{"id":"b144b829-c581-4fc3-8cc6-b058c48ec77c","arxiv_id":"2411.11747","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Anisotropic Gaussian smoothing with step-dependent covariance matrices is inserted into GD, SGD, and Adam, and convergence bounds are derived that generalize the isotropic case.","lead":"This paper replaces the standard gradient in GD, SGD, and Adam with a gradient of an anisotropically Gaussian-smoothed function, using a possibly different covariance matrix at each step. It proves convergence bounds for the smoothed methods and tests one variant against CMA on five benchmark functions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The stated assumptions on smoothing matrices are insufficient: bounded (or merely norm-decaying) Sigma_t does not make the stated error terms vanish, so the advertised convergence to minimizers/stationary points of f is unsupported.","rationale":"I read the paper as attempting to prove finite-time convergence bounds for anisotropic Gaussian smoothing variants of GD, SGD, and Adam, then to conclude convergence of the true gradient. The descent-based architecture of Theorems 3.1 and 3.2 is reasonable, and the finite-time statements can be made correct if one posits a valid switching bound B. The load-bearing defect is at the level of the smoothing-schedule hypotheses. The paper repeatedly asserts that boundedness of ||Sigma_t|| makes the right-hand sides vanish as T grows, which is arithmetically false for constant positive Sigma_t, and the switching terms require explicit decay or summability of B. For AGS-Adam, the final passage from grad f_{Sigma_t}(x_t) -> 0 to grad f(x_t) -> 0 requires ||Sigma_t||^2 ||Sigma_t^{-1}|| -> 0 by Lemma 2.3(b); this is not implied by ||Sigma_t|| -> 0. These are not cosmetic issues: they are precisely the conditions linking the smoothed iterates to the original problem. The experiments in Section 6 use a CMA update on unimodal benchmark functions and provide no verification of these conditions, so they do not supply independent support for the asymptotic claims. No machine-checked proof or reproducible theoretical artifact offsets this gap. A corrected version would need to state explicit decay and conditioning assumptions on Sigma_t, prove the switching bounds under those assumptions, and show the chosen CMA schedule satisfies them. Therefore I see no reason to change the reader's rejection.","tokens_in":21414,"tokens_out":20790,"duration_ms":195880,"concrete_test":"Set Sigma_t = epsilon I_d (epsilon > 0) in the bound of Theorem 3.1 for an L-smooth convex f whose smoothed minimizer differs from the true one, e.g. f(x) = ||x||^2/2 + a sin(x_1) with |a| small. Letting T go to infinity, the theorem's right-hand side tends to (Ld/4)epsilon^2 > 0, directly disproving the claim after Theorem 3.2 that boundedness of (Sigma_t) suffices for convergence to f(x*). Alternatively, verify the same failure in Theorem 4.1 by taking eta_t = 1/t and B(Sigma_t, Sigma_{t-1}) = c > 0, for which the switching term tends to c rather than 0.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The convergence theorems are not supported by the stated assumptions on (Sigma_t). In Theorem 3.1 the bound contains (Ld/(4T)) sum_{t=1}^T ||Sigma_t||^2 and (1/T) sum_{t=1}^{T-1} t B(Sigma_{T-t}, Sigma_{T-t+1}); under only boundedness these need not vanish. For Sigma_t = epsilon I, the first term tends to (Ld/4)epsilon^2 > 0, and the switching term vanishes only if B itself decays. The same defect appears in Theorem 4.1, where sum B(Sigma_t, Sigma_{t-1}) / sum eta_t can be bounded away from 0, e.g. with B = c > 0 and eta_t = 1/t. In Theorem 5.1, even the assumption ||Sigma_t|| -> 0 is insufficient for the final step: Lemma 2.3(b) controls ||grad f_Sigma(x) - grad f(x)|| by L ||Sigma||^2 ||Sigma^{-1}||, so one needs ||Sigma_t||^2 ||Sigma_t^{-1}|| -> 0, which does not follow from ||Sigma_t|| -> 0 (e.g. Sigma_t = diag(t^{-3}, t^{-1}) gives product ~ t^2). The CMA schedule in Section 6 is not shown to satisfy any such condition, so the numerical experiments do not validate the asymptotic claims.