{"id":"0d3fb124-e534-47b7-b5da-eb954044a121","arxiv_id":"2608.03001","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A new pathwise Lyapunov-Perron framework proves almost sure saddle avoidance for stochastic recursions without unit excitation, covering SGD, mirror descent, proximal stochastic gradient, and random reshuffling.","lead":"This paper proves that a broad class of stochastic optimization methods, including stochastic gradient descent, mirror descent, a proximal method, and random reshuffling, almost surely avoid saddle points without requiring the common 'unit excitation' noise assumption. It gives a new proof framework based on a pathwise Lyapunov-Perron argument, and shows the noise can vanish or lie in low-dimensional subspaces while avoidance still holds.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified.","rationale":"The reader identifies Assumption 2.10 as the weakest assumption, and I agree that it is the most delicate condition in the paper. However, I do not elevate it to a load-bearing objection against the central claim. The assumption is explicitly verified in the two main sampling regimes the paper targets: i.i.d. finite-moment oracles in Proposition 2.11 and without-replacement sampling in Proposition 2.12. The pathwise Lyapunov-Perron argument is carefully structured, with the key algebraic identity (10) and the contraction estimates (Lemma 3.7) checking out. The countable-cover argument in Theorem 2.14 is valid because Z* is covered by countably many balls centered at unstable zeros, and the preimage of each null set under a locally Lipschitz map is null. The only point I found worth a second look is a presentation gap: the proof chooses delta in (17) independently of the seed, but the constants in Lemma 3.3 are stated for a fixed reindexed sequence and could in principle depend on the reindexing start m. In fact, an inspection of Lemma A.2 shows the constants are uniformly bounded for large m because the finite-prefix contribution B_kappa is bounded by the vanishing tail sum r_m and the Jordan block sizes are fixed. Thus the argument is sound, but this uniformity should be stated explicitly. For that reason I recommend keeping the ACCEPT verdict and flagging the uniformity issue as a minor presentational refinement rather than a correctness risk.","tokens_in":41552,"tokens_out":36431,"duration_ms":392310,"concrete_test":"Verify the hidden uniformity behind the seed-independent choice of delta: fix H and the original step-size sequence, then compute the constants C_H from Lemma 3.3 for the reindexed sequences {alpha_{m+j}} over an increasing sequence of m values. If sup_m C_H is finite, the choice of delta in (17) is justified as written; if C_H grows without bound, the proof needs an additional argument or a seed-dependent delta, though the theorem could still be recovered via a seed-dependent countable cover.","verdict_should_be":"UNCHANGED","load_bearing_attack":"After a full pass over the proof of Theorem 2.14, I do not find a load-bearing flaw in the central claim. The pathwise change of variables U_k = I - S_k and the Lyapunov-Perron graph argument in Section 3 are internally consistent, and the countable-cover/Fubini step is valid. Assumption 2.10 is genuinely strong, but it is not an unverified black box: Proposition 2.11 verifies it for i.i.d. finite-moment oracles and Proposition 2.12 verifies it for arbitrary permutation sequences in random reshuffling, covering the paper's claimed applications. The one place I would scrutinize is the seed-independent choice of delta in inequality (17), which presumes that the constants C_H, kappa, nu in Lemma 3.3 are uniform in the reindexing start m; this uniformity follows from the fixed Jordan block sizes and the vanishing tail sum r_m, but it is not stated explicitly in the paper.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the stochastic recursion z_{k+1}=z_k-\\alpha_k G(z_k;\\xi_k) with mean field F, and proves an almost sure avoidance theorem for unstable zeros of F under assumptions that replace unit excitation by pathwise conditions on the accumulated noise and linearization errors (Assumption 2.10). The proof fixes a seed realization, applies the path-dependent change of variables U_k=I-\\sum_{i=k}^\\infty \\alpha_i J_i, and uses a Lyapunov-Perron argument to show that initial states whose trajectories stay near an unstable zero form a Lipschitz graph over the center-stable subspace, hence a Lebesgue null set; Fubini's theorem then gives the almost sure statement. The framework is applied to stochastic mirror descent (Corollary 4.8), a normal map-based proximal stochastic gradient method (Corollaries 4.16 and 4.17), and random reshuffling (Corollary 5.3). The applications verify the oracle assumptions under i.i.d. finite-moment sampling and under without-replacement permutation