{"id":"191809a7-5ea6-42a1-b152-5bdc375b4d3b","arxiv_id":"2506.13531","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"In multivariate nonlinear time series models, impulse response functions are only partially identified, and identifiable summaries such as pseudo impulse responses depend on the chosen shock definition and universe of variables.","lead":"This paper shows that in nonlinear dynamic models, the shocks used to compute impulse responses are not uniquely determined by the data: many different shock definitions produce the same observed series. The result tells macroeconomists when impulse response estimates are credible and when they depend on assumptions about shock structure.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 5's identifiability claim is derived from a linear IRF formula and does not extend to the nonlinear setting, leaving the paper's proposed remedy for partial identification unproven.","rationale":"The reader's verdict of CONDITIONAL is appropriate. The central underidentification claim (Proposition 2) is well-supported, so no rejection is warranted. The reader identified the transition-density identification and smoothness as the weakest assumptions, which are real but partially acknowledged in the paper. However, a more concrete and load-bearing issue is Proposition 5, where the proof and statement use the linear IRF formula in a nonlinear context. This is a specific internal inconsistency that can be checked analytically. Since Proposition 5 is an proposed remedy, the paper's constructive claims are less secure, but the core negative result is not affected. Thus the verdict remains CONDITIONAL, with the focus shifted to Proposition 5.","tokens_in":23279,"tokens_out":1201,"duration_ms":13210,"concrete_test":"Take a simple nonlinear conditionally Gaussian model, y_t = m(y_{t-1}) + D(y_{t-1})ε_t with m nonlinear, and derive the IRF at horizon h=2. Then solve the constrained maximization in (4.1) for this IRF. If the maximum depends on y_{t-1} or on the shape of m, then Proposition 5's formula does not hold in the nonlinear setting, confirming the concern.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that nonlinear IRFs are partially identified, with the identified set characterized in Proposition 2, is credible and internally consistent. The reader's weakest assumption about smoothness is real, but the paper itself acknowledges in Appendix A.2 that non-smooth transformations add multiplicity; this weakens the exact description of the identified set but does not destroy the partial-identification message. The more pressing concern is Proposition 5, which is used to argue that identifiable functions of the IRF exist. Its proof and statement rely on the linear IRF formula IRF(h,δ,a) = a'A^h Dδ, which is valid only for the linear Gaussian VAR (2.8). In the nonlinear model (2.1), the IRF defined in (2.7) is nonlinear in δ, so the optimization over δ in (4.1) does not lead to the stated closed-form sqrt(a'Φ^h DD'Φ'^h a). Thus Proposition 5's identifiability claim is unproven for the nonlinear framework that the paper is about. This is an internal inconsistency, not a disagreement with external consensus. Because Proposition 5 is one of the proposed remedies for partial identification, the practical contribution is weakened, although the central negative result (underidentification) stands.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies identification of impulse response functions (IRFs) in nonlinear multivariate dynamic models. It introduces a Gaussian (or uniform) nonlinear autoregressive representation of a Markov process and shows that, in dimension n≥2, this representation is not unique: the identified set consists of transformations that leave the standard normal or uniform distribution invariant. Proposition 2 characterizes this set as smooth invertible transformations with Jacobian determinant ±1, and Proposition 3 provides examples based on radius-dependent rotations. The paper then discusses remedies: non-Gaussian innovations and independent component analysis (Proposition 6), factor models with sources having different dynamics (Propositions 8 and 9), pseudo-IRFs (Proposition 7), and the role of the universe of variables (Section 6). The overall message is that nonlinear innovations and IRFs are generally partially identified, not point identified, unless additional restrictions are imposed.","tokens_in":23495,"tokens_out":16444,"duration_ms":158133,"significance":"If the results hold, the paper makes a useful conceptual contribution by formalizing the extent of underidentification of nonlinear IRFs. The proof of Proposition 2 in Appendix A.3 is self-contained, and the rotation examples in Proposition 3 are correct. The discussion of the dependence of Markov properties, innovations, and IRFs on the universe of variables (Section 