{"id":"bed93281-3d6c-4304-8853-f265ae590cee","arxiv_id":"2411.13939","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For a class of unimodal maps with state-dependent dynamical and observational noise, the paper proves filter stability, a central limit theorem, and large deviations; extreme value and Poisson results are proven only when the observational noise has constant amplitude.","lead":"This paper studies systems whose hidden state evolves nonlinearly with noise that depends on the state, and whose observations add more state-dependent noise. It proves that a Bayesian filter for these systems converges to the same distribution regardless of its starting guess, and it derives statistical laws for the observed process.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Filter stability rests entirely on Main Assumption, which is verified only under the global small-noise bound (Remark 3.8) not checked for the calibrated financial parameters; that gap needs a numerical check before the headline claim is accepted.","rationale":"The stress-test pass confirms the structure of the paper's main proof: the cone construction in Section 3.2.2 and Prop. 3.6 are internally coherent, and the contraction argument via Birkhoff's theorem is standard. The identification of the load-bearing assumption is the Main Assumption in Section 3.2.3, which is explicitly stated as an abstract hypothesis and verified only in Lemma 3.7 under conditions A and B. The paper does not verify these conditions for the empirical financial parameters of Figure 1. This is not an internal inconsistency, but it is a real scope gap: the abstract claims that the filtering result applies to the financial model, yet the only sufficient condition supplied (Remark 3.8) is untested there. I agree with the reader's assessment that this makes the paper correctly CONDITIONAL rather than fully ACCEPT. The other concerns noted by the reader (EVT/Poisson overclaim, in-preparation reference [29] for Prop. 8.2, conjectural extremal index for non-constant s) are real but secondary: they affect claims beyond the central filter-stability theorem and do not undermine the main proof. The concrete numerical test proposed would settle the primary scope question directly. A non-finding would be inappropriate because the gap is concrete and the paper itself supplies no check of the relevant bound for the advertised application.","tokens_in":25969,"tokens_out":1794,"duration_ms":15972,"concrete_test":"For the calibrated financial parameters used in Figure 1, compute or numerically estimate the quantities appearing in the sufficient condition of Remark 3.8: estimate min_x sigma(x) (the dynamical-noise amplitude on the support I_Gamma), delta (the support width of the dynamical noise density), and eps*max_x s(x) for a realistic observational-noise amplitude. If eps*max s(x) < min(|I|/10, min_x sigma(x)*delta/10) holds for the parameter choices, the gap is closed numerically and the theorem applies to the advertised model. If it fails, the paper should either restrict its main claims to the small-observation-noise regime or provide a separate verification of conditions A and B of Lemma 3.7 at the empirical parameters.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's headline claim is the filter-stability theorem: under the Main Assumption in Section 3.2.3, the filtering cocycle admits a unique absolutely continuous equivariant measure, so the optimal filter forgets its initial condition (Prop. 3.6). The proof uses Birkhoff's contraction in Hilbert metrics and is correct in structure, but the Main Assumption is an extra hypothesis that is not a consequence of Assumption TM. The only verification offered for the finance model is Lemma 3.7, whose conditions A and B are guaranteed by the global bound eps*max s(x) <= min(|I|/10, min_x sigma(x)*delta/10) (Remark 3.8). The paper never checks this bound for the parameters used in Figure 1 (gamma0=15.969, alpha=1.64, Sigma_eps=2.7e-5), nor does it state the relevant sigma(x), delta, or eps for that model. If the observational noise is not small compared to the dynamical noise, the cone-contraction event E={Z_n in I' and J_{Z_{n+1}} subset F_{Z_n}} may have zero frequency, and no filter stability is claimed. This is a scope gap rather than an internal contradiction: the theorem may be perfectly true under the global small-noise bound, but the paper's abstract and introduction phrase the result without this caveat, while the empirical application is presented as the motivation. The reader's weakest-assumption analysis identifies exactly this gap, and it is load-bearing because the only advertised application of the filter-stability theorem is the financial model with unverified parameters.