{"id":"965f00c0-f176-4b25-8422-bf14a8bbb054","arxiv_id":"2505.03069","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper defines robust invertibility for nonlinear dynamics using contraction and bi-Lipschitzness, and constructs biLipREN, a recurrent equilibrium network whose forward and inverse are both contracting and bi-Lipschitz by construction.","lead":"Researchers introduce a class of recurrent neural networks called biLipREN whose forward map and inverse map are both guaranteed to be stable and well-conditioned. This lets the network reconstruct inputs from noisy outputs, which is useful in control and generative modeling.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's second bound in (11) is not implied by the stated assumptions: it requires h^{-1}(x,y) to be Lipschitz in x with constant 1/gamma1, which is not a consequence of h being (gamma1,gamma2)-bi-Lipschitz in x.","rationale":"The reader's verdict is CONDITIONAL and identifies Theorem 1's second bound as the weakest point. I agree: this is the single load-bearing concern. The paper's central claim is that biLipREN is robustly invertible by construction, supported by Theorem 2; the conceptual definition of robust invertibility and its advertised error bounds rest on Theorem 1. The first inequality of (11) is the one used in the experiments and appears sound. The second inequality, however, is not derived. Its stated coefficient kappa2*nu/(gamma1*sqrt(1-alpha2^2)) requires the inverse output map h^{-1} to be Lipschitz in x with constant 1/gamma1, which is not implied by h being (gamma1,gamma2)-bi-Lipschitz in x. This is not merely a presentational gap: for h(x,u)=2x+sin(x)+u, h is (1,3)-bi-Lipschitz in x while h^{-1} has x-Lipschitz constant 3, so the precise constant in the theorem cannot follow. A corrected version can either add an explicit assumption on h^{-1}, or state the second bound with the Lipschitz constant of h^{-1} in place of 1/gamma1. Since the biLipREN parameterization itself appears sound and the gap is fixable, the appropriate verdict remains conditional rather than reject; no further verification concern outweighs this one.","tokens_in":13282,"tokens_out":22958,"duration_ms":219807,"concrete_test":"Re-derive the second inequality of (11) by writing out the analogue of (12)-(14) for the inverse system (2), making explicit every use of h^{-1}. In particular, check whether the bound |h^{-1}(x,y)-h^{-1}(x',y)| <= (1/gamma1)|x-x'| follows from the stated (gamma1,gamma2)-bi-Lipschitz assumption on h; testing the scalar h(x,u)=2x+sin(x)+u gives Lipschitz constant 3, so it does not. If the re-derivation requires adding this assumption or replacing 1/gamma1 by the actual Lipschitz constant L of h^{-1} in x, the theorem statement must be amended.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The first inequality of (11) is proved in (12)-(14). The second inequality is dismissed with 'the same logic.' Repeating the argument for the inverse system (2) requires its output map h^{-1}(x,y) to be Lipschitz in the state x, with constant 1/gamma1, so that initial-state differences of the inverse dynamics translate into bounded differences in reconstructed inputs. The stated assumption is only that h(x,u) is (gamma1,gamma2)-bi-Lipschitz in x: gamma1|x-x'| <= |h(x,u)-h(x',u)| <= gamma2|x-x'|. That gives no control over |h^{-1}(x,y)-h^{-1}(x',y)|, which also depends on how h varies in u. Example: h(x,u)=2x+sin(x)+u is (1,3)-bi-Lipschitz in x, but h^{-1}(x,y)=y-2x-sin(x) has x-Lipschitz constant 3, not 1/gamma1=1. Thus the stated second bound cannot be obtained from the hypotheses; a different valid bound can be obtained by chaining the first bound with the forward Lipschitz constant, but that changes the coefficient. The gap sits in the formal robustness certificate, though the first inequality and the biLipREN construction are not directly affected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies robust invertibility of discrete-time nonlinear state-space models. It defines a system to be robustly invertible if both the forward model and its causal inverse are contracting and bi-Lipschitz, and it derives explicit reconstruction error bounds in terms of initial-state mismatch and input/output perturbations (Theorem 1). The authors then propose the biLipREN, a recurrent equilibrium network parameterization, and prove