{"id":"1ec15c91-5ae3-47c3-b639-b6931301b6ce","arxiv_id":"1908.00845","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Backward iterations of dependent random maps converge under contraction in conditional expectation, yielding stationary ergodic solutions for a broad class of nonlinear autoregressions with exogenous covariates.","lead":"Nonlinear autoregressions that include outside covariates are shown to have unique stationary and ergodic solutions under a conditional contraction condition that is weaker than full independence but stronger than a Lyapunov exponent condition. The paper also provides dependence bounds that feed central limit theorems, illustrated on GARCH, Poisson, binary choice, and categorical models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central iteration theorem is sound; the concrete flaw is the GARCH verification in §6.8, where h_t is made to depend on ε_t and Y_t uses h_{t-1} instead of h_t.","rationale":"I focused on the central claim identified in the reader's strongest_claim, namely Theorems 2 and 4. The proof of Theorem 2 is essentially correct: the block-contraction argument in §6.1 gives summable Lp increments, which implies a.s. convergence because L1 norms are dominated by Lp norms; the stationarity/ergodicity transfer is a standard shift-of-variables argument; and the uniqueness argument uses the uniform moment bound in the right way. Theorem 4 correctly reduces the q-lag recursion to the first-order setup via the companion matrix B and the spectral radius condition ρ(A_1+...+A_q)<1. The reader's weakest_assumption, A2, is indeed strong, but the paper explicitly demonstrates in §2.2 that a more natural tail condition does not suffice in dependent cases; this is a limitation of scope rather than a defect in the theorem. The concrete problem I found is in the GARCH application: the state recursion in §6.8 is inconsistent with model (12), since h_t is predictable and should not contain ε_t, while Y_t^± should be multiplied by h_t, not h_{t-1}. This is a real, localized soundness issue in a headline example, and it supports the reader's CONDITIONAL verdict. Because the central iteration theorems are not invalidated, I do not recommend moving the verdict to REJECT or strengthening it to ACCEPT; UNCHANGED is the appropriate outcome.","tokens_in":29924,"tokens_out":42498,"duration_ms":418310,"concrete_test":"Analytical check: re-derive Proposition 4 from the correct state vector X_t=(h_t,(Y_t^+)^δ,(Y_t^-)^δ) for model (12), with h_t predictable and Y_t^±=(ε_t^±)^δ h_t. Compute the B3 matrices from the true recursions, ensuring F_1 contains no ε_t terms and F_2,F_3 depend on h_t rather than h_{t-1}, and verify whether ρ(Σ A_j)<1 is equivalent to γ<1 from G2. If the corrected matrices have the same spectral radius, Proposition 4 can be repaired; if not, the illustrated GARCH application is unsupported as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theorems 2 and 4 are internally coherent: given A1-A2 (respectively B1-B3), the proofs of Lp/a.s. convergence, stationarity/ergodicity, and uniqueness in §6.1 and §6.4 go through; even the a.s.-convergence step from summable Lp increments is valid because the L1 norm is bounded by the Lp norm. The load-bearing assumption A2 is strong, but the paper's own §2.2 shows a tail condition does not suffice, so this is a known limitation rather than a hidden flaw. The one concrete defect I find is in the GARCH application, §6.8. For model (12), h_t is F_{t-1}-predictable and satisfies h_t = π'Z_{t-1}+Σ_{j=1}^q β_j h_{t-j}+Σ_{j=1}^q α_j^+(Y_{t-j}^+)^δ+α_j^-(Y_{t-j}^-)^δ, with no current ε_t term. The proposed state recursion defines F_1 with (α_1^+(ε_t^+)^δ+α_1^-(ε_t^-)^δ)y_{1,1}, which inserts ε_t into h_t, and defines F_2,F_3 as (ε_t^±)^δ y_{1,1}, which would make Y_t^±=(ε_t^±)^δ h_{t-1} instead of (ε_t^±)^δ h_t. The B3 matrices in the proof are therefore not the matrices for the recursion (12); the assertion that γ<1 implies B3 for this model is not established as written. This affects the headline GARCH example but does not undermine the central iteration theorem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a general existence and uniqueness theory for stationary, ergodic, and moment-bounded solutions of nonlinear