EconBase
← Back to paper

Identification in (Endogenously) Nonlinear SVARs Is Easier Than You Think

Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.

95,844 characters · 16 sections · 80 citation commands

Rendered from LaTeX for readability, not typeset faithfully. Citation keys are highlighted; maths is left as source; figures, tables and equation environments are summarised rather than reproduced; unrecognised commands are greyed out so nothing is silently dropped. Email addresses are removed.

Identification in (Endogenously) Nonlinear SVARs Is Easier Than You Think

\global\long\def\uwrite#1#2{\underset{#2}{\underbrace{#1}} }

\global\long\def\blw#1{\ensuremath{#1}}

\global\long\def\abv#1{\ensuremath{\overline{#1}}}

\global\long\def\vect#1{\mathbf{#1}}

\global\long\def\smlseq#1{\{#1\} }

\global\long\def\seq#1{\left\{ #1\right\} }

\global\long\def\smlsetof#1#2{\{#1\mid#2\} }

\global\long\def\setof#1#2{\left\{ #1\mid#2\right\} }

\global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long

\global\long \global\long \global\long \global\long \global\long \global\long\def\Ellp#1{\ensuremath{\mathcal{L}^{#1}}}

\global\long \global\long \global\long \global\long \global\long

\global\long\def\abs#1{\ensuremath{\left|#1\right|}}

\global\long\def\smlabs#1{\ensuremath{\lvert#1\rvert}}

\global\long\def\bigabs#1{\ensuremath{\bigl|#1\bigr|}}

\global\long\def\Bigabs#1{\ensuremath{\Bigl|#1\Bigr|}}

\global\long\def\biggabs#1{\ensuremath{\biggl|#1\biggr|}}

\global\long\def\norm#1{\ensuremath{\left\Vert #1\right\Vert }}

\global\long\def\smlnorm#1{\ensuremath{\lVert#1\rVert}}

\global\long\def\bignorm#1{\ensuremath{\bigl\|#1\bigr\|}}

\global\long\def\Bignorm#1{\ensuremath{\Bigl\|#1\Bigr\|}}

\global\long\def\biggnorm#1{\ensuremath{\biggl\|#1\biggr\|}}

\global\long\def\floor#1{\left\lfloor #1\right\rfloor } \global\long\def\smlfloor#1{\lfloor#1\rfloor}

\global\long\def\ceil#1{\left\lceil #1\right\rceil } \global\long\def\smlceil#1{\lceil#1\rceil}

\global\long \global\long \global\long \global\long \global\long \global\long\def\clsr#1{\ensuremath{\overline{#1}}}

\global\long \global\long \global\long \global\long

\global\long\def\smlinprd#1#2{\ensuremath{\langle#1,#2\rangle}}

\global\long\def\inprd#1#2{\ensuremath{\left\langle #1,#2\right\rangle }}

\global\long \global\long

\global\long \global\long \global\long \global\long \global\long \global\long \global\long

\global\long \global\long \global\long\def\sigf#1{\mathcal{#1}}

\global\long\global\long \global\long\def\flt#1{\mathcal{#1}}

\global\long\global\long \global\long \global\long \global\long \global\long \global\long

\global\long \global\long \global\long \global\long \global\long \global\long \global\long

\global\long \global\long \global\long \global\long \def\independenT#1#2{\mathrel{\rlap{$#1#2$}\mkern2mu{#1#2}}}

\global\long \global\long \global\long \global\long \global\long \global\long\def\inprobu#1{\ensuremath{\overset{#1}{\ensuremath{\rightarrow}}}}

\global\long \global\long \global\long\def\inLp#1{\ensuremath{\overset{\Ellp{#1}}{\ensuremath{\rightarrow}}}}

\global\long \global\long \global\long \global\long\def\wkcu#1{\overset{#1}{\ensuremath{\rightsquigarrow}}}

\global\long

\global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long

\global\long \global\long \global\long

\global\long \global\long \global\long \global\long \global\long \global\long\def\cv#1{\left\langle #1\right\rangle }