in Theorem 2 that a matrix inequality certifies well-posedness, contraction, and bi-Lipschitzness for both the forward and inverse REN. They also show how to compose biLipRENs with static and dynamic orthogonal layers to obtain more general bi-Lipschitz dynamics, including a nonlinear analogue of inner-outer factorization, and demonstrate the approach on a nonlinear time-delay system and a four-cart mass-spring-damper system.","tokens_in":13509,"tokens_out":7492,"duration_ms":71016,"significance":"The proposed biLipREN construction is a potentially valuable contribution: it gives a parameterized recurrent network whose inverse is guaranteed to be well-posed, contracting, and bi-Lipschitz, which is relevant for learning invertible dynamics for control and generative modeling. The idea of certifying both forward and inverse properties through a single LMI and a congruence transformation is elegant, and the composition with orthogonal layers to build more expressive bi-Lipschitz models is useful. The paper also provides numerical evidence supporting the bounds. However, the formal robustness certificate in Theorem 1 has a gap in one of the two advertised error bounds, and the proof of Theorem 2 relies on an unstated external theorem; both issues must be resolved before the results can be accepted as stated.","major_comments":[{"comment":"The second bound in (11) does not follow from the stated assumptions. Repeating the proof for the inverse system requires h^{-1}(x,y) to be Lipschitz in x with constant 1/γ1, but the hypothesis that h(x,u) is (γ1, γ2)-biLipschitz in x gives no control over h^{-1}(x,y) in x. For example, h(x,u)=2x+sin(x)+u is (1,3)-biLipschitz in x, yet h^{-1}(x,y)=y-2x-sin(x) has x-Lipschitz constant 3, not 1/γ1=1. The proof says 'The second inequality follows the same logic,' but the analogous chaining yields a different coefficient unless an additional assumption (e.g., that h^{-1} is Lipschitz in x with the stated constant) is imposed. This gap is load-bearing for the advertised robust-invertibility certificate, so please correct the statement or the proof.","section":"§III.A, Theorem 1, Eq. (11), second inequality"},{"comment":"The proof of Claim 1 cites [38, Thm. 1] for contraction and strong input-output monotonicity, but the relevant theorem and its hypotheses are not stated in the paper. Since Theorem 2 is the central construction, please include the statement of the invoked result or provide a self-contained derivation, and verify explicitly that the LMI (28) satisfies the hypotheses of [38, Thm. 1] for the forward REN and, after the congruence transformation with Ψ, for the inverse REN (25). A quick check of the supply-rate matrices and the transformation of the (3,3) block would make the proof verifiable without access to [38].","section":"§IV.B, Theorem 2"}],"minor_comments":[{"comment":"The first line of (11) is missing a closing parenthesis: it should read ∥G^{-1}_b(G_a(u+δu))−u∥_T.","section":"§III.A, Eq. (11)"},{"comment":"In equations (12) and (13), some norms are written without the truncated-norm subscript T; please make the notation uniform.","section":"§III.A, proof of Theorem 1"},{"comment":"The clause 'for all a∈R^n' appears in Definition 4 but the inequality does not involve a; clarify that the operator G is indexed by initial state a, or remove the quantifier.","section":"§II, Definition 4"},{"comment":"The term '− 2 µ+νI' is rendered ambiguously; write it as −(2/(μ+ν))I and explicitly display the sign to avoid confusion.","section":"§IV.B, Eq. (28)"},{"comment":"The sentence 'By left- and right-multiplying (28) with Ψ^T and Ψ' should be accompanied by the resulting transformed blocks, or at least a note that the definitions in (26) are used; as written, the reader must reproduce the entire congruence calculation to verify the claim.","section":"§IV.B, proof of Theorem 2"},{"comment":"There are typos: 'reconstructs the the input sequences' should be 'reconstructs the input sequences', and the paragraph at the end of Section V.A ends with a stray 'sequences.' fragment.","section":"§I.A and §V.A"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper proposes a new recurrent model class, biLipREN, with certified contraction and bi-Lipschitzness for both the forward map and its explicit inverse. The central algebraic contribution—parameterizing the inverse REN via the congruence transform Psi in (27) and showing it inherits the same LMI conditions—is genuinely new and, as far as I can tell, correct. That alone is worth a serious look from anyone working in control-oriented learning or invertible sequence models.