{"id":"9a9670cc-aee4-4acc-8ef5-2874fcaf3bca","arxiv_id":"2412.21095","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Adaptive deep neural network controllers with Lyapunov-based weight updates are claimed to keep the tracking error of stochastic nonlinear systems uniformly bounded in probability, even with non-vanishing noise.","lead":"This paper designs a controller for nonlinear systems with random noise, using three deep neural networks that learn unknown dynamics online while a Lyapunov analysis gives a probability bound on the tracking error. It promises bounded tracking error in probability even when the noise does not vanish at the equilibrium, a common simplifying assumption in prior work.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 1 is false: a scalar Ornstein-Uhlenbeck process satisfies its hypotheses but violates the claimed bound (exact probability 1 vs. bound 0.05), so Theorem 1's probabilistic guarantee is unsupported.","rationale":"The reader's verdict identified Lemma 1 as the load-bearing step and noted its proof is invalid. Our analysis sharpens this: Lemma 1 is actually false, as the one-dimensional Ornstein-Uhlenbeck counterexample shows. This is decisive because Theorem 1 applies Lemma 1 verbatim to convert the deterministic inequality LVL <= -c VL + b into the probabilistic statement (31). Without a valid Lemma 1, no escape-risk bound is established, and the UUB-p guarantee collapses. The simulation section does not test the escape probability, so it cannot repair the proof. The paper's high-level architecture is coherent and the simulations are reasonable, but the central correctness risk is fatal to the main claim. The reader's REJECT verdict is therefore appropriate; no adjustment is needed. The smoothness gap on g2 is real but secondary, since it could be fixed by strengthening Assumption 2, whereas no simple fix repairs Lemma 1.","tokens_in":20067,"tokens_out":12657,"duration_ms":117239,"concrete_test":"Run the scalar OU test: simulate or analytically solve dz = -z dt + 0.1 dW with z(0) = 0 and estimate P(sup_{0 <= s <= T} z(s)^2 >= 0.1) for increasing T. Recurrence implies this probability tends to 1, whereas Lemma 1 predicts <= 0.05. If the simulation confirms the limit is 1, Lemma 1 is false and Theorem 1 lacks a valid proof.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 1's UUB-p claim rests entirely on Lemma 1, and the lemma is not merely unproved; it is false. Counterexample: let z satisfy dz = -z dt + 0.1 dW, V(z) = z^2, m = 1, lambda = 0.1, t = 0, z(0) = 0. Then LV = -2V + 0.01 <= -2V + 0.01 on {V < 1}, so the generator inequality holds with kappa1 = 2, kappa2 = 0.01. The 1D OU process is positive recurrent, so it exits (-1,1) with probability 1, implying P(sup_{s>=0} V(z(s)) >= 0.1) = 1. Lemma 1's bound gives (1/m)V(0) + (1/lambda)V(0)exp(-kappa1 t) + kappa2/(kappa1 lambda) = 0 + 0 + 0.05 = 0.05, a contradiction. The Appendix's eq. (53) is the source of the error: after tau_m, V(~z) = 0 while V(z) is not, so the asserted equality of escape probabilities is false; the arbitrary partition into supermartingale/submartingale/martingale intervals is unjustified. Consequently, the probability bound (31) in Theorem 1 is unsupported. A secondary gap is the smoothness required for the Taylor expansion (12), which exceeds the stated continuity of g2.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers control-affine stochastic nonlinear systems of the form dx = (f(x)+g1(x)u)dt + g2(x)Σ(t)dω with unknown, unstructured drift f and diffusion g2. It proposes a controller (22) driven by three Lyapunov-based DNN elements whose weights are updated online by the laws (18)-(20). The main result, Theorem 1, claims that under gain condition (28) and feasibility condition (30), the concatenated error state is uniformly ultimately bounded in probability with the explicit probability bound (31), where the escape risk ϑ is given in (29). The proof relies on Lemma 1, a general escape-probability estimate for Itô processes satisfying LV ≤ −κ1 V + κ2. A five-dimensional simulation study is presented to illustrate tracking performance.","tokens_in":20312,"tokens_out":9608,"duration_ms":101183,"significance":"The proposed adaptive structure is conceptually attractive: three online-updated DNNs target the drift and diffusion uncertainties separately, the design avoids the common assumption that the noise vanishes at the equilibrium, and the stability argument aims to produce an explicit, computable escape-risk bound. The derivation of the update laws is systematic, and no constants appear to be fitted