keep the accelerated rate, but each step requires solving an implicit system, which is the practical obstacle.","The explicit Euler discretization diverges for every $p$, so it cannot be used for long horizons; the explicit-implicit scheme is the recommended middle ground.","In higher dimensions only the largest eigenvalue of $A$ enters the bound, so a single $L$-smoothness constant controls the stability window."],"supporting_citations":[{"why":"Supplies the Euler-Lagrange ODE, its $O(1/t^p)$ convergence rate, and the observed divergence of the naive discretization that motivates the stability question.","marker":"[8]"},{"why":"Provides the notion of stiffness and the A-stability distinction used to frame why explicit Euler diverges and implicit Euler converges.","marker":"[2]"},{"why":"Supplies recent stabilized gradient descent ideas that the end-behavior analysis adapts.","marker":"[1]"}],"fun_headline_variants":["Stability window computed for explicit-implicit Euler optimizer","Iteration cap guarantees stability for accelerated method","For p>2, accelerated optimizer diverges after k steps bound","Implicit Euler converges; explicit diverges in accelerated scheme"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result rests on assuming that a simplified product inequality is enough to guarantee stability, but the proof's triangle-inequality step only shows the reverse direction—that exact stability implies the simpler inequality—so the sufficiency of the simpler condition is the load-bearing unproved premise.","fun_headline_variants_meta":{"raw":{"variants":["Stability window computed for explicit-implicit Euler optimizer","Iteration cap guarantees stability for accelerated method","For p>2, accelerated optimizer diverges after k steps bound","Implicit Euler converges; explicit diverges in accelerated scheme"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1537,"prompt_tokens":987,"completion_tokens":550,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":485}},"tokens_in":603,"tokens_out":550,"duration_ms":5737,"temperature":1.0,"reasoning_tokens":485,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:45:39.809551+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a one-dimensional quadratic with $p=3$, $L=10$, $\\delta=0.01$, choose $C<1/(\\epsilon L)$, and evaluate the exact eigenvalues of $M_k$ at $k$ around $44,445$. If the exact stability inequality $y+\\sqrt{y^2-4x}\\le 4$ fails at or below the claimed bound (equivalently, $R(M_k)>1$), the theorem's guarantee is false; a direct run of the update equations would settle the question empirically.","supporting_citations":[{"cited_title":"Wilson, and Michael I","cited_arxiv_id":null,"evidence_quote":"Supplies the Euler-Lagrange ODE, its $O(1/t^p)$ convergence rate, and the observed divergence of the naive discretization that motivates the stability question."},{"cited_title":"Stiffness of odes","cited_arxiv_id":null,"evidence_quote":"Provides the notion of stiffness and the A-stability distinction used to frame why explicit Euler diverges and implicit Euler converges."},{"cited_title":"Explicit Stabilised Gradient Descent for Faster Strongly Convex Optimisation","cited_arxiv_id":"1805.07199","evidence_quote":"Supplies recent stabilized gradient descent ideas that the end-behavior analysis adapts."}],"review_version":1}