\n\nWhat I like: the composition with orthogonal layers gives a nice nonlinear analogue of inner-outer factorization, and the Mazur–Ulam remark in Remark 4 is a smart way to justify staying with linear inner factors. The paper is also honest that biLipREN has larger fitting error than a contracting REN, and the experiments, while only illustrative, at least attempt to compare the theoretical reconstruction bound against adversarial perturbations.\n\nThe soft spot is real. Theorem 1's second inequality in (11) is not implied by the stated hypotheses. The proof skips it with 'the same logic,' but repeating the argument for the inverse requires h^{-1}(x,y) to be Lipschitz in x with constant 1/gamma1. That is not a consequence of h being (gamma1,gamma2)-bi-Lipschitz in x. The stress-test example is correct: h(x,u)=2x+sin(x)+u gives h^{-1}(x,y)=y-2x-sin(x), whose x-Lipschitz constant is 3, not 1. So as written, the advertised error bound for state mismatch in the second inequality can be too optimistic. This does not invalidate the biLipREN construction in Theorem 2—that stands on the LMI congruence argument—but it does mean the paper's formal robustness certificate needs a repair. The fix is probably to either add an explicit Lipschitz assumption on h^{-1} in x, or state the bound with the actual constant, which will be uglier.\n\nThe paper also leans heavily on [37] and [38] and is not fully self-contained; that is acceptable since those are published, but a referee should ask for a cleaner derivation of (11). No code or data is provided, which is a minor drawback for a methods paper; the experiments are suggestive, not statistical.\n\nVerdict: worth sending to peer review. The main construction is a useful contribution and the gap in Theorem 1 is local and fixable. I'd accept a revision that either completes the proof under a proper assumption or weakens the bound accordingly.","headline":"The biLipREN construction is a useful new model class, but Theorem 1's second robustness bound has a genuine gap and needs a fix or a weaker bound.","tokens_in":14090,"tokens_out":6031,"would_cite":true,"duration_ms":56938,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B17","93C10","93D25"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a recurrent network class, the biLipREN, can be made robustly invertible by construction, with a single matrix inequality certifying contraction and bi-Lipschitz bounds for both the forward model and its explicit…","keywords":["robust invertibility","contraction analysis","bi-Lipschitz maps","recurrent equilibrium networks","input reconstruction","inner-outer factorization","strong input-output monotonicity","nonlinear system inversion"],"falsifier":"Take a scalar output map $h(x,u)=x+u^3$ inside an otherwise contracting, bi-Lipschitz system: $h$ is $(1,1)$-bi-Lipschitz in $x$, but its inverse $h^{-1}(x,y)=(y-x)^{1/3}$ is not Lipschitz in $x$, so initial-state differences in the inverse system can produce unbounded reconstruction errors. A simulation or analytic check of the second inequality in (11) for such a system would settle whether the missing assumption is really needed.","tokens_in":12997,"feed_emoji":"🔁","tokens_out":9852,"duration_ms":87754,"temperature":0.7,"pith_summary":"The paper proposes that the right notion of robust invertibility for a nonlinear dynamical system is contraction plus bi-Lipschitzness of the input-output map and of its inverse: small input changes produce small output changes, and different inputs produce outputs that remain distinguishable. It then constructs a recurrent neural model, the biLipREN, in which a single matrix inequality guarantees the forward model is well-posed, contracting, and $(\\mu,\\nu)$-bi-Lipschitz, and that its explicitly defined inverse is also well-posed, contracting, and $(1/\\nu,1/\\mu)$-bi-Lipschitz. A sympathetic reader would care because such a model can reconstruct the input sequence from a noisy, initial-state-mismatched output with an explicit worst-case error bound, without relying on relative degree or normal-form constructions. The paper also shows