{"id":"ddace06b-7b1d-4126-8cda-cd6ba866c98a","arxiv_id":"2608.02443","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"Two-timescale discrete-time fixed-point iterations converge linearly whenever the timescale parameter δ stays below an explicit bound δ̄ = c_S(1−c_F)/(L_FIX·L_S·(c_S+L_R)) built from operator contraction and Lipschitz constants.","lead":"This paper derives explicit formulas for how slowly one part of a two-timescale dynamical system can update while the other part runs fast, so that the combined system still provably converges. It covers deterministic and stochastic settings and shows the formula working on feedback-optimization controllers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 3 is not paracontraction: it requires a global uniform contraction rate c_F<1, without which Eq. (9) is not computable. This overstates the framework's generality and makes the advertised 'readily checkable' bounds unavailable for typical nonlinear plants.","rationale":"I read the main proofs carefully. The comparison-matrix argument in Appendix A is sound, and the M-matrix condition correctly yields the stated δ̄ in Eq. (9). The verification of the feedback-optimization constants in Proposition 1 is consistent, and the stochastic extension in Theorem 3 follows from the same comparison argument with expectations. The reader's weakest assumption is on target: Assumption 3 is a global, quantitative uniform contraction toward the fast fixed-point set, which is strictly stronger than the 'paracontraction' cited in the introduction. The one-dimensional example F(z)=z−z^3 makes this precise: it is paracontractive yet fails Eq. (5) for every c_F<1. This is not merely a wording issue: it means the theorems do not apply to paracontractive fast operators, and the advertised 'readily checkable' bound cannot even be formed without a known global c_F. For nonlinear plants, such global constants are rarely available, so the central contribution is narrower than claimed. I also noticed a minor technical gap in Lemma 1: c_S as defined is not real for parameter values satisfying (1−α)/(α μ^2)>1, but the same inequality forces X_fix=X in non-degenerate cases, so this is not a load-bearing concern. My read therefore does not change the reader's CONDITIONAL verdict: the mathematics is correct under the stated assumptions, but the framing and practical transfer need revision.","tokens_in":20928,"tokens_out":24955,"duration_ms":223103,"concrete_test":"Let X={0}, Z=[−1,1], F(x,z)=z−z^3, Z_FIX(x)={0}. Verify that d(F(z),Z_FIX)=|z−z^3| < |z| for all z∈(−1,1)\\{0} (paracontraction), while sup_{z≠0} d(F(z),Z_FIX)/d(z,Z_FIX)=sup(1−z^2)=1, so Assumption 3 fails for every c_F<1. This one-dimensional check isolates the gap between the stated hypothesis and the advertised one.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — explicit timescale-separation bounds for operator interconnections — rests on Assumption 3 (Eq. 5), which demands a global, uniform contraction toward the fast fixed-point manifold: d_F(F(x,z),Z_FIX(x)) ≤ c_F d_F(z,Z_FIX(x)) with c_F<1 for every slow state x. The text and introduction label this 'paracontraction', but standard paracontraction (e.g., the reference cited in the paper) only requires strict non-uniform decrease. A concrete counterexample to the equivalence: F(x,z)=z−z^3 on X×[−1,1] with Z_FIX(x)={0} is paracontractive (|F(z)|<|z| for z≠0), but |F(z)|/|z|=1−z^2 tends to 1 as z→0, so no c_F<1 satisfies (5). Thus Theorem 1 does not cover paracontractive fast operators; it needs a quantitatively uniform bound. Moreover, even where such a c_F exists, evaluating δ̄ in Eq. (9) requires global knowledge of c_F, c_S, L_FIX, L_S, L_R; for nonlinear plants these constants are typically local or state-dependent. The numerical sections use linear plants with explicit constants, so they do not demonstrate the advertised 'readily checkable' property in the general nonlinear case. This does not disprove the theorems, but it materially narrows the paper's central contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies discrete-time interconnected systems of the form x_{t+1}=x_t+δS(x_t,z_t), z_{t+1}=F(x_t,z_t), viewed as a slow/fast operator interconnection. It shows that if the fast operator contracts uniformly (with rate c_F) toward a slow-state-dependent fixed-point set Z_FIX(x), and the reduced operator x+R(x) is either contractive (Theorem 1) or averaged and metrically subregular (Theorem 2), then for δ below an explicit threshold δ̄ the full interconnection converges linearly to the fixed-point set. The results are extended to stochastic operators with i.i.d. noise (Theorem 3, Corollary 1), yielding convergence in expectation and almost surely. The theory is applied to deterministic and stochastic feedback optimization, producing explicit tuning rules for the timescale parameter. Appendices contain complete proofs of all main results.","tokens_in":21207,"tokens_out":16740,"duration_ms":155355,"significance":"If the results hold, the paper provides a useful operator-theoretic counterpart to continuous-time singular perturbation theory for discrete-time