\global\long\def\smlcv#1{\langle#1\rangle}

\global\long\def\qv#1{\left[#1\right]}

\global\long\def\smlqv#1{[#1]}

\global\long \global\long \global\long \global\long \global\long\global\long \global\long\global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long

comment\global\long\def\objlabel#1#2{\{\}} \global\long\def\objref#1{\{\}}

\global\long \global\long\def\mset#1{\mathcal{#1}}

\global\long\def\largedec#1{\mathbf{#1}}

\global\long \newcommandx\Ican[1][usedefault, addprefix=\global, 1=]{I_{#1}^{\ast}}

\global\long \global\long \global\long \global\long \global\long\def\b#1{\boldsymbol{#1}}

\global\long \global\long \global\long \global\long \global\long \global\long\def\smldblangle#1{\ensuremath{\savebox{\@brx}{\(\m@th{\langle}\)} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}#1\savebox{\@brx}{\(\m@th{\rangle}\)} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}}

\global\long \global\long

\global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long\def\set#1{\mathscr{#1}}

\global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long \global\long

\footnotetext[1]{Department\ of Economics and Corpus Christi College; [email removed]}

\footnotetext[2]{Department of Economics and University College; [email removed]}

\setcounter{footnote}{0}

abstractWe study identification in structural vector autoregressions (SVARs) in which the endogenous variables enter nonlinearly on the left-hand side of the model, a feature we term endogenous nonlinearity, to distinguish it from the more familiar case in which nonlinearity arises only through exogenous or predetermined variables. This class of models accommodates asymmetric impact multipliers, endogenous regime switching, and occasionally binding constraints. We show that, under weak regularity conditions, the model parameters and structural shocks are (nonparametrically) identified up to an orthogonal transformation, exactly as in a linear SVAR. Our results have the powerful implication that most existing identification schemes for linear SVARs extend directly to our nonlinear setting, with the number of restrictions required to achieve exact identification remaining unchanged. We specialise our results to piecewise affine SVARs, which provide a convenient framework for the modelling of endogenous regime switching, and their smooth transition counterparts. We illustrate our methodology with an application to the nonlinear Phillips curve, providing a test for the presence of nonlinearity that is robust to the choice of identifying assumptions, and finding significant evidence for state-dependent inflation dynamics.

(ref) supersedes a result that first appeared, with rather stronger assumptions, as Theorem A.1 in DM24.

\thispagestyle{plain}

\pagenumbering{roman}

\thispagestyle{plain}

\setcounter{tocdepth}{2}

{\singlespacing

}

\pagestyle{fancy}

\pagenumbering{arabic}

Introduction

For more than four decades, following the seminal contribution of Sims80, the linear structural vector autoregression (SVAR)

align[align omitted — 182 chars of source]

has played a central role in empirical macroeconomics. This is a dynamic linear simultaneous equations model (SEM), in which the $p$ endogenous variables $z_{t}$ are jointly determined as a function of their past values and the $p$ (mutually uncorrelated) structural shocks $\varepsilon_{t}$. The latter are regarded as the exogenous inputs to the system, so that causality is understood to run from these shocks to current and future values of $z_{t}$, and a key object of interest is the mapping between $\varepsilon_{t}$ and $z_{t+h}$ for $h\geq0$: the (structural) impulse response function.

In this context, a fundamental result that characterises the extent to which the data is informative about the model parameters, and thus also about those impulse responses, may be phrased heuristically as follows:

quote\begin{enumerate}[label=(ID),ref=(ID)] • Data on $\{z_{t}\}$ is sufficient to identify the linear SVAR parameters $(c,\{\Phi_{i}\}_{i=0}^{p})$, and the structural shocks $\varepsilon_{t}$, up to, and only up to, an orthogonal matrix $Q$. \end{enumerate}

In light of this, what might be termed the `SVAR identification problem' becomes one of finding sufficient additional restrictions on that matrix $Q$, so as to pin down, wholly or partially, the model parameters. The literature since has developed a variety of ways of using macroeconomic theory to generate such restrictions, based e.g.\ on the relative timing of shocks, the signs of their effects on impact, their medium- and long-run effects, and their correlation with external instruments (for a textbook discussion of which, see KL17book).