compositions with orthogonal linear layers give a nonlinear analogue of inner-outer factorization, demonstrated on a time-delay system and a coupled mass-spring-damper.","feed_headline":"Recurrent networks made robustly invertible by construction","feed_subtitle":"One matrix inequality certifies both a model and its inverse are contracting, so noisy outputs reveal their inputs.","key_machinery":"The load-bearing object is the recurrent equilibrium network (REN), a feedback interconnection of a linear time-invariant system with a slope-restricted static nonlinearity whose output is defined by an implicit equilibrium equation. The biLipREN is a REN constrained by the matrix inequality (28), which combines a Lyapunov-like matrix $P$ and a positive diagonal scaling $\\Lambda$ into a single certificate. That certificate simultaneously enforces well-posedness of the implicit equation, contraction of the state dynamics, and strong input-output monotonicity; Lemma 1 converts strong monotonicity into a $(\\mu,\\nu)$-bi-Lipschitz bound. Because the inverse REN is the same model conjugated by the transformation $\\Psi$ in (27), the identical inequality, left- and right-multiplied by $\\Psi^\\top$ and $\\Psi$, certifies the inverse's contraction and reciprocal bi-Lipschitz constants. This single-inequality, duality-by-congruence mechanism is what makes robust invertibility a construction guarantee rather than a post-hoc verification.","core_discovery":"The central claim is Theorem 2: if there exist $P=P^\\top\\succ 0$ and $\\Lambda\\in\\mathbb{D}_+$ satisfying the matrix inequality (28), then the recurrent equilibrium network (22) is well-posed, contracting, and $(\\mu,\\nu)$-bi-Lipschitz, and its inverse (25) is well-posed, contracting, and $(1/\\nu,1/\\mu)$-bi-Lipschitz. The inverse is not learned separately; it is obtained by an explicit coordinate transformation $\\Psi$, and conjugating the inequality by $\\Psi$ turns the forward certificate into the inverse certificate. Consequently robust invertibility is a structural property of the parameterization: every model produced within the class inherits the guarantee, and the reconstruction error is governed by the bounds in (11), with constants built from the contraction rates, overshoots, and bi-Lipschitz constants.","pith_inferences":["Beyond the paper, the same certificate could plausibly be reused in other implicit architectures, such as deep equilibrium networks, because the proof only exploits the linear-fractional structure of the REN and the equilibrium equation.","The distortion bound $\\nu/\\mu$ is a tunable design knob: smaller distortion should improve reconstruction accuracy while shrinking the expressive class, so a testable extension is to sweep $\\nu/\\mu$ and chart the fitting-versus-invertibility trade-off.","A further extension would use the certified inverse as an encoder in sequence-to-sequence learning or as a learned feedforward controller for non-minimum-phase plants, where the explicit error bounds could serve as a safety certificate; this is not stated in the paper."],"forward_implications":["Every biLipREN obtained from the direct parameterization comes with a guaranteed inverse that is itself a REN, so input reconstruction is a forward simulation of that inverse rather than a numerical optimization.","The robust-invertibility certificates are available before training, so unconstrained gradient-based learning can be used while keeping the guarantees intact.","The error bounds in (11) give explicit worst-case reconstruction performance as a function of initial-state mismatch and output noise, with constants computable from the certificate and the output map.","Compositions of biLipRENs with static orthogonal layers remain bi-Lipschitz with multiplicative constants, and compositions with dynamic orthogonal (all-pass) layers realize a nonlinear inner-outer factorization in the style of minimum-phase/all-pass decompositions.","In the experiments, a biLipREN fits the nonlinear mechanical system with only mildly larger error than a contracting REN, but unlike the contracting REN it recovers the input sequence under measurement noise and uncertain initial states."],"supporting_citations":[{"why":"Supplies the recurrent equilibrium network model and the theorem that converts the matrix inequality into contraction and strong input-output