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces anisotropic Gaussian smoothing as a modification of GD, SGD, and Adam. In place of the gradient of f, the algorithms use the gradient of the Gaussian-smoothed function f_Sigma_t defined in (2), with a time-varying symmetric invertible matrix Sigma_t, and the paper presents finite-time bounds for convex and non-convex L-smooth objectives (Theorems 3.1 and 3.2), a stochastic bound for AGS-SGD (Theorem 4.1), and an almost-sure stationarity result for AGS-Adam (Theorem 5.1). The bounds are expressed with a generic switching penalty B(Sigma,T) bounding |f_Sigma(x)-f_T(x)|, and the paper claims that, as T tends to infinity, the right-hand sides tend to zero whenever the sequence (||Sigma_t||) is bounded, thereby recovering the unsmoothed GD/SGD/Adam results as Sigma_t tends to 0. The numerical section replaces nabla f_Sigma_t by the Monte Carlo estimator (7) and generates Sigma_t with a CMA-type update, testing five benchmark functions.","tokens_in":21736,"tokens_out":28969,"duration_ms":243877,"significance":"The generalization from isotropic (scalar or identity) smoothing to a time-varying matrix Sigma_t is natural and potentially useful: it allows direction-dependent smoothing, and the abstract formulation of the switching cost B(Sigma,T) cleanly isolates the new difficulty. The paper is written in the proof-template style of the authors' prior isotropic work [21,22]; the appendices give complete proofs of the background lemmas, and the constants in the bounds are explicit. If the convergence claims were supported by the stated assumptions, this would be a solid contribution to the smoothing and zero-order optimization literature. However, the central asymptotic claims are not merely underspecified: they are false under the stated hypotheses (see major comments), and the proof of Theorem 5.1 does not establish its conclusion. The experimental section demonstrates the behavior of a heuristic implementation (CMA plus Monte Carlo) that the theorems do not cover, so it does not compensate for the theoretical gaps.","major_comments":[{"comment":"The assertion that the right-hand sides of Theorems 3.1, 3.2, and 4.1 converge to zero whenever (||Sigma_t||) is bounded is false. In Theorem 3.1, the term (Ld/(4T)) sum_{t=1}^T ||Sigma_t||^2 with the constant sequence Sigma_t = sigma I converges to Ld sigma^2 / 4 > 0 rather than to 0, and the switching term (1/T) sum_{t=1}^{T-1} t B(Sigma_{T-t}, Sigma_{T-t+1}) equals (1/T) sum_{s=1}^{T-1} (T-s) B(Sigma_s, Sigma_{s+1}); for Sigma_t alternating between two fixed matrices with different spectra, B is a positive constant c and this term grows like c(T-1)/2, so the bound diverges. Even under B(Sigma_t, Sigma_{t+1}) = O(1/t), the weighted mean is O(log T) and does not vanish. In Theorem 3.2, the second term requires (1/T) sum_{t=1}^T ||Sigma_t||^4 ||Sigma_t^{-1}||^2 to tend to 0, and the third term requires (1/T) sum_{t=0}^T B(Sigma_{t+1}, Sigma_t) to tend to 0; boundedness of (||Sigma_t||) alone gives neither, especially since ||Sigma_t^{-1}|| is uncontrolled. In Theorem 4.1, with eta_t = 1/t and B identically c > 0, the final term sum_{t=1}^T B(Sigma_t, Sigma_{t-1}) / sum_{t=1}^T eta_t behaves as c T / log T and diverges. The theorems as stated therefore do not