The linearity of (ref) is convenient, but inherently limiting as to the nature of the dynamics that can be modelled. In particular, it has the rather undesirable implication that the response of the economy to shocks must be the same irrespective of the phase of the business cycle: so that e.g.\ an aggregate demand shock has exactly the same effect on unemployment and inflation in the depths of a recession, when there is considerable slack in the labour market, as it would during periods of expansion. The substantial literature on nonlinear (S)VAR models has arisen partly to address these limitations (see e.g.\ Chan09; TTG10; HT13, for surveys). These allow the parameters of the SVAR at time $t$ to depend on an exogenous (or if not wholly exogenous, at least predetermined) regime-switching process $s_{t-1}$, as e.g.\ in\footnote{In many treatments of these models, the regime indicator in (ref) is denoted as $s_{t}$, rather than $s_{t-1}$. However, a feature of these models is that the regime is always determined prior to the realisation of $\varepsilon_{t}$, and may thus be regarded as measurable with respect to time-$(t-1)$ information; we have written $s_{t-1}$ to make this clearer.}

equation[equation omitted — 121 chars of source]

where often $s_{t}\in\{1,\ldots,L\}$ takes finitely many values, and each $\Phi_{i}(s)=\sum_{\ell=1}^{L}\pi_{\ell}(s)\Phi_{i}^{(\ell)}$ switches, or smoothly transitions, between the parameters of the $L$ `regimes'; here each $\pi_{\ell}(s)\in[0,1]$, with $\sum_{\ell=1}^{L}\pi_{\ell}(s)=1$. The evolution of $\{s_{t}\}$ may be modelled as an exogenous Markov chain (as in a Markov switching model), possibly with state-dependent transition probabilities, or as a function of certain predetermined variables (such as an element of $z_{t-i}$ for some $i\geq1$, as in a typical `threshold autoregressive' model); but in any case, $s_{t-1}$ must be determined prior to the realisation of $\varepsilon_{t}$. We therefore refer to these henceforth as exogenous regime-switching SVARs. (This characterisation applies to time-varying parameter VARs, in which $\{s_{t}\}$ is also some exogenous but possibly nonstationary process, such as a random walk.)

While models of the form (ref) enjoy greatly enriched dynamics relative to (ref), here the possibility of regime switching exacerbates the identification problem. Indeed the counterpart of (ref) for general Markov-switching models is that, conditional on the regime $s_{t-1}=s$, the parameters of (ref) are identified up to an orthogonal matrix $Q(s)$. Since this matrix may vary with $s\in\{1,\ldots,L\}$, the number of unidentified parameters, and thus the number of restrictions needed to deliver (exact) identification, scales proportionally with the number of states $L$. In practice, this may necessitate replicating a common set of $p(p-1)/2$ restrictions across all $L$ regimes (see e.g.\ RRWZ05, Sec.\ II; SZ06AER, Sec.\ III), yielding $Lp(p-1)/2$ restrictions in total. Similarly, in their two-regime STVAR model, AG12AEJ impose a Cholesky ordering on the elements of $z_{t}$ in each regime.

The exogeneity of the regime (i.e.\ of $s_{t-1}$) moreover restricts the kinds of nonlinearities that may be exhibited by the model's impulse responses. Notably, since each regime is itself a linear SVAR, the immediate effects of the shocks (i.e.\ the impact multipliers) must be linear in $\varepsilon_{t}$: which in particular rules out the possibility of sign-dependent asymmetries. It also renders (ref) unable to accommodate occasionally binding constraints, such as the zero lower bound (ZLB) constraint on short-term nominal interest rates, because the model requires the regime (whether `constrained' or `unconstrained') to be determined prior to realising the value of the potentially constrained variable -- whereas, as a matter of logic, it ought to be the value of that variable which determines whether the model is in fact in the constrained or unconstrained regime (see AMSV21).