autoregressions with exogenous covariates, viewed as backward iterations of dependent random maps. The central result, Theorem 2, replaces the classical average contraction condition with a contraction in conditional expectation, yielding Lp and almost sure convergence of backward iterations, uniqueness among non-anticipative solutions with finite p-th moment, and stationarity/ergodicity. Theorem 4 extends this to q-lag recursions under a matrix spectral-radius condition. Section 3 provides bounds on Wu's functional dependence measure and a central limit theorem. The framework is applied to CHARN, GARCH, Poisson autoregressions, binary choice, and categorical time series models.","tokens_in":30231,"tokens_out":10413,"duration_ms":107499,"significance":"If the main theorems are correct, this is a useful and broadly applicable framework: it gives explicit, checkable sufficient conditions for stationarity, ergodicity, moments, and functional-dependence estimates for nonlinear time series with covariates, and it covers both predetermined and strictly exogenous regressors. The paper is honest about the strength of its uniform contraction assumption, and Section 2.2 gives a concrete example showing that a weaker tail condition does not in general suffice for L1 convergence. The central proof strategy is coherent and the statements of Theorems 2 and 4 are internally consistent; no fitted parameters or reverse-engineered predictions appear. However, the GARCH verification in Section 6.8 contains a substantive mismatch with the stated model, so the application section needs correction before the paper can be accepted as is.","major_comments":[{"comment":"The verification of Assumption B3 for the GARCH model does not correspond to the recursion (12). In (12), h_t is F_{t-1}-predictable and satisfies h_t = π'Z_{t-1} + Σ_{j=1}^q β_j h_{t-j} + Σ_{j=1}^q α_j^+(Y_{t-j}^+)^δ + α_j^-(Y_{t-j}^-)^δ, with no current ε_t term. But the proof defines F_2(y,ζ_t) = (ε_t^+)^δ y_{1,1}, F_3(y,ζ_t) = (ε_t^-)^δ y_{1,1}, and includes (α_1^+(ε_t^+)^δ + α_1^-(ε_t^-)^δ) y_{1,1} in F_1. This makes the state component corresponding to Y_t^± equal to (ε_t^±)^δ times the previous value y_{1,1}, rather than (ε_t^±)^δ h_t. The displayed matrices A_1 and A_j are therefore not the Lipschitz matrices for the recursion X_t = ((Y_t^+)^δ, (Y_t^-)^δ, h_t) arising from (12), and the assertion that G1-G2 with γ<1 implies B3 is not established as written. A corrected verification should define F_2 and F_3 through the current h_t coordinate (or, with a suitable state ordering, through F_1 containing no current ε_t term) and recompute the spectral radius; the rank-one structure of the resulting matrix sum suggests that γ<1 may still be the right condition, but this needs to be proved.","section":"§6.8, Eq. (12) and Proposition 4"}],"minor_comments":[{"comment":"The display 'Σ_{s,t∈Z, s≤t} ‖d(y, f_t^s(x))‖_p < ∞' cannot hold as written over the infinite bi-infinite index set; the intended statement is presumably that the telescoping series is summable uniformly in t, e.g. sup_{t∈Z} Σ_{i≥0} ‖d(f_t^{t-i}(x), f_t^{t-i-1}(x))‖_p < ∞.","section":"§6.1, Eq. (22)"},{"comment":"The sentence 'An immediate consequence of Assumption A4' should refer to Assumption A3, since A4 is introduced later.","section":"§3.1, after Assumption A3"},{"comment":"In the definition H_t = ((Y_t^+)^δ, (Y_t^+)^δ, h_t), the second coordinate is presumably (Y_t^-)^δ; otherwise the state vector has a duplicate component.","section":"§4.2, Proposition 4"},{"comment":"The text 'Using Assumptions A5-B5' refers to assumptions that are not named in the paper; it should reference Assumptions A4 and B4.","section":"§6.5, Proof of Proposition 2"},{"comment":"The sentence 'A2 is satisfied with d = d' appears to contain a typographical omission; the new metric should be denoted by a different symbol, such as d with a tilde, to distinguish it from the original distance d.","section":"§6.4, Proof of Theorem 4"}],"recommendation":"major_revision","confidential_remarks":"The GARCH verification error is localized and likely fixable, and it does not undermine the central Theorems 2 and 4. I would encourage the authors to correct the state-space verification and to re-examine the other applications for the same type of state-coordinate mismatch. The paper fits the journal's scope and makes a genuine contribution if the applications are repaired."