monotonicity, which the proof applies directly to both forward and inverse models.","marker":"[38]"},{"why":"Introduces monotone bi-Lipschitz network layers and their composition with orthogonal maps, the static building block that the biLipREN extends to dynamical systems.","marker":"[37]"},{"why":"Establishes the well-posedness condition for the equilibrium network's implicit equation, used to ensure that both the REN and its inverse are well-defined.","marker":"[47]"},{"why":"Supplies the contraction theory used to define incremental exponential stability of the forward and inverse dynamics.","marker":"[17]"},{"why":"Defines inner-outer factorization of nonlinear operators, which the composition of biLipREN with dynamic orthogonal layers is said to generalize.","marker":"[21]"},{"why":"Provides the coupled mass-spring-damper system used as the robust-inversion experiment.","marker":"[48]"},{"why":"Defines output-input stability, the prior notion of inverse stability that the paper strengthens with incremental and bi-Lipschitz guarantees.","marker":"[28]"}],"fun_headline_variants":["BiLipREN: one inequality certifies forward and inverse contraction","One inequality guarantees contracting models with contracting inverses","Invertible neural dynamics via a single matrix inequality","Robustly invertible recurrent networks from a single certificate","Recurrent nets with certified invertible inverses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The second error bound in Theorem 1 assumes that the inverse of the output map responds in a bounded way to changes in the state; this is not implied by the stated assumption that the forward output map is bi-Lipschitz in the state, and the advertised bound depends on it.","fun_headline_variants_meta":{"raw":{"variants":["BiLipREN: one inequality certifies forward and inverse contraction","One inequality guarantees contracting models with contracting inverses","Invertible neural dynamics via a single matrix inequality","Robustly invertible recurrent networks from a single certificate","Recurrent nets with certified invertible inverses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000894,"raw_usage":{"total_tokens":3858,"prompt_tokens":955,"completion_tokens":2903,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":2825}},"tokens_in":571,"tokens_out":2903,"duration_ms":19335,"temperature":1.0,"reasoning_tokens":2825,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:00:11.934548+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a scalar output map $h(x,u)=x+u^3$ inside an otherwise contracting, bi-Lipschitz system: $h$ is $(1,1)$-bi-Lipschitz in $x$, but its inverse $h^{-1}(x,y)=(y-x)^{1/3}$ is not Lipschitz in $x$, so initial-state differences in the inverse system can produce unbounded reconstruction errors. A simulation or analytic check of the second inequality in (11) for such a system would settle whether the missing assumption is really needed.","supporting_citations":[{"cited_title":"Recurrent equilibrium networks: Flexible dynamic models with guaranteed stability and robustness,","cited_arxiv_id":null,"evidence_quote":"Supplies the recurrent equilibrium network model and the theorem that converts the matrix inequality into contraction and strong input-output monotonicity, which the proof applies directly to both forward and inverse models."},{"cited_title":"Monotone, bi- lipschitz, and polyak-lojasiewicz networks,","cited_arxiv_id":null,"evidence_quote":"Introduces monotone bi-Lipschitz network layers and their composition with orthogonal maps, the static building block that the biLipREN extends to dynamical systems."},{"cited_title":"Inner-outer factorization of nonlinear operators,","cited_arxiv_id":null,"evidence_quote":"Defines inner-outer factorization of nonlinear operators, which the composition of biLipREN with dynamic orthogonal layers is said to generalize."},{"cited_title":"A convex parameterization of robust recurrent neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the coupled mass-spring-damper system used as the robust-inversion experiment."},{"cited_title":"Output-input stability and minimum-phase nonlinear systems,","cited_arxiv_id":null,"evidence_quote":"Defines output-input stability, the prior notion of inverse stability that the paper strengthens with incremental and bi-Lipschitz guarantees."}],"review_version":1}