Recently, SM21 and AMSV21 introduced the first SVAR models involving what we here refer to as endogenous regime switching, which are notably distinguished from the earlier literature on the nonlinear SVARs of the form (ref) in that they permit the autoregressive `regime' to be determined jointly with the values of the endogenous variables. For example, the `censored and kinked SVAR' (CKSVAR) of SM21 takes the form \[ \phi_{0}^{+}y_{t}^{+}+\phi_{0}^{-}y_{t}^{-}+\Phi_{0}^{x}x_{t}=c+\sum_{i=1}^{k}[\phi_{i}^{+}y_{t-i}^{+}+\phi_{i}^{-}y_{t-i}^{-}+\Phi_{i}^{x}x_{t-i}]+\varepsilon_{t}. \] where $y_{t}^{+}$ and $y_{t}^{-}$ denote the positive and negative parts of $y_{t}$ (a scalar), and $x_{t}$ is $(p-1)$-dimensional. In this model there are two contemporaneous regimes: one associated with $y_{t}>0$ (the `unconstrained' regime, in the ZLB setting), and the other with $y_{t}\leq0$ (the `constrained' regime), and in every period the model is solved simultaneously for the current values of $y_{t}$ and $x_{t}$, and for the applicable regime (as depends on the sign of $y_{t}$). Thus in situations where the $\varepsilon_{t}=0$ would entail a solution of $y_{t}=0$ (or approximately so), this allows the impact multipliers of $\varepsilon_{t}$ to be asymmetric, being dependent on which regime they push the model into.

Building on these developments, this paper proposes a new class of nonlinear SVAR models, which have the general form

align[align omitted — 94 chars of source]

where each $f_{i}:\mathbb{R}^{p}\ensuremath{\rightarrow}\mathbb{R}^{p}$ is a continuous, possibly nonlinear function, with $f_{0}$ being invertible; we refer to these models as `endogenously nonlinear', in view of the nonlinearities on the l.h.s. Because it is not tied to any particular functional form, (ref) also offers a great deal of flexibility in its dynamics, comparable to that offered by (ref). This framework readily encompasses the CKSVAR, which corresponds to a special case in which each $f_{i}$ is piecewise linear. More general models with endogenous switching between several regimes may be straightforwardly encompassed within the framework (ref), by specifying

equation[equation omitted — 152 chars of source]

to be an invertible, (continuous) piecewise affine function, where $\{\set Z^{(\ell)}\}_{\ell=1}^{L}$ is a convex partition of $\mathbb{R}^{p}$, and the current regime $\ell_{t}$ corresponds to the element of that partition for which $z_{t}\in\set Z^{(\ell_{t})}$.

The principal contribution of this paper is to characterise observational equivalence in the setting of the following, slightly more general formulation of (ref),

align[align omitted — 172 chars of source]

where $\b z_{t-1}=(z_{t-1}^{\top},\ldots,z_{t-k}^{\top})^{\top}$, and $\b{f}_{1}:\mathbb{R}^{kp}\ensuremath{\rightarrow}\mathbb{R}^{p}$ need not be separable in the lags of $z_{t}$ ((ref)). Remarkably, despite the far greater flexibility afforded by the parametrisation of (ref) relative to the linear SVAR (ref), the fundamental identification result (ref) carries over to (ref) essentially unchanged. Under weak conditions on the functions $(f_{0},\b{f}_{1})$ and the distribution of the shocks, we have ((ref)):

quote\begin{enumerate}[label=(ID$^\prime$),ref=(ID$^\prime$)] • Data on $\{z_{t}\}$ is sufficient to identify the nonlinear SVAR parameters $(f_{0},\b{f}_{1})$, and the structural shocks $\varepsilon_{t}$, up to, and only up to, an orthogonal matrix $Q$. \end{enumerate}

This is a nonparametric identification result, in the sense that we do not suppose that $(f_{0},\b{f}_{1})$ have any particular (known) parametric form. While its proof draws upon the microeconometrics literature on nonlinear SEMs (see in particular Matz09Ecta,Matz15Ecta; BH18Ecta; CGHP21JPE), it constitutes a genuinely novel result within that setting. (ref) has the powerful implication that most of the existing identification results for linear SVARs apply directly to the endogenously nonlinear SVAR, since in both cases exact identification is a matter of imposing $p(p-1)/2$ restrictions sufficient to pin down $Q$.

There follows a discussion of the $L$-regime endogenous regime-switching SVAR, which arises when $f_{0}$ is specified to have the piecewise affine form (ref), and of how to verify the conditions of (ref) in this case ((ref)). (Here we also suppose, mostly to provide a practically convenient parametrisation, that the SVAR has the time-separable form (ref), with each $\{f_{i}\}_{i=1}^{k}$ also specified to have the same functional form as (ref).) In this context, our results imply that it is sufficient, for the purposes of exact identification, to impose identifying restrictions in only one of those $L$ regimes, or even to distribute these in some way across those regimes. To obtain smooth transitions between adjacent regimes, we propose to convolve $f_{0}$ with a smooth kernel. This has the considerable advantage of preserving the invertibility of $f_{0}$, whereas this may fail if one attempts to smooth $f_{0}$ by the usual device of replacing each indicator function in (ref) by a smooth, cdf-like function (as is very commonly done to produce `smooth transition' (S)VARs).