fixed-point iterations. The explicit δ̄ formulas (9), (13), (23) in terms of operator constants are a genuine strength, as is the M-matrix/Schur comparison argument, which avoids Lyapunov-function searches. The proofs are self-contained, and the extension to averaged/subregular and stochastic reduced operators broadens the range of algorithms covered. The feedback-optimization application gives concrete, non-asymptotic tuning rules. However, the advertised coverage of 'paracontractive' fast operators is broader than the actual uniform-contraction assumption, and the 'readily checkable' claim should be tempered for general nonlinear plants.","major_comments":[{"comment":"Assumption 3 requires a global, uniform contraction rate c_F∈(0,1) toward the fixed-point manifold, d_F(F(x,z),Z_FIX(x))≤c_F d_F(z,Z_FIX(x)). The text labels this 'paracontractive' and states that paracontractivity is weaker than contractivity. Standard paracontraction (strict decrease of distance to each fixed point) does not imply (5). For example, on X×[-1,1] with F(x,z)=z−z^3 and Z_FIX(x)={0}, |F(z)|<|z| for z≠0, but |F(z)|/|z|=1−z^2→1 as z→0, so no uniform c_F exists. Thus Theorems 1–3 do not cover general paracontractive fast operators. The introduction's claim that the results hold 'under a paracontraction property of the fast operator' is inaccurate and should be revised to 'uniform contraction' or accompanied by verifiable sufficient conditions for the uniform bound.","section":"Section II, Assumption 3 (Eq. 5); Section I"},{"comment":"The paper's central claim that the bounds are 'readily checkable' is only demonstrated on linear plants with known contraction rates (z_{t+1}=z_t+κ(u_t−z_t) and z_{t+1}=z_t+r_t(u_t−z_t)). For a general nonlinear plant satisfying Assumptions 2–5, the constants c_F, L_FIX, L_S, L_R are global and often state-dependent; no procedure, example, or discussion is provided for obtaining them. This does not invalidate the theorems, but it materially narrows the practical scope of the advertised checkability. The manuscript should either provide a nonlinear example with computable constants or add a remark clarifying that the constants must be known a priori and may be difficult to estimate in practice.","section":"Section VI-C, Eqs. (42), (44); Conclusions"},{"comment":"The stochastic framework assumes the same fixed-point manifold Z_FIX(x) for every realization r of the noise, i.e., all stochastic fast operators share a common equilibrium manifold. This is a strong requirement: in many stochastic operator interconnections (random coordinate updates, packet drops, asynchronous activations), the fixed-point set depends on the realization. The application in Proposition 2 satisfies the assumption by construction, but the paper does not discuss this restriction or relate it to existing stochastic-approximation settings. Since the stochastic results are advertised as covering 'random updates, asynchronous activations, or packet losses,' this limitation should be explicitly acknowledged.","section":"Section V, Assumption 9 (Eq. 16)"}],"minor_comments":[{"comment":"The quantification 'for all x,x'∈X and r∈D' is unclear in (21b), where an expectation over r appears. Please state explicitly that (21a) holds for every r, while (21b) is an averaged condition over the distribution of r.","section":"Section V, Assumption 13 (21b)"},{"comment":"There are several LaTeX formatting artifacts, e.g., '¯x x' instead of a projection notation, and inconsistent use of d_F(z,¯z_x) versus d_F(z,Z_FIX(x)) in Appendix A. These should be cleaned up for readability.","section":"Theorems 2 and 3, Appendix B"},{"comment":"The proof writes E[·] without conditioning on the current state. Since x_t,z_t are random at later steps, the one-step bound should be stated as a conditional expectation given the history, and the final inequality (24) obtained by taking total expectation. The argument is correct but the presentation should be made precise.","section":"Appendix E (proof of Theorem 3)"}],"recommendation":"major_revision","confidential_remarks":"The mathematical core is sound under the stated assumptions: the comparison-matrix proof is correct, and the explicit formulas for δ̄ are a useful contribution. The main weakness is the gap between the framing claims (paracontraction, readily checkable bounds) and the actual uniform-contraction assumption, which should be fixed before publication. The heavy self-citation pattern (e.g., [10], [36]–[38], [43], [45]) is not inappropriate, but the novelty relative to the preliminary version [43] could be stated more crisply."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things up front. First, the central math is solid: Theorems 1–3 follow from the stated assumptions via a clean comparison argument, the M-matrix/Schur condition gives exactly the δ̄ formulas in (9), (13), and (23), and the application constants in Props. 1–2 check out. Second, the paper oversells the fast-operator assumption. The intro and the text call it paracontraction, but Assumption 3 really demands a global uniform contraction rate c_F < 1 toward the fixed-point manifold. That is strictly stronger than paracontraction. The stress-test example—F(z) = z − z³, Z_FIX = {0}—is paracontractive but admits no such c_F, so the framework does not cover the class the text advertises.