support the advertised convergence to the minimizer in the convex case or to stationary points in the non-convex case; explicit decay conditions, such as ||Sigma_t|| tending to 0 in the relevant Cesaro sense and weighted summability of B(Sigma_t, Sigma_{t+1}) (e.g., B = o(1/t)), are missing.","section":"Section 3, after Theorem 3.2; Section 4, after Theorem 4.1"},{"comment":"The final step of the proof, which states that since ||Sigma_t|| -> 0 we have lim_{t->infinity} ||nabla f(x_t)||^2 = 0 almost surely, is not justified. To pass from lim ||nabla f_{Sigma_t}(x_t)|| = 0 to lim ||nabla f(x_t)|| = 0, one must control ||nabla f_{Sigma_t}(x_t) - nabla f(x_t)||, and Lemma 2.3(b) bounds this by L ||Sigma_t||^2 ||Sigma_t^{-1}|| ((3+d)/2)^{3/2}. The assumption ||Sigma_t|| -> 0 does not imply ||Sigma_t||^2 ||Sigma_t^{-1}|| -> 0; for example, for d = 2 and Sigma_t = diag(t^{-3}, t^{-1}), one has ||Sigma_t|| = t^{-1} -> 0 but ||Sigma_t||^2 ||Sigma_t^{-1}|| = t^{-2} * t^3 = t -> infinity. The hypotheses of the theorem must be strengthened (for instance, to ||Sigma_t||^2 ||Sigma_t^{-1}|| -> 0) or an additional argument must be given; as it stands, the proof of Theorem 5.1 is incomplete.","section":"Section 5, proof of Theorem 5.1"},{"comment":"The statement of Theorem 3.1 is inconsistent with its proof regarding the learning rate. The stated bound is (1/(2T)) ||x_0 - x^*||^2, while the proof in Appendix B obtains (1/(2 lambda T)) ||x_0 - x^*||^2; the factor 1/lambda is dropped in the statement. In addition, the proof uses the inequality lambda (L lambda / 2 - 1) <= -lambda/2, which is valid only if lambda <= 1/L, but no step-size restriction appears in Theorem 3.1 or Theorem 3.2. The displayed proof of Theorem 3.2 also loses the 1/lambda factor, and its telescoping step bounds f_{Sigma_1}(x_1) - f_{Sigma_{T+1}}(x_{T+1}) by f(x_0) - f*, which does not follow for non-convex f because the inequality f_Sigma(x) >= f(x) from Lemma 2.1(b) is unavailable in that setting and f_{Sigma_1}(x_0) can exceed f(x_0) by up to (Ld/4) ||Sigma_1||^2.","section":"Section 3, Theorems 3.1 and 3.2 and Appendix B"},{"comment":"The bound in Lemma 2.5(b) is either false or unproved as written. The proof sets Sigma^2 = P D_Sigma P^T with D_Sigma = diag(sigma_1, ..., sigma_d), so sigma_i and tau_i are treated as eigenvalues of Sigma^2 and T^2; but then the estimate with the squared terms ||(D_Sigma - D_T)_+||^2 and ||(D_T - D_Sigma)_+||^2 does not follow from Lemma 2.3(a), because applying that lemma to the smoothing matrix sqrt((D_Sigma - D_T)_+) yields (Ld/4) ||(D_Sigma - D_T)_+|| without the square. If instead sigma_i and tau_i are meant to be eigenvalues of Sigma and T, the claimed bound is false: for f(x) = (1/2)||x||^2, Sigma = 2I, and T = I, the true difference is |f_Sigma(x) - f_T(x)| = 3d/4, while the claimed bound is (Ld/4)(2-1)^2 = d/4. The statement of the lemma does not define sigma_i and tau_i, and no proof in Appendix A establishes the displayed formula.","section":"Section 2, Lemma 2.5(b)"}],"minor_comments":[{"comment":"The text refers to results shown in Figure 1a through Figure 1e for the five benchmark tests, but the benchmark results appear in Figure 2; the caption ordering of Figure 2 also does not match the F1-F5 enumeration in the text, and the names Powell function F18 and Rosenbrock function F6 introduce a third numbering for the same test functions.","section":"Section 6"},{"comment":"The inequality f_{Sigma_t}(x_t) <= f_{Sigma_t}(x_{t-1}) <= f_{Sigma_{t-1}}(x_{t-1}) + B(Sigma_t, Sigma_{t+1}) should read B(Sigma_t, Sigma_{t-1}); the subsequent sum writes ||Sigma_{T+1}|| where ||Sigma_{t+1}|| is meant, and the notation f_{sigma T - t} should be