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a hidden Markov model in which the latent process is a unimodal map perturbed by heteroscedastic Markov-chain noise and the observation is further corrupted by heteroscedastic observational noise. The central result is a filtering theorem: under the Main Assumption of Section 3.2.3, the filtering cocycle admits a unique absolutely continuous equivariant probability measure, so the optimal filter forgets its initial prior (Proposition 3.6). The proof uses Birkhoff contraction in Hilbert metrics on random cones. The paper also claims a CLT and large deviation principle for the observed process (Proposition 4.1), concentration inequalities under a constant modulation assumption (Proposition 5.1), Gumbel extreme-value laws (Propositions 7.1 and 7.3), and Poisson statistics for rare events (Propositions 8.1 and 8.2). The motivating application is a financial leverage model from the authors' previous work.","tokens_in":26520,"tokens_out":5407,"duration_ms":58879,"significance":"If the main theorem is correct, it is a valuable contribution: it gives filter stability for a genuinely nonlinear, non-uniformly hyperbolic hidden Markov model with heteroscedastic noises, and the random-cone method is an elegant way to obtain an equivariant measure. The paper is honest in stating the Main Assumption as a sufficient condition, and the proof of Proposition 3.6 is structurally sound. The advertised financial application is a potentially important selling point, but it is not verified for the calibrated parameters. The EVT and Poisson sections are more conditional and contain admitted gaps, especially when the modulation term is not constant. Overall the paper contains a substantial core result together with several less-developed peripheral claims that need to be either proved or explicitly delimited.","major_comments":[{"comment":"The advertised application to the financial model is not supported by the verified hypotheses. Lemma 3.7 establishes the Main Assumption only when conditions A and B hold, and Remark 3.8 gives the sufficient global bound ε max_x s(x) ≤ min{|I|/10, min_x σ(x)δ/10}. The paper never checks this bound for the parameters used in Fig. 1 (γ0=15.969, α=1.64, Σε=2.7×10^-5), and it does not state the corresponding σ(x), δ, or ε. It is therefore unknown whether the observational noise is small enough relative to the dynamical noise for filter stability to hold in the financial model. The abstract and introduction should either state this small-noise restriction explicitly or provide the missing numerical verification.","section":"§3.2.3, Lemma 3.7, Remark 3.8, Fig. 1"},{"comment":"Proposition 7.1 is conditional on β>0, but the positivity of the extremal index β is not established in Section 7.1. It is proved only later, in Section 7.2, under the additional Assumption EI that s is constant; the text explicitly says 'When s is not a constant, we do not have a proof.' Thus the Gumbel law for the general heteroscedastic observational-noise model is not demonstrated. In addition, the proof relies on inequality (33), which requires the stationary density h to be strictly positive on the set {x : x+s(x)ε ∈ B_t}; this strict positivity is asserted without proof in Section 7.1.","section":"§7.1 and §7.2"},{"comment":"Propositions 8.1 and 8.2 are stated under Assumption EI, i.e., s is constant, so they do not cover the general heteroscedastic case advertised in the title. Moreover, the proof that l(λ)=1 is only sketched with the phrase 'By the same argument used in the preceding section,' and Proposition 8.2 relies on the unpublished reference [29]. The compound-Poisson intermediate statement does not by itself prove standard Poisson statistics. A complete proof or a published reference is needed before these results are claimed.","section":"§8.1 and §8.2"}],"minor_comments":[{"comment":"The displayed inequality in Eq. (25) and the subsequent display contain mismatched parentheses and a duplicated event; the notation P×Pε should be used consistently and the parentheses should be balanced.","section":"§5.1, Eq. (25)"},{"comment":"The convention P(Z0=0)=1 appears to conflict with the later use of a stationary process (Xn,Zn) with a nontrivial Z0 distribution; the authors should clarify how these two viewpoints are reconciled.","section":"§3.1, Eq. (18)"},{"comment":"The symbol W_t is used both for the extreme-value probability in (29)-(30) and for the limiting random variable in Section 8.1; please use distinct notation.","section":"§7.1 and §8.1"},{"comment":"The statement 'Supposing that l(λ) is continuous at 0' should be justified, since identifying the characteristic-function limit requires control of l(λ) on a neighborhood of 0, not merely continuity at a single point.","section":"§8.1"},{"comment":"The reference [18] is likely misspelled as 'Brockner' rather than 'Brocker'; please check.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The filtering theorem is the main contribution and appears sound under its stated assumptions, but the paper oversells the financial applicability without checking the small-noise bound. The EVT and Poisson sections are more conditional than the abstract suggests, and the use of an 'in preparation' reference for a stated proposition is a concern for the editor. I would recommend major revision with a clear request to either prove or explicitly restrict the claims in Sections 7-8."