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my take.\n\nThe real contribution is Theorem 2: replacing the Lyapunov-exponent condition of Elton/Borovkov with a uniform conditional-expectation contraction on blocks of length m. That gives Lp and a.s. convergence of backward iterates, moments, stationarity, ergodicity, and uniqueness, and Theorem 4 extends it to q-lag recursions via spectral radius of the sum of the Lipschitz matrices. The proofs are straightforward and I believe correct. Section 2.2's counterexample showing that a tail condition on the contraction coefficient is not sufficient is honest and useful – that is a stated limitation, not a hidden flaw.\n\nThe functional dependence bounds in Propositions 1 and 2 are a genuine addition, and the examples (CHARN, PARX, binary choice, categorical) go beyond what is in the literature.\n\nThe soft spot is the GARCH application. I agree with the stress-test note: Section 6.8 defines the state recursion inconsistently with model (12). In (12) h_t is predictable, but F1 contains an epsilon_t term, and F2/F3 use y_{1,1} where they should use h_t. The A_j matrices in the proof are not the matrices for (12); the claim that gamma < 1 gives B3 is not established as written. This matters because GARCH is a headline example, but it is localized – the core iteration theorem does not depend on it, and I expect the fix is a redefinition of the state vector.\n\nAlso, the assumption numbering in Section 3.1 is sloppy: A4 is used before it is defined, and the proof of Proposition 2 cites A5-B5. Minor, but an editor should have someone clean it up.\n\nVerdict: conditional accept, as the reader says. The central argument holds up; the flaw is in one application. This deserves a serious referee. I would engage with it.","headline":"A genuinely new and mostly sound framework for dependent random maps with exogenous covariates, but the GARCH application is mis-specified as written and the core theorems deserve a careful referee.","tokens_in":30822,"tokens_out":4028,"would_cite":true,"duration_ms":36893,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M10","60G05","60G10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Backward iteration of dependent random maps yields unique stationary, ergodic nonlinear autoregressions with exogenous covariates.","keywords":["time series","random maps","ergodicity","dependence","exogenous covariates","conditional contraction","backward iterations","functional dependence measure"],"falsifier":"For the linear recursion $X_t=\\kappa(Z_{t-1})X_{t-1}+\\varepsilon_t$, take a stationary ergodic covariate process with $E\\log\\kappa(Z_0)<0$, $P(\\kappa(Z_0)\\ge1)>0$, and dependence such that $E[\\prod_{i=1}^j\\kappa(Z_{t-i})]$ does not decay to zero, as constructed in Section 2.2. The paper predicts the series solution fails to converge in $L^1$; verifying whether the backward iterations still converge in $L^1$ for such a process would delineate exactly how far beyond A2 the method can go. For $q$-lag models, the analogous test is to find a covariates-dependent model where $\\rho(A_1(z)+\\cdots+A_q(z))<1$ for every $z$ but $\\sup_z\\rho(A_1(z)+\\cdots+A_q(z))\\ge1$, and ask whether a stationary non-anticipative solution with a finite moment still exists.","tokens_in":29667,"feed_emoji":"📈","tokens_out":10487,"duration_ms":97267,"temperature":0.7,"pith_summary":"This paper establishes a general existence, uniqueness, stationarity, and ergodicity theorem for nonlinear autoregressive processes with exogenous covariates. The driving mechanism is the convergence of backward iterations of dependent random maps: under a conditional contraction in expectation that is uniform over blocks of $m$ iterates, the backward iterates $f_t^{t-s}(x)$ converge in $L^p$ and almost surely to a limit independent of the starting point $x$, and that limit defines the stationary solution $X_t$. The same conclusion is extended to $q$-lag recursions $X_t=F(X_{t-1},\\ldots,X_{t-q},\\zeta_t)$ under the spectral-radius condition $\\rho(A_1+\\cdots+A_q)<1$. The paper also bounds the functional dependence measure of the