f_{Sigma_{T-t}}.","section":"Appendix B, proof of Theorem 3.1"},{"comment":"The symbol lambda is used both for the step size in Algorithms 1-2 and Theorems 3.1-3.2 and for the variance bound E[||nabla f_k||^2] <= lambda in Theorem 4.1; these two roles should be separated to avoid confusion.","section":"Sections 3 and 4"},{"comment":"The abstract and introduction state that the algorithms approach global minima more effectively, but the theorems only provide conditional convergence to a minimizer in the convex case and to stationary points in the non-convex case; the global-minimum language should be qualified.","section":"Abstract and Introduction"},{"comment":"The experiments implement the Monte Carlo estimator (7) and a CMA-based update for Sigma_t, whereas the theorems concern exact gradients nabla f_{Sigma_t} and require conditions on (Sigma_t) that are not verified for the CMA sequence; a sentence stating that the experiments are heuristic and not covered by the theorems would be appropriate.","section":"Section 6"}],"recommendation":"reject","confidential_remarks":"The manuscript is an incremental extension of the authors' own isotropic-smoothing papers [21,22], and the anisotropic lemmas in Section 2 are the genuinely new content. The main obstacle to publication is not the incremental nature but the fact that the headline convergence claims are contradicted by simple counterexamples (constant or alternating Sigma_t) and that the proof of Theorem 5.1 has a genuine gap. I found no circularity or hidden parameter fitting: the bounds are derived from stated lemmas with explicit constants. The numerical section (CMA plus Monte Carlo) is too disconnected from the theory to count as validation of the theorems. The fit with math.OC is fine, though the abstract's claims about escaping local minima are promotional and not supported by the results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — the thing to know about this paper is that the core idea is genuinely a generalization of isotropic Gaussian smoothing, but the advertised convergence claims are not supported by the assumptions the authors state. The paper replaces the scalar smoothing parameter with a matrix and analyzes GD, SGD, and Adam; the framework is coherent and the proofs are explicit, with no hidden fitted constants. Lemma 2.5(a), the composition result, is new and correct as far as I can see, and the algorithmic extension to CMA-style updates is a sensible direction.\n\nThe soft spots are load-bearing. In Theorems 3.1 and 3.2, the error terms include sums of switching costs B(Sigma_t, Sigma_{t+1}) weighted by t (or summed without decay). The authors claim the bounds vanish when the Sigma_t are bounded, but that is false: if B is constant c, the term in Theorem 3.1 is roughly (c/T) sum t = c(T-1)/2, which diverges; Theorem 3.2 has a term 4/(T lambda) sum B, which tends to a constant, not zero. Theorem 4.1 has the same issue: with B = c and eta_t = 1/t, the ratio sum B / sum eta_t can be bounded away from zero. So the 'converges to 0' statements simply don't follow from boundedness. The fix is to add explicit decay assumptions on B, e.g., B(Sigma_t, Sigma_{t+1}) -> 0 fast enough, but the paper doesn't do that.\n\nIn Theorem 5.1, even ||Sigma_t|| -> 0 is insufficient: the final step needs ||Sigma_t||^2 ||Sigma_t^{-1}|| -> 0 (by Lemma 2.3(b)), and a matrix like diag(t^{-3}, t^{-1}) has that product grow like t^2. The assumption in the statement is just ||Sigma_t|| -> 0, so the conclusion is not justified.