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version. The paper's filter-stability theorem (Prop 3.6) is a genuine new result: for a class of unimodal maps with heteroscedastic dynamic noise plus heteroscedastic observation noise, they prove that the optimal filter forgets its prior and converges to a unique equivariant measure. The machinery is random cones and Birkhoff contraction, and the proof is careful; the Main Assumption is honestly stated, and Lemma 3.7 gives checkable sufficient conditions. The CLT and large-deviation result in Prop 4.1 are also new for this observation-noise setup, and the concentration inequality is a nice addition. I think the paper is a real contribution, and the self-citation to [1,2] is fine because the earlier stationary-measure theory is exactly the input.\n\nThe soft spots are real, but they are scope gaps, not internal contradictions. The abstract says \"no matter one's initial guess\" without the caveat. The Main Assumption is not a consequence of Assumption TM; it requires observation noise to be small compared with dynamical noise. For the bank-leverage model, Lemma 3.7's conditions A and B are guaranteed by the global bound in Remark 3.8, and the paper never checks that bound for the parameters used in Figure 1. So the flagship application is not verified. That should be fixed by a short numerical check. The EVT/Poisson sections are weaker: Prop 7.1 rests on Assumptions S/S' and an extremal index that is proven only when s is constant; for non-constant s it is explicitly conjectural. Prop 8.2 relies on a reference marked \"In preparation.\" These are honest limitations, but they should be flagged in the abstract as conditional.\n\nBottom line: this deserves serious refereeing. I'd send it to a good probability/dynamical systems journal with a request to verify the finance parameters numerically and to mark the EVT/Poisson results as conditional. If I were working on nonlinear filtering or rare events for noisy random maps, I'd cite the filter-stability theorem.","headline":"The filter-stability result is real and well-proved under its Main Assumption, but the flagship financial application is not verified because that assumption is never checked for the calibrated parameters.","tokens_in":26818,"tokens_out":2554,"would_cite":true,"duration_ms":24884,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G35","62M20","37A50","60F05","60F10"],"pacs":[],"model":"deepseek-v4-flash","headline":"For a class of chaotic unimodal maps with two state-dependent noises, the filtering distribution converges to a unique limit independent of the initial guess, carrying CLT, large-deviation, concentration, Gumbel, and Poisson laws.","keywords":["filtering stability","unimodal maps","heteroscedastic noise","observational noise","equivariant measure","random cones","Hilbert projective metric","bank leverage systemic risk"],"falsifier":"Simulate the bank-leverage chain with the parameters of Figure 1 using two different initial priors for the hidden state, and compare the filtering distributions after a long common observation sequence; if they remain at positive distance on a recurrent set of observations, the Main Assumption fails for those parameters. A more direct check is to evaluate $|T(J_{z_0})| < \\sigma(x)\\delta/2$ for a point $z_0$ in the support of $\\mu_Z$ and for all $x \\in J_{z_0}$; condition A of Lemma 3.7 failing means the paper's sufficient condition is not met.","tokens_in":25797,"feed_emoji":"🎯","tokens_out":11834,"duration_ms":114120,"temperature":0.7,"pith_summary":"This paper treats a hidden process that evolves by a chaotic unimodal map plus state-dependent dynamical noise, then passes through a second, independent state-dependent observational noise; both noises are heteroscedastic, meaning their size varies with the state. The main claim is filter stability: the conditional distribution of the hidden state given the whole observed history converges, as observations accumulate, to a unique absolutely continuous distribution that does not depend on the initial prior, so different agents starting from different guesses eventually make the same forecasts. The proof uses a random-cone contraction for the filtering cocycle, and the same spectral machinery yields a central limit theorem, a large-deviation principle, concentration inequalities, a Gumbel extreme-value law, and Poisson statistics for the observed process. The authors connect these results to a bank-leverage model of systemic risk, where the hidden state is the leverage of a representative bank.","feed_headline":"Filter converges to one law, whatever your starting guess","feed_subtitle":"For chaotic maps with state-dependent noises, the filter forgets its initial guess and limit theorems follow.","key_machinery":"The central object is the filtering cocycle $P((x,z),\\nu)=(\\sigma(x,z),P_z\\nu)$ on fibres over the