sequences.","tokens_in":41693,"tokens_out":25459,"duration_ms":273189,"significance":"If correct, the paper is a substantial advance: it removes the unit-excitation assumption that is standard in stochastic saddle avoidance and extends the measure-zero basin mechanism to sampled maps that do not share a common fixed point or common linearization. The pathwise Lyapunov-Perron framework is original and internally consistent, and the proof is unusually detailed; Assumption 2.10, while strong, is explicitly verified for the two principal sampling models in Propositions 2.11 and 2.12. The paper also provides the first asymptotic almost sure strict saddle avoidance statement for random reshuffling without injected perturbations. The central theorem involves no fitted parameters or circular reasoning, and the external convergence theorem [61] used for the NSGD applications is cleanly separable from the avoidance proof.","major_comments":[],"minor_comments":[{"comment":"The seed-independence of the radius \\delta in (17), and hence of \\delta_0 in Theorem 2.13, is asserted but not fully justified: Lemma 3.3 supplies constants C_H, \\kappa, \\nu for a fixed reindexed step-size sequence, while the same \\delta is used for every tail \\{\\alpha_{m+\\ell}\\}_{\\ell\\ge0}. Please add an explicit uniformity argument (e.g., using the fixed Jordan block sizes and the fact that the finite tail sums \\sum_{\\ell=0}^{K-1}\\alpha_{m+\\ell}^2 vanish as m\\to\\infty) to show that C_H, \\kappa, and \\nu can be chosen uniformly over all sufficiently large reindexing starts m. The claim is true, but the present text leaves this point implicit.","section":"Section 3.3-3.4, Eq. (17)"},{"comment":"The step F_\\lambda(z_k)\\to0 almost surely is delegated to [61, Theorem 3.6], but the hypotheses of that theorem are not reproduced. Please either quote the theorem or state precisely how Assumption 4.15 and the step-size condition (31) imply its hypotheses; otherwise the reader cannot verify the conversion from convergence of prox_{\\lambda\\varphi}(z_k) to convergence of z_k in the proof of Corollary 4.16.","section":"Corollary 4.16 and [61, Theorem 3.6]"},{"comment":"The symbol m is used for the reindexing start in Section 3.4 and also for the Jordan block size in Lemma A.2; renaming one of the two uses would avoid confusion for the reader.","section":"Notation, Section 3.4 and Lemma A.2"}],"recommendation":"minor_revision","confidential_remarks":"The only editorial concern is the reliance on the authors' own preprint [61] for a key convergence step in the NSGD application. This is not circular for Theorem 2.14, but I recommend asking the authors to state the status of [61] and to reproduce the exact theorem they invoke, so that Corollaries 4.16 and 4.17 are self-contained enough for verification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a serious paper. It removes unit excitation from stochastic saddle avoidance and gets the first almost-sure avoidance result for random reshuffling. The pathwise Lyapunov-Perron graph argument is new and, as far as I can tell, correct. The main theorem is proven in detail, and the applications verify the pathwise conditions under standard i.i.d. sampling and without-replacement sampling. The comparison table is honest about what Beneventano does and doesn't do.\n\nWhat's good: the four-step proof (change of variables, transition bounds, LP graph, null-set transfer) is well constructed. The path-dependent change of variables is the right fix for the missing common fixed point. The countable-cover/Fubini step is valid. The martingale-type noise verification (Prop 2.11) and the RR verification (Prop 2.12) actually cover the advertised applications. The authors are also careful to state that Assumption 2.10 is the real restriction, and they verify it rather than leaving it abstract.\n\nSoft spots, in proportion: Assumption 2.10 is strong. It requires controlled weighted tails of the value errors and linearization errors, pathwise. That is not a flaw, but it means the theorem's reach beyond the two verified sampling models is untested. The proof of Theorem 2.13 uses a seed-independent choice of delta in (17); the constants in Lemma 3.3 are uniform in the reindexing start, but the paper doesn't explicitly say so. It follows, but a referee should ask for it to be spelled out. The NSGD application leans on [61, Theorem 3.6] for iterate convergence; the hypotheses aren't reproduced and the overlap in authorship is worth disclosing, but the reliance is external to the avoidance proof and is properly flagged. One more minor presentation issue: the essential smoothness of the mirror-map composite in Section 4.1 is asserted rather than fully checked; probably fine, but it deserves a line of verification.