Our methodology is illustrated by estimating an endogenously regime-switching SVAR (in the log vacancy--unemployment ratio and inflation), to investigate the possibility of a nonlinear Phillips curve relationship ((ref)) that was recently proposed by BenignoEggertsson2023 to explain the recent post-pandemic inflation surge. In particular, our identification results allow us to examine the evidence for nonlinearity in a manner that is robust to alternative identification assumptions, thus shedding light on the recent debate between BenignoEggertsson2023 and BeaudryHouPortier2025.

Finally, we provide an extension of our results to the augmented model

align*[align* omitted — 214 chars of source]

where $(\b z_{t-1}^{(1)},\b z_{t-1}^{(2)})$ is some partitioning of $\b z_{t-1}$, and $\{v_{t}\}$ is a strictly exogenous process, in the sense of being independent of $(\b z_{0},\{\varepsilon_{t}\})$ ((ref)). Here $\sigma(\cdot)$ is a diagonal matrix (with strictly positive entries), which allows the conditional variances of the structural shocks to depend on certain predetermined variables. In this setting, we show that (ref) continues to provide a valid characterisation of the identification of $(f_{0},\b{f}_{1})$, and that $Q$ may moreover be subject to further restrictions, if there is sufficient variability in the (diagonal) entries of $\sigma(\b z_{t-1}^{(2)},v_{t-1})$; these correspond to exactly the restrictions familiar from the linear SVAR literature on `identification by heteroskedasticity'. Our main result here ((ref)) not only accommodates both: (i) ARCH-type heteroskedasticity; and (ii) the possible dependence of the r.h.s.\ of the SVAR on an exogenous process $\{v_{t}\}$; but also (iii) permits $\b{f}_{1}$ to be discontinuous in some of its arguments (specifically, $\b z_{t-1}^{(2)}$ and $v_{t-1}$).

notation*$e_{m,i}$ denotes the $i$th column of an $m\times m$ identity matrix; when $m$ is clear from the context, we write this simply as $e_{i}$. $\smlnorm{\cdot}$ denotes the Euclidean norm on $\mathbb{R}^{m}$. Matrix norms are always those induced by the corresponding vector norm. For a function $g:\mathbb{R}^{m}\ensuremath{\rightarrow}\mathbb{R}^{n}$, $Dg(u_{0})=[(\partial g_{i}/\partial u_{j})(u_{0})]$ denotes the $(n\times m)$ Jacobian (matrix) of $g(u)$ at $u=u_{0}$. A `density' is always a density with respect to Lebesgue measure, unless otherwise stated.

Observational equivalence and identification

The linear SVAR: a brief review

Our point of departure is the linear SVAR, in which the observed series $\{z_{t}\}$ are regarded as being generated linearly from an underlying $p$-dimensional sequence of structural shocks $\{\varepsilon_{t}\}$, each of which have an economic interpretation (as e.g.\ an aggregate supply shock, a monetary policy shock, etc.) That is, for some $k\in\ensuremath{\mathbb{N}}$,

equation[equation omitted — 145 chars of source]

where $z_{t}$ and $\varepsilon_{t}$ are $\mathbb{R}^{p}$-valued, and to permit a more compact presentation we have defined $\b{\Phi}_{1}\coloneqq[\Phi_{1},\ldots,\Phi_{k}]$ and $\b z_{t-1}\coloneqq(z_{t-1}^{\top},\ldots,z_{t-k}^{\top})^{\top}$.