\n\nThe genuinely new content: explicit, checkable bounds on the timescale parameter expressed in operator constants, the averaged-plus-metrically-subregular extension in Theorem 2 (Lemma 1's contraction constant is correctly derived), and the stochastic extension in Theorem 3. The appendices contain complete proofs, and the paper is honest about its relation to the earlier conference version [43]. The authors do not fit parameters; simulations use δ = 0.999δ̄ without tuning.\n\nThe soft spots are real but are mostly about framing and transfer, not validity. The uniform-contraction requirement is the main one, and it is not a minor terminological quibble: without a global c_F, Eq. (9) cannot be computed. The paper's other advertised selling point—'readily checkable' bounds—also depends on global knowledge of c_F, c_S, L_FIX, L_S, L_R, α, and μ. For nonlinear plants these constants are typically local or state-dependent, and the numerics only treat linear plants with explicit constants, so the checkability claim is not actually demonstrated in the general nonlinear setting. A secondary point: no code or seeds are provided for the simulations.\n\nNone of this undermines the stated theorems. But it does narrow the practical reach of the contribution, and the authors should fix the paracontraction mislabel and discuss how the required constants could be estimated in realistic problems.\n\nWho is this for? People working on operator-theoretic analysis of two-timescale discrete-time algorithms—feedback optimization, distributed optimization, singular perturbation in discrete time. The paper deserves a serious referee: the core results are correct, the proofs are complete, and the constant-aware δ bounds are useful enough to justify the revision work. I would recommend sending it to peer review rather than desk-rejecting, with the understanding that a revision must address the Assumption 3 framing and the global-constant practicality gap.","headline":"The theorems are correct and the explicit δ̄ bounds are a real contribution, but the paper mislabels its fast-operator assumption as paracontraction when it actually requires a global uniform contraction rate, which narrows the advertised scope.","tokens_in":21874,"tokens_out":2051,"would_cite":true,"duration_ms":21544,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C70","47H10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that an interconnected fast-slow discrete-time system converges linearly whenever the timescale separation parameter δ is smaller than an explicit threshold built from the contraction rates and Lipschitz constants of its f","keywords":["timescale separation","singular perturbation","operator theory","fixed-point iterations","discrete-time systems","stochastic operators","feedback optimization","linear convergence"],"falsifier":"Instantiate the linear comparison system at the core of the proof—x_{t+1}=(1−δc_S)x_t+δL_S e_t and e_{t+1}=δL_FIX L_R x_t+(c_F+δL_FIX L_S)e_t—and simulate it with δ slightly below δ̄: the theorem predicts the error decays as Cρ^t with ρ the spectral radius of M(δ). Any admissible nonlinear instance satisfying Assumptions 1–6 whose error fails to obey the bound (10) at δ<δ̄, or any numerical divergence in that linear system, would refute the claim.","tokens_in":20671,"feed_emoji":"⚡","tokens_out":5911,"duration_ms":56228,"temperature":0.7,"pith_summary":"The paper bridges singular perturbation ideas and operator theory for discrete-time systems. It considers a slow update x_{t+1}=x_t+δS(x_t,z_t) coupled to a fast update z_{t+1}=F(x_t,z_t), where for each frozen slow state the fast operator has a fixed-point manifold Z_FIX(x). The authors show that if the fast map contracts uniformly toward that manifold and the reduced slow map (with z pinned to the manifold) contracts, then the full interconnection converges linearly for every δ below an explicit threshold δ̄ built from the operator constants. The same mechanism extends to averaged and metrically subregular reduced maps, and to stochastic counterparts with almost sure convergence. The threshold is practical: it turns a qualitative separation assumption into a number one can compute and use to tune a controller.","feed_headline":"One formula sets the speed limit for fast-slow iterations","feed_subtitle":"Linear convergence holds whenever the tunable step stays below a closed-form threshold built from operator constants.","key_machinery":"The workhorse is a 2×2 comparison matrix M(δ) that tracks two errors jointly: the slow-state error measured as distance to the reduced fixed-point set, and the fast-state error measured as distance to the fast fixed-point manifold. One-step bounds on each error produce the linear recurrence e_{t+1} ≤ M(δ)e_t, and the formula for δ̄ is precisely the condition that M(δ) have spectral radius below one. For the broader case, a lemma converts averagedness plus metric subregularity into an explicit contraction