\n\nWho this is for: researchers in saddle avoidance, stochastic approximation, and random reshuffling theory. It deserves a serious referee. The central claim is novel, the proof is rigorous enough to warrant the time, and the limitations are honestly stated rather than buried.","headline":"A genuinely new pathwise framework that removes unit excitation and delivers the first a.s. saddle-avoidance result for random reshuffling; the proofs are detailed and the soft spots are minor.","tokens_in":42195,"tokens_out":2075,"would_cite":true,"duration_ms":22273,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C15","90C26","65K05","62L20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that stochastic recursions almost surely avoid strict saddles without unit excitation, replacing the noise-excitation mechanism with verifiable pathwise tail conditions and a Lyapunov–Perron graph argument.","keywords":["saddle avoidance","unit excitation","stochastic mirror descent","random reshuffling","Lyapunov-Perron","nonsmooth optimization","stochastic approximation","strict saddle"],"falsifier":"Choose a two-dimensional recursion $z_{k+1} = z_k - \\alpha_k(H z_k + b_k + J_k z_k)$ with $H$ having one negative eigenvalue and adversarial deterministic sequences $(b_k, J_k)$ for which the Assumption 2.10 tails diverge — say, $b_k = \\alpha_k^{-1}$ on sparse blocks or $J_k$ chosen so that $S_k$ does not converge — and test numerically whether a positive-measure set of initial points converges to $0$; a positive-measure basin would show that the tail conditions are necessary rather than merely sufficient.","tokens_in":41367,"feed_emoji":"🎯","tokens_out":8545,"duration_ms":86714,"temperature":0.7,"pith_summary":"This paper claims that a broad class of stochastic recursions $z_{k+1}=z_k-\\alpha_k G(z_k;\\xi_k)$ almost surely avoid the unstable zeros of their mean field $F$, without unit excitation: the noise is never required to have a uniformly positive projection in every direction. The authors prove that, along almost every random seed path, the set of initial points whose trajectories can stay trapped near an unstable zero has Lebesgue measure zero. This replaces the center-stable manifold argument used in deterministic saddle avoidance with a path-dependent change of variables and a Lyapunov–Perron contraction argument, and it works even though the sampled maps share neither a common fixed point nor a common linearization. If correct, the theorem yields almost sure strict saddle avoidance for stochastic mirror descent, for a normal map-based proximal stochastic gradient method on nonsmooth composite objectives, and for random reshuffling, and it converts iterate-convergence guarantees into convergence to local minimizers.","feed_headline":"Stochastic algorithms can avoid saddle points without unit excitation","feed_subtitle":"Pathwise tail conditions replace unit excitation, covering mirror descent, proximal SGD, and random reshuffling.","key_machinery":"The central object is the path-dependent change of variables $U_k = I - S_k$ with $S_k = \\sum_{i=k}^\\infty \\alpha_i J_i$, where $J_i = DG(z_*;\\xi_i) - DF(z_*)$; applying it to the centered recursion removes the sample-dependent linear perturbation from the leading term, leaving the common matrix $H = DF(z_*)$ in control of the spectral splitting. The Lyapunov–Perron operator $T_\\zeta$ acts on a weighted sequence space $Y_\\theta$ and expresses the center-stable component forward and the unstable component backward as a contraction; its fixed points form a Lipschitz graph $h: E^{cs} \\to E^u$ over the center-stable subspace, and that graph has Lebesgue measure zero. This graph is exactly the set of transformed initial points whose trajectories can remain in a bounded neighborhood of the saddle.","core_discovery":"The paper's central claim is Theorem 2.14: under local regularity of the sampled fields (Assumption 2.2) and the pathwise oracle condition (Assumption 2.10), the recursion (3) satisfies $P(\\lim_{k\\to\\infty} z_k \\in Z^*)=0$, where $Z^*$ is the set of unstable zeros of $F$ — points where $F(z_*)=0$ and $DF(z_*)$ has an eigenvalue with negative real part. Equivalently, if $z_0$ has a density, a trajectory cannot converge to a strict saddle of the mean-field dynamics. The authors show this is the natural stochastic replacement for the deterministic center-stable manifold theorem, and they verify the required conditions for i.i.d. martingale-type noise and for without-replacement finite-sum sampling.","pith_inferences":["Beyond the paper: any oracle that verifies the two pathwise series conditions in Assumption 2.10 — not only i.i.d. or without-replacement sampling — would make the corresponding algorithm inherit Theorem 2.14; the paper's examples are therefore a template rather than an exhaustive list.","Beyond the paper: the pathwise formulation suggests that saddle avoidance does not require a noise component in the unstable direction at all; vanishing noise (interpolation) and low-dimensional noise (finite-sum with $n