Observational equivalence in this setting being well understood (see e.g.\ Hamilton94, Ch. 11; Lut07, Ch. 9), our purpose here is to briefly review this in a manner that facilitates the comparison with our results for the endogenously nonlinear SVAR, which are developed in (ref) below. To simplify the problem, we suppose that $\{\varepsilon_{t}\}_{t\in\mathbb{Z}}$ is i.i.d., with a (Lebesgue) density $\varrho\in\mathscr{R}$, normalised to have $\ensuremath{\mathbb{E}}\varepsilon_{t}=0$ and $\ensuremath{\mathbb{E}}\varepsilon_{t}\varepsilon_{t}^{\top}=I_{p}$. By the Markov property the joint density of $\{z_{t}\}_{t=1}^{T}$, conditional on $\b z_{0}$, is simply the product of the conditional densities of $z_{t}\mid\b z_{t-1}$, for $t\in\{1,\ldots,T\}$. Under our assumptions, this density is time-invariant, and equals \[ \varphi_{z_{t}\mid\b z_{t-1}}(\xi\mid\b{\xi}_{-1})=\varrho(\Phi_{0}\xi-\b{\Phi}_{1}\b{\xi}_{-1})\cdot\smlabs{\det\Phi_{0}}, \] where $\xi\in\mathbb{R}^{p}$, $\b{\xi}_{-1}\in\mathbb{R}^{kp}$. Accordingly, we say that two alternative parameterisations of the linear SVAR, $(c,\Phi_{0},\b{\Phi}_{1},\varrho)$ and $(\tilde{c},\tilde{\Phi}_{0},\tilde{\b{\Phi}}_{1},\tilde{\varrho})$, are observationally equivalent if they imply identical conditional densities $\varphi_{z_{t}\mid\b z_{t-1}}$; in which case they also yield identical (conditional) likelihoods, for every possible realisation of $\{z_{t}\}$.

We then have the following well-known result, that data on $\{z_{t}\}$ identifies the SVAR coefficients $(c,\Phi_{0},\b{\Phi}_{1})$ up to, and only up to, an orthogonal transformation. Let $\mathbb{O}(p)$ denote the set of $p\times p$ orthogonal matrices.

thmLet $(\tilde{c},\tilde{\Phi}_{0},\tilde{\b{\Phi}}_{1})\in\mathbb{R}^{p}\times\mathbb{R}^{p\times p}\times\mathbb{R}^{p\times kp}$. Then there exists a $\tilde{\varrho}\in\mathscr{R}$ such that $(\tilde{c},\tilde{\Phi}_{0},\tilde{\b{\Phi}}_{1},\tilde{\varrho})$ is observationally equivalent to $(c,\Phi_{0},\b{\Phi}_{1},\varrho)$ in the model (ref), if and only if there exists a $Q\in\mathbb{O}(p)$ such that \begin{align*} \tilde{c} & =Qc & \tilde{\Phi}_{0} & =Q\Phi_{0} & \tilde{\b{\Phi}}_{1} & =Q\b{\Phi}_{1}. \end{align*}
rem\refstepcounter{subremark} (\roman{subremark}). Versions of this result, or equivalent characterisations thereof, have long been utilised in the analysis of linear SVARs, and linear simultaneous equations models (SEMs). This particular characterisation leads naturally to the `orthogonal reduced-form parametrisation' (ARRW18Ecta, Sec.\ 2.3) of the SVAR, in terms of the (unidentified) $Q\in\mathbb{O}(p)$ and the (identified) reduced form parameters ($\Phi_{0}^{-1}\b{\Phi}_{1}$ and $\Phi_{0}^{\top}\Phi_{0}$), which has proved fruitful for the analysis of sign-restricted SVARs (Faust98CR; Uhlig98CR,Uhlig05JME; ARRW18Ecta), and the formulation of rank conditions for global identification (RRWZ10REStud). \refstepcounter{subremark} (\roman{subremark}). The preceding follows as a corollary to (ref) below, albeit that result is proved under stronger regularity conditions on the allowable set of densities $\mathscr{R}$. Because of the linearity of (ref), the same result also holds under weaker conditions on the model than we have maintained here. For example, we could require $\{\varepsilon_{t}\}$ merely to be stationary white noise, since all that is really needed to identify the reduced form parameters is the orthogonality of $\varepsilon_{t}$ from $\b z_{t-1}$. On the other hand, the assumption that $\{\varepsilon_{t}\}$ is an i.i.d.\ process, often with a known (often Gaussian) distribution is common in empirical work, particularly in the context of Bayesian SVARs, and even in discussions of identification in these models (as in e.g.\ RRWZ10REStud). \refstepcounter{subremark} (\roman{subremark}). Here we have maintained only that the individual elements of $\varepsilon_{t}=(\varepsilon_{1t},\ldots,\varepsilon_{pt})^{\top}$ are contemporaneously orthogonal, rather than being independent. We thereby exclude the possibility, highlighted in a strand of the linear SVAR literature (e.g.\ LMS17JoE; GMR20REStud), of exploiting the additional restrictions available when the shocks are independent and non-Gaussian, to strengthen the above result to one in which the SVAR coefficients are identified up to an unknown (signed) permutation matrix. \refstepcounter{subremark} (\roman{subremark}). Let $k_{0}$ denote the true lag order of the SVAR, i.e.\ the greatest $i\in\ensuremath{\mathbb{N}}$ such that $\Phi_{i}\neq0$. We have implicitly maintained that this is less than or equal to $k$, which may therefore be interpreted as an upper bound on the true lag order of the model. In this sense, (ref) does not assume knowledge of the true lag order $k_{0}$ of the SVAR, but merely of some finite upper bound $k$ thereof.