to the theorem. However, the paper's central novelty, the quantified probabilistic ultimate bound, rests entirely on Lemma 1, and Lemma 1 is contradicted by a simple one-dimensional example satisfying all of its hypotheses. Consequently, the main claim (31) is not established. The contribution is therefore conditional on a substantial repair of the probabilistic argument, not on local polishing.","major_comments":[{"comment":"Lemma 1 is false as stated, and this is load-bearing for Theorem 1. Consider the scalar Ornstein-Uhlenbeck process dz = -z dt + 0.1 dW, with V(z) = z^2, m = 1, λ = 0.1, t = 0, and z(0) = 0. On Q_m = {|z| < 1}, LV = -2z^2 + 0.01 = -2V + 0.01, so the generator hypothesis holds with κ1 = 2 and κ2 = 0.01. The bound in (57) gives P(sup_{s≥0} V(z(s)) ≥ 0.1) ≤ 0.05. But the one-dimensional nondegenerate OU process is recurrent, so it exits (-1,1) with probability one and must cross the level |z| = sqrt(0.1) ≈ 0.316 on the way; hence the left-hand side is 1, a contradiction. The source of the error is visible in Eq. (53): after τ_m, the stopped process satisfies V(\\tilde z(s)) = 0 while V(z(s)) need not be zero, so the two escape probabilities are not equal. The partition of [τ_B, ∞) into supermartingale, submartingale, and martingale intervals is also unjustified. Since Theorem 1 invokes Lemma 1 to obtain (31), the central probabilistic guarantee of the paper is unsupported.","section":"Section V (Lemma 1) and Appendix, Eqs. (43)-(57)"},{"comment":"The Taylor expansion vec(g2(x)) = Ψ(e, xd)e + vec(g2(xd)) is asserted with Ψ a C^∞ function, while the standing assumption in Section III is only that g2 is continuous. For a merely continuous g2, the difference vec(g2(x)) - vec(g2(xd)) need not be representable with a C^∞ matrix-valued Ψ; at least C^1 regularity is needed for the integral/Taylor representation, and C^∞ regularity is required for the stated smoothness of Ψ. This is not a purely technical point, because F1 and F2 are defined through Ψ and are the functions approximated by the DNNs in (14)-(16). The assumption on g2 should be strengthened consistently, or the regularity of Ψ should be proved from weaker hypotheses.","section":"Section III, Eq. (12)"}],"minor_comments":[{"comment":"The symbols b and c are used in Definition 1 before they are introduced later in the same section, and c is also used as a generic constant in the UUB-p definition; this makes the definition difficult to parse and should be reordered or renamed.","section":"Section V, Definition 1"},{"comment":"The filtration notation in Section II is self-referential and unclear: the text says 'Ft is a complete filtration given by the family of σ-algebras up to time t, i.e., FS : FS ⊆ Ft ∀t ∈ [0,t]', which uses t on both sides and appears to contain a typographical error.","section":"Section II"},{"comment":"In the simulation section, the state map is written as x : R≥0 → R^3, although the simulation state is five-dimensional; this appears to be a typo.","section":"Section VI"},{"comment":"The simulations report RMS tracking errors for individual sample paths but do not estimate the escape probability or compare it with the theoretical bound in (31), which is the paper's distinctive probabilistic claim.","section":"Section VI"}],"recommendation":"reject","confidential_remarks":"The paper has a clear and testable central claim, and the Lyapunov-based DNN adaptive structure is original, but Lemma 1 is demonstrably false: a standard one-dimensional Ornstein-Uhlenbeck process satisfies all hypotheses and violates the conclusion. This is a mathematical error in the main theorem rather than a presentation issue. I recommend rejection. A resubmission that replaces the escape-risk lemma with a correct estimate, or weakens the theorem accordingly, could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: the paper extends the authors' Lb-DNN framework to stochastic systems. The high-level idea is coherent: three DNNs compensate drift and diffusion terms, weights are updated by Lyapunov-derived laws, and the result is a UUB-in-probability bound with an explicit escape risk. That is a genuine step beyond the cited prior work, including the only earlier stochastic Lb-DNN result [35]. The relaxation of the vanishing-noise assumption is real, and the 5D simulation is clearly described.