solution, so standard limit theorems apply; this yields concrete conditions for CHARN, GARCH, Poisson autoregressions, dynamic binary choice, and categorical time series.","feed_headline":"One bound makes nonlinear time series with covariates stationary","feed_subtitle":"Backward iterations of dependent random maps give a unique ergodic solution with finite moments.","key_machinery":"The central object is the backward iteration of dependent random maps, $f_t^s=f_t\\circ\\cdots\\circ f_s$, whose limit $f_t^{-\\infty}(x)=\\lim_{s\\to\\infty}f_t^{t-s}(x)$ is used to build the stationary solution. The load-bearing identity is the block contraction in Assumption A2: for each $t$, almost surely, $E[d^p(f_t^{t+m-1}(x),f_t^{t+m-1}(y))\\mid\\mathcal{F}_{t-1}]\\le\\kappa^p d^p(x,y)$ for a fixed $\\kappa<1$, together with a one-step $L$-Lipschitz bound. Iterating this bound block by block gives geometric decay of $L^p$ distances between backward iterates, which yields convergence, independence of the starting state, and uniqueness. For $q$-lag recursions the same mechanism is run through the companion matrix $B$ built from $A_1,\\ldots,A_q$; the condition $\\rho(A_1+\\cdots+A_q)<1$ implies $\\rho(B)<1$, producing a block contraction on the product state space.","core_discovery":"On its own terms, the paper proves Theorem 2: whenever Assumptions A1 and A2 hold, for every $t$ and every starting state $x$ the backward iterates $f_t^{t-s}(x)$ converge almost surely and in $L^p$ to a limit $X_t(x)$ that does not depend on $x$; the process $((X_t,\\zeta_t))$ is stationary, is ergodic whenever $(\\zeta_t)$ is, and is the unique non-anticipative solution of $X_t=f_t(X_{t-1})$ satisfying $\\sup_t E[d^p(X_t,x_0)]<\\infty$. Theorem 4 transfers the statement to $q$-lag recursions through a companion-matrix argument, with the checkable condition $\\rho(A_1+\\cdots+A_q)<1$. The paper further shows that the functional dependence coefficients of $X$ decay geometrically in the lag, up to a convolution term inherited from the covariate process, and gives a central limit theorem for partial sums of functions of the solution. The intended advance is to replace the classical contraction-on-average or Lyapunov-exponent condition by a conditional expectation contraction that is straightforward to verify when exogenous regressors enter the dynamic.","pith_inferences":["The paper notes A1-A2 do not require stationarity of $(\\zeta_t)$; an editor's extension is that the same backward-iteration construction may define nonstationary or locally stationary solutions, although laws of large numbers would need separate assumptions.","For categorical models the convergence of backward iterations is literal path coalescence; this suggests a constructive perfect-simulation or regeneration algorithm for categorical time series with covariates, extending the simple $q=1$ illustration.","The linear counterexample indicates that the dependence structure of $\\kappa(Z_t)$ matters as much as its marginal moments; a sharper research target is a condition on the tail of products along stationary paths that is between $\\|\\kappa(Z_0)\\|_\\infty<1$ and $E\\log\\kappa(Z_0)<0$.","The functional-dependence bounds rely only on Lipschitz dependence of the map on the covariate coordinate (A3' or B4), so the same route should yield CLTs for quasi-likelihood estimation in PARX, GARCH-X, and related observation-driven models with covariates not listed in the paper."],"forward_implications":["Any recursion satisfying A1-A2 has exactly one stationary non-anticipative solution with a finite $p$-th moment, and initializing from that solution makes the process stationary and ergodic.","For $q$-lag autoregressions, existence and uniqueness reduce to checking $\\rho(A_1+\\cdots+A_q)<1$, which is explicit for CHARN, GARCH with covariates, PARX, and other models in Section 4.","The functional dependence coefficients of the solution decay at least geometrically with the lag, up to a convolution term coming from the covariate process, so the central limit theorem in Theorem 5 applies to partial sums of Lipschitz functions of the process.","The framework works with predetermined regressors — the noise at time $t$ independent of the past, but possibly influencing future covariates — rather than requiring strict exogeneity.","Under strict exogeneity, Theorem 3 replaces the uniform contraction by the log-moment condition $E\\log\\kappa(Z_0)<0$, at the cost of losing unconditional moments and simple higher-lag conditions."],"supporting_citations":[{"why":"Supplies the $L^p$ contraction framework for iterated independent random maps that the paper adapts to dependent maps.","marker":"Wu and Shao (2004)"},{"why":"Gives the Lyapunov-exponent almost-sure convergence criterion for dependent Lipschitz maps, restated as Theorem 1 and used for categorical and binary examples.","marker":"Elton (1990)"},{"why":"Introduces the functional dependence measure used in Section 3 to control weak dependence and to derive the central limit theorem.","marker":"Wu (2005)"},{"why":"Provides the affine stochastic recursion series solution used in Theorem 3's proof and in the counterexample showing a log-moment condition is not enough.","marker":"Brandt (1986)"},{"why":"Studies stationarity of volatility models with covariates via affine random maps; Proposition 4 extends it to moment existence.","marker":"Francq and Thieu (2019)"},{"why":"Dynamic binary choice stationarity and mixing results; Proposition 6 gives weaker or complementary conditions.","marker":"de Jong and Woutersen (2011)"},{"why":"Introduces PARX Poisson autoregressions with covariates under Markovian covariates; Proposition 5 removes the Markov restriction.","marker":"Agosto et al. (2016)"},{"why":"Treats Markov chains in random environments; Theorem 3 improves by avoiding uniform contraction in the environment.","marker":"Stenflo (2001)"}],"fun_headline_variants":["Backward iterations prove unique ergodic solution for covariate-driven nonlinear AR","Conditional-expectation contraction yields ergodic nonlinear AR with covariates","Backward iteration guarantees stationarity for nonlinear AR with exogenous covariates","Ergodic solutions for nonlinear AR with covariates via backward iteration","Dependent random maps: a single condition for stationarity in nonlinear AR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire existence and uniqueness proof depends on Assumption A2, namely that the same contraction factor $\\kappa<1$ bounds the conditional expectation of the distance after every block of $m$ iterates, uniformly over all states and all times. If that uniform bound fails, the argument collapses even when the covariate process is stationary and ergodic; the paper's linear example shows that a weaker log-moment condition $E\\log\\kappa(Z_0)<0$ is not enough for $L^1$ convergence.","fun_headline_variants_meta":{"raw":{"variants":["Backward iterations prove unique ergodic solution for covariate-driven nonlinear AR","Conditional-expectation contraction yields ergodic nonlinear AR with covariates","Backward iteration guarantees stationarity for nonlinear AR with exogenous covariates","Ergodic solutions for nonlinear AR with covariates via backward iteration","Dependent random maps: a single condition for stationarity in nonlinear AR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00102,"raw_usage":{"total_tokens":4275,"prompt_tokens":889,"completion_tokens":3386,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":3295}},"tokens_in":505,"tokens_out":3386,"duration_ms":23320,"temperature":1.0,"reasoning_tokens":3295,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:32:06.207820+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the linear recursion $X_t=\\kappa(Z_{t-1})X_{t-1}+\\varepsilon_t$, take a stationary ergodic covariate process with $E\\log\\kappa(Z_0)<0$, $P(\\kappa(Z_0)\\ge1)>0$, and dependence such that $E[\\prod_{i=1}^j\\kappa(Z_{t-i})]$ does not decay to zero, as constructed in Section 2.2. The paper predicts the series solution fails to converge in $L^1$; verifying whether the backward iterations still converge in $L^1$ for such a process would delineate exactly how far beyond A2 the method can go. For $q$-lag models, the analogous test is to find a covariates-dependent model where $\\rho(A_1(z)+\\cdots+A_q(z))<1$ for every $z$ but $\\sup_z\\rho(A_1(z)+\\cdots+A_q(z))\\ge1$, and ask whether a stationary non-anticipative solution with a finite moment still exists.","supporting_citations":[],"review_version":1}