The (endogenously) nonlinear SVAR

We now seek to extend (ref) to the setting of the following (endogenously) nonlinear SVAR

equation[equation omitted — 92 chars of source]

where $f_{0}:\mathbb{R}^{p}\ensuremath{\rightarrow}\mathbb{R}^{p}$ is invertible, and $\b{f}_{1}:\mathbb{R}^{kp}\ensuremath{\rightarrow}\mathbb{R}^{p}$. (As a convenient location normalisation, we set $f_{0}(0)=0$.) This model evidently nests (ref), by taking $f_{0}(z_{t})=\Phi_{0}z_{t}$ and $\b{f}_{1}(\b z_{t-1})=c+\sum_{i=1}^{k}\Phi_{i}z_{t-i}$. Another important special case arises when $\b{f}_{1}(\b z_{t-1})=\sum_{i=1}^{k}f_{i}(z_{t-i})$ is additively time-separable, as considered in DM24. But while this separability facilitates an extension of the Granger--Johansen representation theorem to these nonlinear SVARs, it is not necessary for the results that follow. The only separability that we require here is between $z_{t}$, $\b z_{t-1}$ and $\varepsilon_{t}$.

We develop the following running example throughout the rest of the paper.

example[nonlinear Phillips curve] The Phillips curve is a key component of any macroeconomic model. It provides a causal link between aggregate output and prices, and is thus essential in modelling the monetary policy transmission mechanism. Its name derives from the seminal contribution of Phillips1958, who proposed the following simple nonlinear relationship between (wage) inflation, $\pi_{t}^{w}$, and labour market tightness as measured by the unemployment rate, $u_{t}$, \begin{equation} \pi_{t}^{w}=a+b\left(\frac{1}{u_{t}}\right)^{c}. \end{equation} Several recent contributions have used the vacancy-to-unemployment ratio, $\theta_{t}\coloneqq v_{t}/u_{t}$, as an alternative measure of tightness, and price (instead of wage) inflation, $\pi_{t}$, see e.g.\ BallLeighMishra2022, BenignoEggertsson2023, and BeaudryHouPortier2025. These papers employ alternative functional forms for (ref), and introduce additional dynamics, inflation expectations, and other shocks.\footnote{BallLeighMishra2022 use a third order polynomial in $\log\theta_{t}$, BenignoEggertsson2023 a piecewise linear function in $\log\theta_{t}$ with a kink at $\theta_{t}=1$, and BeaudryHouPortier2025 consider both these specifications.} Here we consider a stylised version of the piecewise linear Phillips curve proposed in BenignoEggertsson2023, \begin{equation} \pi_{t}-\pi=\begin{cases} \kappa\log\theta_{t}+\eta_{\pi,t}, & if \theta_{t}\leq1 (`normal')\\ \kappa^{\mathrm{tight}}\log\theta_{t}+\eta_{\pi,t}, & if \theta_{t}>1 (`labour shortage') \end{cases} \end{equation} where $\pi$ denotes steady state or target inflation, and $\eta_{\pi,t}$ an exogenous shock. Despite differences in specification, the fundamental identification problem in all such models remains the same. Insofar as inflation and tightness may plausibly be determined \emph{simultaneously}, the r.h.s.\ of (ref) cannot (in general) be identified as though it were a (nonlinear) regression. Simultaneous causation can instead be addressed by incorporating both (ref), and the corresponding reverse (causal) model for the effect of inflation on tightness, into an ($\mathbb{R}^{2}$-valued) nonlinear function $f_{0}(z_{t})$, where $z_{t}=(\log\theta_{t},\pi_{t})^{\top}$, yielding a specification for the l.h.s.\ of (ref).