\n\nThe problem is the central guarantee. Lemma 1 is load-bearing and its appendix proof is not valid. The claimed partition of the time axis into supermartingale, submartingale, and martingale intervals after the process reaches its ultimate bound is unjustified for a general continuous strong Markov process. More concretely, equation (53) equates the escape probabilities of the stopped and original processes after the stopping time, which is false: V(tilde z) is zero while V(z) is not, so the difference is not zero. The stress-test counterexample is decisive: a 1D Ornstein-Uhlenbeck process dz = -z dt + 0.1 dW with V = z^2 satisfies the generator inequality LV <= -2V + 0.01 on {V < 1}, exits (-1,1) with probability 1, while the lemma's bound gives 0.05. The lemma is not merely missing proof; it is false. Theorem 1's probability bound (31) is therefore unsupported.\n\nThere is a secondary gap: the Taylor expansion (12) requires g2 smooth enough to produce a C-infinity Psi, while Section III only assumes g2 continuous. That is minor compared to the lemma. The simulations also lack error bars and have no baseline against [35], so the empirical claims are suggestive rather than strong.\n\nWhat the paper does well: the three-DNN architecture is a credible contribution, and the projection-based update laws plus the explicit treatment of nonvanishing noise are well motivated. The writing is clear, and the citation practice is fair, including proper attention to [35]. This is not a circular or sloppy paper; it is a serious attempt with a specific technical flaw.\n\nWho this is for: researchers in adaptive stochastic nonlinear control and Lyapunov-based learning. The paper deserves a serious referee rather than a desk reject, because the concept is valuable and a corrected proof of the probability bound would make it a solid contribution. As it stands, the main theorem does not hold, so any acceptance would require heavy revision.\n\nBest.","headline":"Extends Lb-DNN control to stochastic systems with a coherent architecture, but the main probability bound rests on a false lemma and the theorem is unsupported as stated.","tokens_in":20921,"tokens_out":1958,"would_cite":false,"duration_ms":20292,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that an adaptive three-DNN controller keeps the tracking error of a nonlinear stochastic system inside a computable ball with an explicit probability, without assuming the noise vanishes at the equilibrium.","keywords":["stochastic systems","Lyapunov methods","deep neural networks","adaptive control","nonlinear control systems","uniformly ultimately bounded in probability","escape probability"],"falsifier":"Take $V$ itself as a mean-reverting square-root diffusion $dV = (1 - V)\\,dt + \\sqrt{V}\\,d\\omega$ on the set $\\{V < m\\}$ with $m$ large and $\\lambda < m$; then $LV = -V + 1$ satisfies the hypotheses of Lemma 1 with $\\kappa_1 = \\kappa_2 = 1$, yet because the process is positive recurrent with non-degenerate interior diffusion, $P(\\sup_{t \\le s < \\infty} V(s) \\ge \\lambda) = 1$ for every $\\lambda < m$, exceeding the lemma's bound $(1/m)V(0) + (1/\\lambda)V(0)e^{-t} + 1/\\lambda$ whenever $\\lambda$ is large enough that this bound is below 1. That directly falsifies Lemma 1 and with it the probability certificate in Theorem 1.","tokens_in":19719,"feed_emoji":"🎯","tokens_out":12010,"duration_ms":109316,"temperature":0.7,"pith_summary":"The paper develops a Lyapunov-based adaptive controller in which three deep neural networks learn, in real time, the unknown state-dependent drift and diffusion terms of a control-affine nonlinear stochastic differential equation. The main result, Theorem 1, states that if the gain and feasibility conditions (28) and (30) hold, then for every initial condition in the set $S$ the concatenated error state $z(t)$ is uniformly ultimately bounded in probability: with probability at least $1-\\vartheta$ the norm of $z$ stays below $\\sqrt{\\lambda/\\alpha_1}$ for all future time, where $\\vartheta$ is an explicit escape risk computed from the initial Lyapunov value, the ultimate bound, and the design gains. The significance is that the guarantee does not require the noise to vanish at the equilibrium and does not assume known structure or known bounds for the drift or diffusion uncertainties, assumptions that pervade earlier stochastic adaptive-control results. A sympathetic reader would care because this turns a black-box deep network into a certified component of a safety-relevant stochastic control loop, with a probability bound that can be evaluated before deployment.","feed_headline":"Stochastic tracking error kept bounded with a quantified probability","feed_subtitle":"Three deep networks learn drift and diffusion online; proof gives a computable escape-risk bound.","key_machinery":"The load-bearing object is the Lyapunov function $V_L(z) = \\tfrac{1}{2} e^\\top e + \\sum_{\\ell=1}^3 \\tfrac{1}{2} \\tilde{\\theta}_\\ell^\\top \\gamma_\\ell^{-1} \\tilde{\\theta}_\\ell$ on the compact set $D$, together with the infinitesimal generator $L V_L$. The argument's other central pieces are: (i) the Taylor expansion (12), $\\mathrm{vec}(g_2(x)) = \\Psi(e,x_d)\\, e + \\mathrm{vec}(g_2(x_d))$, which converts the diffusion term into state-dependent terms that the DNNs can absorb, with the constant term $\\mathrm{vec}(g_2(x_d))$ entering the ultimate bound; (ii) the three Lb-DNNs $\\Phi_1, \\Phi_2, \\Phi_3$ with the universal-approximation identities (14)--(16), updated by the projection-based laws (18)--(20) with $\\sigma$-modification; and (iii) Lemma 1, which bounds $P(\\sup V \\ge \\lambda)$ by $(1/m)V(z(0)) + (1/\\lambda)V(z(0))e^{-\\kappa_1 t} + \\kappa_2/(\\kappa_1 \\lambda)$ using a partition of the post-hitting interval into supermartingale/submartingale/martingale stretches and Doob's maximal inequality. Equation (21) expands $\\Phi_\\ell - \\hat\\Phi_\\ell$ to first order in the weight error, producing the higher-order terms $\\Delta_\\ell$ that enter the gains and the ultimate bound $b$.","core_discovery":"On the paper's own terms, the central claim is Theorem 1. For the system (10), with the controller (22) and the DNN weight update laws (18)--(20), if the gain condition (28) and the feasibility condition (30) hold, then for every initial condition $z(0) \\in S$ the closed-loop solution satisfies\n$$P\\!\\left(\\sup_{t \\le s < \\infty} \\|z(s)\\| < \\sqrt{\\$\\lambda$/\\alpha_1}\\right) \\ge 1 - \\vartheta,$$\nwhere $\\vartheta$ is the explicit escape risk in (29). The theorem is the first in this line to combine three Lyapunov-based deep networks with a probabilistic ultimate-bound guarantee for both drift and diffusion uncertainties. The proof works by taking the infinitesimal generator of the Lyapunov function (24), using the Taylor expansion (12) of $\\mathrm{vec}(g_2)$ about the desired trajectory to fold the diffusion term into DNN-compensable expressions, and applying Lemma 1 to convert the generator inequality $LV_L \\le -c V_L + b$ into a maximal-inequality bound on the probability that $V_L$ ever exceeds a level $\\lambda$. Because the Taylor expansion leaves the residual term $\\mathrm{vec}(g_2(x_d))$ in the bound rather than requiring $g_2(0)=0$, the noise need not vanish at the equilibrium.","pith_inferences":["Editorial inference: the escape-probability methodology of Lemma 1, if its hypotheses can be verified for a specific closed-loop process, would apply as a general template to any Itô system satisfying $LV \\le -\\kappa_1 V + \\kappa_2$, making it a candidate probabilistic substitute for classical deterministic UUB certificates.","Editorial inference: because $\\vartheta$ depends on the initial Lyapunov value, an immediate extension is to invert (29) to compute the largest admissible initial-condition set for a prescribed risk level, or to adjust $k_e$ online to shrink the ultimate bound $\\sqrt{\\lambda/\\alpha_1}$ at the cost of control effort.","Editorial inference: the smoothness gap between the standing assumption ($g_2$ continuous, Section III) and the requirement that $\\mathrm{vec}(g_2)$ be $C^\\infty$ for the Taylor expansion (12) suggests a concrete mathematical extension using a Peano-type remainder under Lipschitz assumptions on $g_2$."],"forward_implications":["The designer can compute, before deployment, the set $S$ of admissible initial conditions and the escape risk $\\vartheta$, giving the controller a quantitative safety certificate.","Because $g_2(0)=0$ is never required, the method covers plants where stochastic noise acts even when the tracking error is zero, a regime excluded by most prior stochastic adaptive results.","No structural or bound knowledge of $f$, $g_2$, or $\\Sigma$ is assumed beyond continuity and the existence of the smooth Taylor expansion (12); the three DNNs absorb both drift and diffusion nonlinearities online.","The simulation on the five-dimensional system shows the tracking-error RMS remains bounded as the noise mean ranges over $[-0.1, 0.1]$ and covariance over $[1,10]$, degrading gradually rather than abruptly, consistent with the exponential-convergence claim."],"supporting_citations":[{"why":"Supplies the comparison theorem (Thm. 4.1) used to solve the Lyapunov inequality $LV \\le -\\kappa_1 V + \\kappa_2$ and obtain the expectation bound (45).","marker":"[2]"},{"why":"The only prior Lb-DNN result for stochastic nonlinear systems; this paper extends it by removing the vanishing-noise assumption