\n\nThere is also a concrete linear-algebra error in the proof of Lemma 2.5(b). In the non-commuting case, the authors set H = diag(alpha,...,alpha) with alpha = min eigenvalues, then prove Sigma^2 - alpha I is PSD, but then bound ||Sigma^2 - H^2|| as if H^2 = alpha I. You can't have both: H^2 = alpha^2 I in general, and Sigma^2 - alpha^2 I need not be PSD. The bound might be repairable with H = sqrt(alpha) I, but as written the proof is inconsistent.\n\nFinally, the experiments don't validate the headline claim of escaping local minima: all five test functions (Sphere, Ellipsoidal, Powell, Rosenbrock, Different Powers) are unimodal or convex. And the CMA schedule for Sigma_t is not shown to satisfy any of the decay conditions the theorems need.\n\nSo: the paper is a serious attempt at a useful generalization, and it deserves a serious referee, but as it stands the main theorems need substantial correction. I'd send it to peer review with major revisions, not desk reject, because the framework is sound at the conceptual level and Lemma 2.5(a) is a contribution worth keeping.","headline":"An anisotropic smoothing framework that is a real generalization but whose main convergence theorems don't follow from the stated assumptions.","tokens_in":22250,"tokens_out":4000,"would_cite":false,"duration_ms":34595,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","90C15","90C30"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that replacing gradients with gradients of an anisotropic Gaussian-smoothed function preserves the convergence guarantees of GD, SGD, and Adam while adding a non-local, basin-escaping search.","keywords":["anisotropic Gaussian smoothing","gradient descent","stochastic gradient descent","Adam","CMA","non-convex optimization","L-smooth functions","zero-order optimization"],"falsifier":"Take $f(x)=\\|x\\|^2$ on $\\mathbb{R}^2$ and run AGS-GD with $\\Sigma_t=\\mathrm{diag}(1,1/t)$. Every matrix is symmetric and invertible and $\\|\\Sigma_t\\|=1$ is bounded, but $\\|\\Sigma_t\\|^4\\|\\Sigma_t^{-1}\\|^2=t^2$, so the smoothing term in Theorem 3.2 grows without bound; checking whether the stated boundedness condition really yields a vanishing right-hand side settles the adequacy of the paper's assumptions. A direct empirical counterpart is to run AGS-SGD under the same schedule with decreasing step sizes and record whether $\\|\\nabla f_{\\Sigma_t}(x_t)\\|$ stays away from zero.","tokens_in":21227,"feed_emoji":"🎯","tokens_out":11089,"duration_ms":95797,"temperature":0.7,"pith_summary":"This paper proposes replacing the gradient in gradient descent, stochastic gradient descent, and Adam with the gradient of a Gaussian-smoothed version of the objective, where the covariance matrix of the smoothing kernel may change every iteration, making the smoothing anisotropic. It proves finite-time convergence bounds for AGS-GD on convex and non-convex functions whose gradients are Lipschitz (L-smooth), stochastic convergence to a noisy ball for AGS-SGD, and an almost-sure gradient-norm convergence theorem for AGS-Adam. The results reduce to unsmoothed GD, SGD, and Adam when $\\Sigma_t=0$ and to isotropic Gaussian smoothing when $\\Sigma_t=\\sigma_t I_d$. The motivation is to escape suboptimal local minima by averaging function values over a directional neighborhood rather than looking at one point's gradient, while keeping the guarantees of gradient-based optimization.","feed_headline":"Directional Gaussian smoothing provably converges for GD, SGD, Adam","feed_subtitle":"Changing the smoothing direction each step still gives O(1/T) convergence and an almost-sure Adam limit.","key_machinery":"The load-bearing object is the $\\Sigma$-Gaussian smoothed function $f_\\Sigma(x)=\\pi^{-d/2}\\int_{\\mathbb{R}^d} f(x+\\Sigma u)e^{-\\|u\\|^2}du$ and its gradient $\\nabla f_\\Sigma(x)=2\\pi^{-d/2}\\Sigma^{-1}\\int_{\\mathbb{R}^d} u f(x+\\Sigma