6) is a valuable reminder. The paper also usefully connects the identification problem to nonlinear ICA and to dynamic factor models. Its contribution is primarily formal and conceptual rather than an estimation theory; no consistency or inference results are claimed. The central negative result—that nonlinear IRFs are partially identified in general multivariate settings—is credible and internally consistent.","major_comments":[{"comment":"The statement 'The function PIRFt is identifiable' is not proved and, as stated, is too strong. The random variable PIRFt(h,Δ,yt) in (4.10) is constructed by feeding the same future innovations ε_{t+1},... into the baseline and perturbed paths, so its full distribution depends on the particular function g selected from the identified set, not only on the transition density. What the text actually establishes in (4.11) is that E[PIRFt(h,Δ,yt)|yt] = E(yt+h|yt+Δ) − E(yt+h|yt), which is a functional of the transition density and hence identifiable. Because Proposition 7 is presented as one of the paper's remedies for partial identification, the proposition should be restated and proved for this conditional expectation (or for another identifiable summary), and the non-identifiability of the stochastic PIRF should be acknowledged.","section":"Section 4.3.3, Proposition 7"},{"comment":"The dimension count in Proposition 4 and Lemma 4 appears incorrect. The conditions (A.4) are the Cauchy-Riemann equations for the map b1 + i b2, so the space of harmonic pairs of total degree at most m has real dimension 2(m+1), not 2m; for m=1 the paper's own parametrization leaves the four parameters b1,00, b2,00, c00, and d11 free. Moreover, the set {b1,h,k, b2,h,k : h+k ≥ m} in the proposition is infinite-dimensional, so the statement as written is not well defined. This does not undermine the paper's qualitative underidentification conclusion, but the proposition and Lemma 4 need to be corrected or replaced by a precise statement.","section":"Section 3.2.2 and Appendix A.4, Proposition 4"}],"minor_comments":[{"comment":"Equation (2.9) and the corresponding PIRF formula in Remark 4 give the cumulative response (Id + Φ + ... + Φ^h), whereas the definition in (2.7) and the formula used in Proposition 5 are the horizon-h responses Φ^h Dδ and Φ^h Δ. Please correct these displays or label them explicitly as cumulative IRFs.","section":"Section 2.3, Eq. (2.9) and Remark 4"},{"comment":"The sentence referring to simulations in 'Section 4.2.2' should point to Section 4.3.2, which is where the simulation consistency argument appears.","section":"Section 4.3.3"},{"comment":"The cross-reference to 'Proposition 6' in Remark 4 appears to be a typo; the relation IRF = PIRF(...) is the content of Proposition 7.","section":"Section 4.3.3, Remark 4"},{"comment":"The codomain of T is written as (0,1]^n, and the Jacobian condition is stated on mixed open/closed domains; please make the domain and codomain of T consistent (presumably [0,1]^n or (0,1)^n).","section":"Section 3.1, Proposition 2"},{"comment":"The proof shows that any two representations yield a transformation T with |det| = 1, but it does not explicitly state the converse: every such T defines an observationally equivalent representation via g̃(yt−1, T(ut)) = g(yt−1, ut). Adding a sentence would make the 'identified up to' characterization complete.","section":"Appendix A.3, proof of Proposition 2"},{"comment":"In display (4.5), the components of the inverse c.d.f. vector are all written as F_1^{-1}(u_i,t); they should be F_i^{-1}(u_i,t) for i=1,...,n to match the statement that Fi is the c.d.f. of wi,t.","section":"Section 4.2, Eq. (4.5)"},{"comment":"It would help to state explicitly that Proposition 5 is a linear-model illustration and that no nonlinear analogue of the 'maximum IRF' identification is claimed; the current transition from the nonlinear underidentification problem to the SVAR(1) formula could be misread as proposing a nonlinear remedy.","section":"Section 4.1"},{"comment":"The word 'pormanteau' should be 'portmanteau'.","section":"Section 6.1.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a theory-oriented contribution, and its central negative result is solid. The main required changes are local: Proposition 7 needs to be restated for the identifiable conditional expectation of the PIRF, and Proposition 4's dimension statement needs correction or removal. The manuscript also relies on a substantial number of the authors' own prior papers for supporting results; most are legitimate background, but the editor may wish to ensure that Propositions 4 and 6 do not depend on unpublished or hard-to-access working papers. The paper could also benefit from a short passage clarifying its novelty relative to the nonlinear ICA results of Hyvarinen and Pajunen (1999)."