and quantifying the escape risk.","marker":"[35]"},{"why":"Universal approximation theorem for deep narrow networks used to justify the existence of the three ideal DNNs and the representation (14)--(16).","marker":"[40]"},{"why":"Provides the first-order Taylor expansion of the DNN output in the weight error (21) and the ideal-weight boundedness Assumption 3.","marker":"[41]"},{"why":"The smooth projection operator used in the update laws (18)--(20), whose property P2 in Thm. 1 is invoked to bound the adaptive terms in (39).","marker":"[43]"},{"why":"Gives Theorem C.4 characterizing supermartingale/submartingale behavior used to partition the post-hitting interval in Lemma 1.","marker":"[48]"},{"why":"Supplies Doob's maximal inequality and Markov's inequality used to obtain the escape-probability bound in Lemma 1.","marker":"[49]"}],"fun_headline_variants":["Three DNNs tame stochastic drift and diffusion with a provable escape bound","Adaptive DNN control bounds stochastic tracking error with a provable probability","Probabilistic guarantee for stochastic control using three Lyapunov-based DNNs","Quantified escape bound: three DNNs handle drift and diffusion online","Stochastic tracking error bounded in probability with Lyapunov-based DNNs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire probability certificate rests on Lemma 1, whose proof assumes that after the Lyapunov function first reaches its ultimate bound the remaining time can be partitioned into intervals where the process is alternately a supermartingale, a submartingale, and a martingale, and that the stopped and unstopped processes have the same tail probability after the stopping time, neither of which is guaranteed for a general continuous strong Markov process.","fun_headline_variants_meta":{"raw":{"variants":["Three DNNs tame stochastic drift and diffusion with a provable escape bound","Adaptive DNN control bounds stochastic tracking error with a provable probability","Probabilistic guarantee for stochastic control using three Lyapunov-based DNNs","Quantified escape bound: three DNNs handle drift and diffusion online","Stochastic tracking error bounded in probability with Lyapunov-based DNNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001303,"raw_usage":{"total_tokens":5349,"prompt_tokens":1014,"completion_tokens":4335,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":4235}},"tokens_in":630,"tokens_out":4335,"duration_ms":29474,"temperature":1.0,"reasoning_tokens":4235,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:03:47.875275+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $V$ itself as a mean-reverting square-root diffusion $dV = (1 - V)\\,dt + \\sqrt{V}\\,d\\omega$ on the set $\\{V < m\\}$ with $m$ large and $\\lambda < m$; then $LV = -V + 1$ satisfies the hypotheses of Lemma 1 with $\\kappa_1 = \\kappa_2 = 1$, yet because the process is positive recurrent with non-degenerate interior diffusion, $P(\\sup_{t \\le s < \\infty} V(s) \\ge \\lambda) = 1$ for every $\\lambda < m$, exceeding the lemma's bound $(1/m)V(0) + (1/\\lambda)V(0)e^{-t} + 1/\\lambda$ whenever $\\lambda$ is large enough that this bound is below 1. That directly falsifies Lemma 1 and with it the probability certificate in Theorem 1.","supporting_citations":[{"cited_title":"Stabilization of stochastic nonlinear systems driven by noise of unknown covariance,","cited_arxiv_id":null,"evidence_quote":"Supplies the comparison theorem (Thm. 4.1) used to solve the Lyapunov inequality $LV \\le -\\kappa_1 V + \\kappa_2$ and obtain the expectation bound (45)."},{"cited_title":"Deep neural networks-prescribed performance optimal control for stochastic nonlinear strict-feedback systems,","cited_arxiv_id":null,"evidence_quote":"The only prior Lb-DNN result for stochastic nonlinear systems; this paper extends it by removing the vanishing-noise assumption and quantifying the escape risk."},{"cited_title":"A sufficiently smooth projection operator,","cited_arxiv_id":null,"evidence_quote":"The smooth projection operator used in the update laws (18)--(20), whose property P2 in Thm. 1 is invoked to bound the adaptive terms in (39)."},{"cited_title":"Øksendal and B","cited_arxiv_id":null,"evidence_quote":"Gives Theorem C.4 characterizing supermartingale/submartingale behavior used to partition the post-hitting interval in Lemma 1."},{"cited_title":"Le Gall, Measure theory, probability, and stochastic processes","cited_arxiv_id":null,"evidence_quote":"Supplies Doob's maximal inequality and Markov's inequality used to obtain the escape-probability bound in Lemma 1."}],"review_version":1}