u)e^{-\\|u\\|^2}du$, which admits a Monte Carlo estimate from function values only. The proofs rest on three inequalities: Lemma 2.1, Gaussian smoothing preserves L-smoothness, convexity, and Lipschitzness; Lemma 2.3, $|f_\\Sigma(x)-f(x)|\\le (Ld/4)\\|\\Sigma\\|^2$ and $\\|\\nabla f_\\Sigma(x)-\\nabla f(x)\\|\\le L\\|\\Sigma\\|^2\\|\\Sigma^{-1}\\|((3+d)/2)^{3/2}$; and Lemma 2.5, which bounds the switching cost $|f_\\Sigma(x)-f_T(x)|$ by $B(\\Sigma,T)$ and is the new ingredient needed when the smoothing matrix changes direction. The switching lemma is what makes the anisotropic case genuinely different, because two anisotropic kernels cannot generally be composed into a single kernel of the same family unless their covariances are ordered.","core_discovery":"The central discovery is that the convergence theory of gradient-based optimization survives a non-local replacement of the gradient, provided the smoothing matrices change slowly enough. Specifically, the paper shows that when the update uses $\\nabla f_{\\Sigma_t}(x_{t-1})$, the smoothed gradient defined by convolution with an anisotropic Gaussian of covariance $\\Sigma_t^2$, the standard descent arguments go through. For convex L-smooth functions, AGS-GD satisfies $f(x_T)-f(x_*)\\le O(1/T)$ up to a smoothing cost involving $\\|\\Sigma_t\\|^2$ and a switching cost $B(\\Sigma_t,\\Sigma_{t+1})$ that bounds $|f_{\\Sigma_t}(x)-f_{\\Sigma_{t+1}}(x)|$. For non-convex L-smooth functions, the minimum squared gradient norm along the trajectory is bounded by a similar expression involving $\\|\\Sigma_t\\|^4\\|\\Sigma_t^{-1}\\|^2$ plus switching costs. In the stochastic setting, AGS-SGD converges to a noisy ball whose radius is set by the step size, gradient variance, and smoothing parameters, and AGS-Adam achieves $\\lim_{t\\to\\infty}\\|\\nabla f(x_t)\\|^2=0$ almost surely under decay conditions on $\\Sigma_t$ and the switching bounds. Unsmoothed GD, SGD, and Adam, and isotropic Gaussian smoothing, are recovered as special cases.","pith_inferences":["The paper leaves open how to choose $\\Sigma_t$ so the theorems apply; enforcing $\\|\\Sigma_t\\|^2\\|\\Sigma_t^{-1}\\|=O(1/t)$ and $B(\\Sigma_t,\\Sigma_{t+1})=O(\\eta_t^2)$ is a concrete sufficient schedule that would connect the theory to the CMA experiments.","Because Lemma 2.5 says $f_H=(f_\\Sigma)_T$ for $H=\\sqrt{\\Sigma^2+T^2}$, anisotropic smoothing can be read as a homotopy through smoothed landscapes; whether rotating the smoothing axes rather than only shrinking them helps escape basins is a testable question.","A practical byproduct is a stopping criterion: monitor $\\|\\Sigma_t\\|^2\\|\\Sigma_t^{-1}\\|$ to decide when the smoothed gradient has come within a desired tolerance of the true gradient."],"forward_implications":["For convex L-smooth objectives, AGS-GD inherits the $O(1/T)$ rate of gradient descent up to a smoothing term $\\frac{Ld}{4T}\\sum_t\\|\\Sigma_t\\|^2$ and a switching term built from $B(\\Sigma_t,\\Sigma_{t+1})$; both vanish as $T\\to\\infty$ when the smoothing matrices are controlled.","For non-convex L-smooth objectives, the same framework bounds the minimum squared gradient norm along the trajectory, with the smoothing cost written in terms of $\\|\\Sigma_t\\|^4\\|\\Sigma_t^{-1}\\|^2$.","AGS-SGD converges to a noisy ball whose radius is set by the learning rate, the stochastic gradient variance, and the smoothing parameters; a decreasing learning rate drives the smoothed gradient to zero.","AGS-Adam attains $\\lim_{t\\to\\infty}\\|\\nabla f(x_t)\\|^2=0$ almost surely when $\\|\\Sigma_t\\|\\to 0$ and the switching bounds $B$ and $\\tilde{B}$ decay fast enough, with no extra smoothing terms in the final bound.","Because the