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper. The central result is real: in a multivariate Markov model, the nonlinear AR representation with Gaussian/uniform innovations is only identified up to measure-preserving diffeomorphisms, so nonlinear IRFs are partially identified rather than point identified. Proposition 2 is self-contained and correct under the stated smoothness assumptions. That is worth knowing for anyone doing nonlinear local projections or generative-model IRFs. The paper also does well to show the one-dimensional uniqueness, the polar-coordinate rotation examples, and the dependence on time unit and universe.\n\nThe soft spots are real but not fatal. Proposition 5 is the biggest. It presents the maximum IRF as sqrt(a'Φ^h DD'Φ'^h a) and calls it identifiable, but that formula only holds for the linear Gaussian VAR—it uses IRF(h,δ,a)=a'A^hDδ, which is linear in δ and history-invariant. In the nonlinear model (2.7) the IRF is nonlinear in δ, stochastic, and history-dependent, so the optimization in (4.1) does not produce that closed form. As written, the proposition is stated for the general framework and is false; at minimum it needs to be explicitly restricted to the linear illustration, and the identified-function argument for nonlinear models needs a separate derivation. Proposition 7 is also overclaimed: the expectation of the pseudo-IRF is identifiable from the transition density, but the full stochastic PIRF as a coupled counterfactual may depend on the chosen representation. The paper should either prove distributional identifiability or state the weaker conditional-expectation result. Minor issues: Proposition 4 leans on an external harmonic-function result and its eigenvalue argument is sketchy, and Appendix A.2 shows non-smooth transformations add more multiplicity, so Proposition 2's identified set is a smoothness-dependent description (the paper acknowledges this). The heavy self-citation is mostly legitimate—the cited results are prior and external in origin.\n\nWho is this for? Applied macroeconomists using local projections or machine-learning reduced forms, and time-series econometricians thinking about nonlinear shocks. The negative result is the contribution; the remedies are less developed. I would send it to a serious referee, with the expectation of major revision. The central finding deserves to be in the literature; the overclaims need to be scrubbed.","headline":"A credible partial-identification result for nonlinear IRFs, with two overclaimed remedies (Props 5 and 7) that should be fixed before publication.","tokens_in":24003,"tokens_out":2647,"would_cite":true,"duration_ms":25642,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M10","60J05"],"pacs":[],"model":"deepseek-v4-flash","headline":"In multivariate nonlinear dynamic models, structural shocks and impulse responses are only partially identified: the same transition density is compatible with every smooth volume-preserving transformation of the underlying shocks.","keywords":["Nonlinear Autoregressive Model","Generative Model","Impulse Response Functions","Nonlinear Independent Component Analysis","Local Projections","Nonlinear Innovations","Partial Identification","Recurrent Markov Process"],"falsifier":"Simulate a bivariate nonlinear Markov process from a known transition density, estimate that density nonparametrically, and fit two representations $g_0$ and $g_1$ that are related by a nontrivial smooth volume-preserving transformation $T$ (for instance a norm-dependent rotation with $|\\det \\partial T/\\partial u'| = 1$). If both fits have the same likelihood while their horizon-2 impulse responses differ materially, Proposition 2's non-identification is confirmed in practice; if instead the transformed innovations fail a test of serial independence, the empirical identified set is smaller than the full orbit the paper describes.","tokens_in":23072,"feed_emoji":"🔀","tokens_out":17901,"duration_ms":154008,"temperature":0.7,"pith_summary":"The paper asks whether structural shocks and impulse response functions can be recovered uniquely in nonlinear multivariate dynamic models, and its answer is that in general they cannot. Any first-order Markov process with a positive transition density admits a nonlinear autoregressive representation $y_t = g(y_{t-1}; \\varepsilon_t)$ with standard normal shocks $\\varepsilon_t$, but once there are at least two series this representation is not unique: the shocks are identified only up to smooth invertible transformations $T$ of the uniform innovation with Jacobian determinant $+1$ or $-1$ everywhere. The nonlinear innovations and the impulse responses they generate are therefore partially identified, not point identified, which matters because impulse responses are the standard tool for saying what a shock does to the economy. The paper then catalogues conditions under which