stationary process $(X_n,Z_n)$, where $P_z$ is the Bayesian likelihood update followed by the Markov-kernel prediction. The paper equips each fibre with the cone $V_0(J_z)$ of integrable nonnegative densities supported on the observational-noise support $J_z$, carrying the Hilbert projective metric $\\Theta_0$. Birkhoff's theorem (Theorem 3.3) turns a linear map sending $V_0(J_z)$ into a smaller cone $V_c(J_{z'})$ into a strict contraction; the Main Assumption guarantees this happens with positive frequency, so the intersection of iterated images is a unique direction, the equivariant measure. The limit theorems are obtained by perturbing the quasi-compact Perron-Frobenius operator $L$: the perturbed operators $L_z$ and the Keller-Liverani spectral theory give the leading-eigenvalue expansion that produces the CLT, large deviations, Gumbel law, and Poisson statistics.","core_discovery":"On the paper's own terms, the discovery is that filtering is well posed for a class of chaotic unimodal maps with two heteroscedastic noises. The filtering cocycle constructed from the Bayesian update and the Markov-chain prediction admits a unique absolutely continuous equivariant probability measure, and any initial density converges to it at an exponential rate in the Hilbert projective metric (Proposition 3.6). The key mechanism is Birkhoff contraction on random cones: when the observation $z$ lies in a set $I'$ of positive measure for which the transition density $\\zeta(x,\\cdot)$ is bounded below by $c>0$ on an interval $F_z$ containing the support $J_z$ of the next observational-noise likelihood, the update maps the cone of nonnegative densities on $J_z$ into a strictly smaller cone $V_{c^4}(J_{z'})$, contracting distances. Because such good observations recur with positive frequency, the composed contractions select one direction independent of the starting point. The same framework gives a central limit theorem and large-deviation principle (Prop 4.1), a concentration inequality for the empirical measure (Prop 5.1), a Gumbel law with extremal index 1 (Prop 7.3), and Poisson distribution and point-process statistics (Props 8.1 and 8.2) under the stated assumptions.","pith_inferences":["Inference: If condition A of Lemma 3.7 fails for the calibrated bank-leverage parameters, the paper's machinery gives no uniqueness; a numerical comparison of $|T(J_{z_0})|$ with $\\min_x \\sigma(x)\\delta/2$ would delimit the regime where filter stability can be expected and might reveal a transition in the noise ratio.","Inference: The GEV detector sketched in Appendix A, where the location parameter should equal $\\log(t \\int s(x)^{-1}\\,d\\mu(x))$, could be tested on simulated time series to estimate the modulation average from data; the paper does not carry out that numerical test.","Inference: The paper leaves the non-constant-modulation case open for the extremal index and Poisson parameter; a natural test is to compute the quantities $q_k$ numerically for a slowly varying modulation and see whether the compound Poisson law reduces to the standard Poisson law."],"forward_implications":["For any two initial priors on the hidden state, the filtering distributions stay close and converge to the same equivariant filter as the observation window grows, so the filtering problem is well posed for these systems.","A central limit theorem holds for additive functionals of the observed process, and large-deviation bounds with a convex rate function follow (Prop 4.1).","If the modulation term is constant, the empirical measure of the observed process converges to the push-forward measure $\\mu'$ with exponential concentration, and its deviations from the stationary measure $\\mu$ are exponentially small (Prop 5.1 and its application).","Under the threshold and ball-shrinking assumptions, rare events of the observed process obey a Gumbel law with extremal index 1, and visit counts converge to a standard Poisson distribution and a Poisson point process (Props 7.3, 8.1, and 8.2).","In the bank-leverage model of systemic risk, when the observational noise is sufficiently small relative to the dynamical noise, these statistical laws apply to the leverage process (Lemma 3.7)."],"supporting_citations":[{"why":"Defines the heteroscedastic-noise Markov chain model, the bank-leverage interpretation, and the empirical parameter values used later.","marker":"[1]"},{"why":"Establishes quasi-compactness of the averaged Perron-Frobenius operator, exponential decay of correlations, and the BV space that this paper perturbs with observational noise.","marker":"[2]"},{"why":"Supplies the iterative nonlinear filtering scheme and filter-stability viewpoint on which the cocycle construction is based.","marker":"[16]"},{"why":"Provides the asymptotic-stability result for optimal filters of random chaotic maps that frames the equivariant-measure problem.","marker":"[18]"},{"why":"Gives Proposition 3.2.5, the formula the paper uses for the posterior update of the filter.","marker":"[17]"},{"why":"Supplies