smoothed gradient admits a Monte Carlo estimate from function values only, all three algorithms operate as zero-order methods, so the convergence guarantees carry over to derivative-free implementations."],"supporting_citations":[{"why":"Supplies Lemma 2.1 and Lemma 2.3: Gaussian smoothing preserves smoothness and convexity and gives the smoothing-error bounds that anchor the convergence proofs.","marker":"[19]"},{"why":"Provides the isotropic GSmooth-GD, GSmooth-SGD, and GSmooth-Adam results that this paper generalizes to matrix-valued smoothing sequences.","marker":"[21]"},{"why":"Gives the stochastic and Adam proof templates and the improved bound for the switching term that AGS-Adam adapts to anisotropic matrices.","marker":"[22]"},{"why":"Supplies the Adam lemmas and almost-sure convergence framework that Theorem 5.1 extends to smoothed gradients.","marker":"[9]"},{"why":"Motivates the CMA covariance-matrix adaptation used in the numerical experiments to produce the smoothing matrices $\\Sigma_t$.","marker":"[7]"}],"fun_headline_variants":["Anisotropic Gaussian smoothing provably improves GD, SGD, Adam","Directional smoothing gives O(1/T) convergence for GD, SGD, Adam","AGS: anisotropic smoothing to escape local minima in gradient methods","Anisotropic smoothing provably accelerates convergence for GD, SGD, Adam","Anisotropic Gaussian smoothing: provable convergence bounds for optimizers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the unstated requirement that the smoothing matrices shrink and switch fast enough for the smoothing error and switching costs to vanish at the right rates, and the paper does not prove that its experimental CMA schedule meets that requirement.","fun_headline_variants_meta":{"raw":{"variants":["Anisotropic Gaussian smoothing provably improves GD, SGD, Adam","Directional smoothing gives O(1/T) convergence for GD, SGD, Adam","AGS: anisotropic smoothing to escape local minima in gradient methods","Anisotropic smoothing provably accelerates convergence for GD, SGD, Adam","Anisotropic Gaussian smoothing: provable convergence bounds for optimizers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001913,"raw_usage":{"total_tokens":7564,"prompt_tokens":1084,"completion_tokens":6480,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":700,"completion_tokens_details":{"reasoning_tokens":6387}},"tokens_in":700,"tokens_out":6480,"duration_ms":47331,"temperature":1.0,"reasoning_tokens":6387,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:12:09.706544+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $f(x)=\\|x\\|^2$ on $\\mathbb{R}^2$ and run AGS-GD with $\\Sigma_t=\\mathrm{diag}(1,1/t)$. Every matrix is symmetric and invertible and $\\|\\Sigma_t\\|=1$ is bounded, but $\\|\\Sigma_t\\|^4\\|\\Sigma_t^{-1}\\|^2=t^2$, so the smoothing term in Theorem 3.2 grows without bound; checking whether the stated boundedness condition really yields a vanishing right-hand side settles the adequacy of the paper's assumptions. A direct empirical counterpart is to run AGS-SGD under the same schedule with decreasing step sizes and record whether $\\|\\nabla f_{\\Sigma_t}(x_t)\\|$ stays away from zero.","supporting_citations":[{"cited_title":"Nesterov and V","cited_arxiv_id":null,"evidence_quote":"Supplies Lemma 2.1 and Lemma 2.3: Gaussian smoothing preserves smoothness and convexity and gives the smoothing-error bounds that anchor the convergence proofs."},{"cited_title":"Completely derandomized self- adaptation in evolution strategies","cited_arxiv_id":null,"evidence_quote":"Motivates the CMA covariance-matrix adaptation used in the numerical experiments to produce the smoothing matrices $\\Sigma_t$."}],"review_version":1}