identification is restored — non-Gaussian independent sources, latent factors with different dynamics, and identifiable summary effects such as the pseudo impulse response — and it shows that the whole analysis depends on the universe of variables under study.","feed_headline":"Nonlinear shocks in multivariate models are not uniquely identified","feed_subtitle":"Same data fit many shock structures; non-Gaussianity or distinct source dynamics restores uniqueness.","key_machinery":"The load-bearing object is the identified set described in Proposition 2: for $n \\geq 2$, the collection of nonlinear autoregressive representations compatible with a given transition density is exactly the orbit of the true representation $(g, u_t)$ under smooth invertible transformations $T$ of the uniform shock with $|\\det \\partial T(u)/\\partial u'| = 1$ everywhere. Existence of at least one representation is supplied by the recursive conditional-quantile inversion $\\varepsilon_t = \\Phi^{-1} \\circ F(y_t \\mid y_{t-1})$, which builds Gaussian innovations from the conditional cumulative distribution function of the transition. The multiplicity is generated by distribution-preserving maps, and in polar coordinates these include norm-dependent rotations of the Gaussian shock, $Q(\\|\\varepsilon_t\\|^2)\\varepsilon_t$; the polar decomposition of the Jacobian (a scale, a rotation, a unit-determinant symmetric part) organizes the group of admissible transformations. This mechanism converts the identification problem into a question about measure-preserving diffeomorphisms of the unit cube, which is what lets the paper import and use results from nonlinear independent component analysis.","core_discovery":"The paper's central discovery is a characterization of the identified set for nonlinear dynamic models, stated as Proposition 2. Fix the transition density $f(y_t \\mid y_{t-1})$ of a first-order Markov process: for a single series ($n=1$) the nonlinear autoregressive representation with Gaussian innovations is essentially unique, up to the sign change $\\varepsilon \\mapsto -\\varepsilon$ and irregular non-smooth cases; for $n \\geq 2$ the uniform shocks $u_t$ are identified only up to a continuously differentiable, invertible transformation $T$ of $[0,1]^n$ whose Jacobian determinant is identically $+1$ or identically $-1$. Any such volume-preserving transformation of the shocks produces a new representation with exactly the same likelihood, so no amount of data can choose among them. A concrete family is rotation of the Gaussian shock vector by an angle that depends on its norm, $\\eta_t = Q(\\|\\varepsilon_t\\|^2)\\varepsilon_t$ with $Q$ a special orthogonal matrix function, the nonlinear analogue of the orthogonal-matrix indeterminacy of linear SVARs. Because the impulse response depends on which representative of this set is selected, the IRF is only partially identified; the identifiable objects are exactly the functionals of the transition density itself, such as the pseudo impulse response from a direct additive shock to $y_t$.","pith_inferences":["The norm-dependent rotations the paper constructs double as a robustness diagnostic: compute the impulse response over a grid of such rotations and report the spread, which measures how much of the reported IRF is structural rather than an artifact of the chosen representative.","Because the identified set is an orbit of measure-preserving transformations, conventional identifying devices — sign restrictions, narrative records, Cholesky-type recursions — amount to selecting one element of the orbit; their identifying content could be audited by checking whether the selected shock survives adversarial volume-preserving perturbations.","The universe dependence implies that nonlinear impulse responses are not coherent under aggregation: a shock identified in a small system need not correspond to any shock in an augmented system, with direct consequences for factor-augmented and network models.","A testable extension: apply the covariance-based Markov tests of Section 6.1.1 to residuals from two representations in the identified set; if one passes and the other fails serial independence, the empirical identified set is strictly smaller than the full orbit of Proposition 2."],"forward_implications":["With two or more series, two estimation exercises with different starting values or algorithms can converge to different functions $g$ inside the identified set and therefore to different impulse responses, all with the same asymptotic fit to the data.","Univariate nonlinear models are identified: for $n=1$ the Gaussian innovation and its IRF are unique up to the sign flip, so the failure of identification is a genuinely multivariate phenomenon.","Independence of non-Gaussian sources restores identification: if at most one source is Gaussian, the mixing structure