factor, so the same matrix argument applies without requiring the reduced map to be a strict contraction.","core_discovery":"The central claim is that convergence of the full two-timescale iteration follows from convergence of its two separated limits, and the separation can be quantified. If δ < c_S(1−c_F)/(L_FIX L_S(c_S+L_R)), the joint error vector measuring distance of the slow state to its target and the fast state to the slow-parameterized fixed-point manifold decays as Cρ^t. The proof reduces one step of the recurrence to a two-by-two comparison matrix whose diagonal entries are the contraction margins and whose off-diagonal entries are the coupling gains; the threshold is exactly the condition that this matrix have spectral radius below one. The same comparison argument covers reduced operators that are on","pith_inferences":["The comparison-matrix structure suggests a modular design principle: one can assemble a convergent two-timescale algorithm by certifying the fast component and the reduced component separately, then computing the safety margin; the constants compose exactly as matrix entries.","Because δ̄ shrinks as L_FIX or L_S grows, the framework points to a quantitative trade-off between coupling sensitivity and admissible update speed, which could guide sensor, actuator, and controller design in feedback systems.","A natural testable extension is to relax the global uniform contraction of the fast map to local or state-dependent rates and ask whether a state-dependent δ(x) can recover linear convergence; the current threshold requires the global constant c_F.","For stochastic operators, the expectation-based contraction is weaker than pathwise contraction, so one could explore whether the result persists under Markov-dependent rather than i.i.d. randomness, since the proof only uses the expected comparison inequality."],"forward_implications":["For any interconnection whose fast map contracts toward a slowly moving manifold and whose reduced map contracts, linear convergence of the joint state is guaranteed as soon as δ is below the explicit threshold; no Lyapunov search is needed.","The formula doubles as a tuning rule: smaller coupling gains (L_FIX, L_S, L_R) or larger contraction margins (c_S, 1−c_F) widen the admissible range of δ.","Averaged and metrically subregular reduced operators—covering many operator-splitting and optimization iterations—inherit the same timescale separation result, with the contraction constant stated explicitly in terms of the averaging and subregularity parameters.","Stochastic versions with expectation-based contraction deliver both convergence in expectation and almost sure convergence, so asynchronous updates, random activations, and packet losses fit the same framework.","In feedback optimization, the scheme with plant dynamics as the fast map and gradient descent on the reduced cost as the slow map converges for δ below the bound, whereas δ=1 can diverge even when the reduced method converges.","The bound is directly checkable from standard operator constants, which is the contrast with purely existential singular-perturbation results."],"fun_headline_variants":["Two-timescale convergence pinned to a single operator formula","Closed-form threshold guarantees joint convergence","Fast-slow iterations: one constant sets the convergence rate","Operator theory sets explicit bound for two-timescale methods","A simple inequality decides when fast-slow iterations converge"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the fast map shrinks the distance to its fixed-point manifold by a fixed factor c_F<1 uniformly over all slow states and all points in the state space, with c_F known and global; if contraction is only local, state-dependent, or unquantified, the threshold δ̄ cannot be evaluated and the theorems do not apply.","fun_headline_variants_meta":{"raw":{"variants":["Two-timescale convergence pinned to a single operator formula","Closed-form threshold guarantees joint convergence","Fast-slow iterations: one constant sets the convergence rate","Operator theory sets explicit bound for two-timescale methods","A simple inequality decides when fast-slow iterations converge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000231,"raw_usage":{"total_tokens":1280,"prompt_tokens":660,"completion_tokens":620,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":404,"completion_tokens_details":{"reasoning_tokens":547}},"tokens_in":404,"tokens_out":620,"duration_ms":5470,"temperature":1.0,"reasoning_tokens":547,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T07:17:41.843236+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Instantiate the linear comparison system at the core of the proof—x_{t+1}=(1−δc_S)x_t+δL_S e_t and e_{t+1}=δL_FIX L_R x_t+(c_F+δL_FIX L_S)e_t—and simulate it with δ slightly below δ̄: the theorem predicts the error decays as Cρ^t with ρ the spectral radius of M(δ). Any admissible nonlinear instance satisfying Assumptions 1–6 whose error fails to obey the bound (10) at δ<δ̄, or any numerical divergence in that linear system, would refute the claim.","supporting_citations":[],"review_version":1}