the Nagaev-Guivarc'h perturbed-operator technique used to prove the central limit theorem and large-deviation principle.","marker":"[23]"},{"why":"Provides the Markov-chain concentration inequality at the core of the exponential bound for the empirical measure.","marker":"[22]"},{"why":"Supplies the Keller-Liverani spectral-perturbation framework used to derive the Gumbel law.","marker":"[6]"},{"why":"Completes the Keller-Liverani rare-events machinery for escape rates and quasistationarity behind the extreme-value argument.","marker":"[7]"},{"why":"Provides the spectral-perturbation route to compound Poisson statistics used for the Poisson law and point-process convergence.","marker":"[15]"}],"fun_headline_variants":["Chaotic maps' filter: one law, any start","No matter your guess, the filter finds one law","State-dependent noise? Filter still converges to a single law","Heteroscedastic chaos: filter forgets its starting guess","One law for chaotic maps, regardless of initial guess"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The filtering-stability conclusion rests on the assumption that, for a positive fraction of observations, the dynamical noise is strong enough relative to the observational noise that the hidden state's transition density is bounded below by a positive constant on an interval containing the support of the next observation's noise; if that ratio is reversed, the contraction argument gives no uniqueness.","fun_headline_variants_meta":{"raw":{"variants":["Chaotic maps' filter: one law, any start","No matter your guess, the filter finds one law","State-dependent noise? Filter still converges to a single law","Heteroscedastic chaos: filter forgets its starting guess","One law for chaotic maps, regardless of initial guess"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000682,"raw_usage":{"total_tokens":3073,"prompt_tokens":899,"completion_tokens":2174,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":2093}},"tokens_in":515,"tokens_out":2174,"duration_ms":14346,"temperature":1.0,"reasoning_tokens":2093,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:44:07.950563+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the bank-leverage chain with the parameters of Figure 1 using two different initial priors for the hidden state, and compare the filtering distributions after a long common observation sequence; if they remain at positive distance on a recurrent set of observations, the Main Assumption fails for those parameters. A more direct check is to evaluate $|T(J_{z_0})| < \\sigma(x)\\delta/2$ for a point $z_0$ in the support of $\\mu_Z$ and for all $x \\in J_{z_0}$; condition A of Lemma 3.7 failing means the paper's sufficient condition is not met.","supporting_citations":[{"cited_title":"Lillo, G","cited_arxiv_id":null,"evidence_quote":"Defines the heteroscedastic-noise Markov chain model, the bank-leverage interpretation, and the empirical parameter values used later."},{"cited_title":"Lillo, G","cited_arxiv_id":null,"evidence_quote":"Establishes quasi-compactness of the averaged Perron-Frobenius operator, exponential decay of correlations, and the BV space that this paper perturbs with observational noise."},{"cited_title":"Chigansky, R","cited_arxiv_id":null,"evidence_quote":"Supplies the iterative nonlinear filtering scheme and filter-stability viewpoint on which the cocycle construction is based."},{"cited_title":"Brockner, G","cited_arxiv_id":null,"evidence_quote":"Provides the asymptotic-stability result for optimal filters of random chaotic maps that frames the equivariant-measure problem."},{"cited_title":"Capp´ e, E","cited_arxiv_id":null,"evidence_quote":"Gives Proposition 3.2.5, the formula the paper uses for the posterior update of the filter."},{"cited_title":"Aimino, M","cited_arxiv_id":null,"evidence_quote":"Supplies the Nagaev-Guivarc'h perturbed-operator technique used to prove the central limit theorem and large-deviation principle."},{"cited_title":"Paulin, Concentration inequalities for Markov chains by Marton couplings and spectral methods, Electronic Journal of Probability, (2015)","cited_arxiv_id":null,"evidence_quote":"Provides the Markov-chain concentration inequality at the core of the exponential bound for the empirical measure."},{"cited_title":"Keller, Rare events, exponential hitting times and extremal indices via spectral perturbation , Dynamical Systems 27.1 (2012), pp","cited_arxiv_id":null,"evidence_quote":"Supplies the Keller-Liverani spectral-perturbation framework used to derive the Gumbel law."},{"cited_title":"Keller, C","cited_arxiv_id":null,"evidence_quote":"Completes the Keller-Liverani rare-events machinery for escape rates and quasistationarity behind the extreme-value argument."},{"cited_title":"Compound Poisson statistics for dynamical systems via spectral perturbation","cited_arxiv_id":"2308.10798","evidence_quote":"Provides the spectral-perturbation route to compound Poisson statistics used for the Poisson law and point-process convergence."}],"review_version":1}