and the sources are recovered up to scale and permutation (linear ICA).","Observed series that are transformations of independent latent factors with different dynamics — different autocovariance functions, or different transition-density derivatives — yield essentially unique factors and identifiable IRFs.","The pseudo impulse response, defined by shocking $y_t$ itself by $\\Delta$, is always identifiable from the transition density, and Proposition 7 shows the structural IRF equals the PIRF plus a horizon-one correction; the unidentified part of the IRF is precisely this correction."],"supporting_citations":[{"why":"Supplies the recursive multivariate quantile transformation used in Appendix A.1 to prove that every first-order Markov process admits the nonlinear autoregressive representation with Gaussian innovations.","marker":"Rosenblatt (1952)"},{"why":"Provides the inversion method for drawing from a distribution that the construction of uniform and Gaussian nonlinear innovations in Section 2.1 is built on.","marker":"Gourieroux and Monfort (1997)"},{"why":"Establishes existence of a solution to nonlinear ICA for continuous random vectors, the static analogue of Corollary 2 and the framing for the multivariate identification multiplicity.","marker":"Hyvarinen and Pajunen (1999)"},{"why":"Gives the linear ICA identifiability result that Proposition 6 rests on: independent non-Gaussian sources are identifiable up to scale and permutation.","marker":"Comon (1994)"},{"why":"Defines the nonlinear multivariate impulse response analysis that the paper's IRF extends and contrasts with, including the conditional-mean innovation the paper argues is unsuitable for shocking.","marker":"Koop, Pesaran and Potter (1996)"},{"why":"Supplies the Gaussian VAR benchmark (Remark 2 and Section 4.1) whose identifiable IRF functionals the paper generalizes to the nonlinear setting.","marker":"Plagborg-Moller and Wolf (2021)"},{"why":"Provides the polar decomposition of matrices with positive determinant used in Lemma 1 to parameterize the group of admissible Jacobian transformations.","marker":"Hall (2015)"},{"why":"Delivers the harmonic-function analysis behind Proposition 4 that fixes the dimension of the bivariate identified set and shows the underidentification is large.","marker":"Gourieroux and Jasiak (2022)"}],"fun_headline_variants":["Nonlinear IRFs identified only up to volume-preserving shock transformations","Multivariate nonlinear shocks: identification fails without extra restrictions","Nonlinear impulse responses: only partial identification possible","Shock identification in nonlinear models: many representations, same likelihood","Nonlinear IRFs: non-unique shocks unless sources are non-Gaussian"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise — flagged by the paper itself in Sections 4.3 and 6 — is that the one-step-ahead conditional distribution of the full observed system is identified from the data and that the series really form a first-order Markov system in the variables observed; if the distribution shifts over time, the Markov order is wrong, or relevant variables are omitted, the described set of equally good shock structures no longer applies, and the characterization counts only smooth transformations, whereas the non-smooth cases in Appendix A.2 add still more multiplicity.","fun_headline_variants_meta":{"raw":{"variants":["Nonlinear IRFs identified only up to volume-preserving shock transformations","Multivariate nonlinear shocks: identification fails without extra restrictions","Nonlinear impulse responses: only partial identification possible","Shock identification in nonlinear models: many representations, same likelihood","Nonlinear IRFs: non-unique shocks unless sources are non-Gaussian"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000612,"raw_usage":{"total_tokens":2831,"prompt_tokens":917,"completion_tokens":1914,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":1829}},"tokens_in":533,"tokens_out":1914,"duration_ms":12694,"temperature":1.0,"reasoning_tokens":1829,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:59:24.913878+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a bivariate nonlinear Markov process from a known transition density, estimate that density nonparametrically, and fit two representations $g_0$ and $g_1$ that are related by a nontrivial smooth volume-preserving transformation $T$ (for instance a norm-dependent rotation with $|\\det \\partial T/\\partial u'| = 1$). If both fits have the same likelihood while their horizon-2 impulse responses differ materially, Proposition 2's non-identification is confirmed in practice; if instead the transformed innovations fail a test of serial independence, the empirical identified set is smaller than the full orbit the paper describes.","supporting_citations":[],"review_version":2}