As noted in the introduction, we term (ref) an endogenously nonlinear SVAR, because of the possible nonlinearity on the l.h.s.\ of the model, i.e.\ in the endogenous variables $z_{t}$. Were the model instead required to be linear in the endogenous variables, so that $f_{0}(z)=\Phi_{0}z$, identification of the model parameters would be as straightforward as it is in the linear SVAR; and along the lines of (ref)(ref) above, the assumption that $\{\varepsilon_{t}\}$ is independent across time could be weakened to one of $\{\varepsilon_{t}\}$ being merely a martingale difference sequence (with respect to the filtration generated by $\{z_{t}\}$). However, in imposing linearity on the l.h.s., we would lose the possibilities for endogenous regime switching, asymmetric impact multipliers, and of handling occasionally binding constraints, which the general model (ref) affords. We accordingly want to permit $f_{0}$ to be nonlinear: a consequence of which is that independence across time of $\{\varepsilon_{t}\}$ becomes necessary for the parameters of (ref) to be identified. (But note that there is no requirement of contemporaneous independence between the elements of $\varepsilon_{t}$.)

We thus continue to maintain that the structural shocks $\{\varepsilon_{t}\}$ are i.i.d.\ with $\ensuremath{\mathbb{E}}\varepsilon_{t}=0$ and $\ensuremath{\mathbb{E}}\varepsilon_{t}\varepsilon_{t}^{\top}=I_{p}$, and a (Lebesgue) density $\varrho\in\mathscr{R}$. The parameter space for the model (ref) then consists of collections $\mathscr{F}_{0}\nif_{0}$ and $\pmb{\mathscr{F}}_{1}\ni\b{f}_{1}$ of functions $\mathbb{R}^{p}\ensuremath{\rightarrow}\mathbb{R}^{p}$ and $\mathbb{R}^{kp}\ensuremath{\rightarrow}\mathbb{R}^{p}$, and a collection of densities $\mathscr{R}\ni\varrho$ supported on $\mathbb{R}^{p}$, which under our regularity conditions, together determine the conditional density

equation[equation omitted — 161 chars of source]

We continue to regard two alternative parametrisations of the model as being observationally equivalent if they yield the same conditional density. For convenience, we shall suppose throughout that $\b z_{0}$ is continuously distributed, with a density that is a.e.\ strictly positive on $\mathbb{R}^{kp}$. Our assumptions below then ensure that this is also true for every successive $\b z_{t}$, and $\varphi_{z_{t}\mid\b z_{t-1}}(\xi\mid\b{\xi}_{-1})$ is thus well defined for almost every $\xi\in\mathbb{R}^{p}$ and $\b{\xi}_{-1}\in\mathbb{R}^{kp}$, for all $t\geq1$.

Our regularity conditions on the model parameter space, which are sufficient to ensure that the conditional density (ref) exists (and is unique up to the usual a.e.\ equivalence), are as follows.

\needspace{4\baselineskip}

{{{{\scalefont{0.76}PS}}}}

assumption$\mathscr{F}_{0}$, $\pmb{\mathscr{F}}_{1}$ and $\mathscr{R}$ collect every function such that: \begin{enumerate}[label={{{\scalefont{0.76}\arabic*.}}}, ref={{{\scalefont{0.76}\arabic*}}}] • $\tilde{f}_{0}\in\mathscr{F}_{0}$ and $\tilde{\b{f}}_{1}\in\pmb{\mathscr{F}}_{1}$ are locally Lipschitz (continuous); • $\tilde{f}_{0}\in\mathscr{F}_{0}$ is a bijection $\mathbb{R}^{p}\ensuremath{\rightarrow}\mathbb{R}^{p}$, $\tilde{f}_{0}(0)=0$, and $\det D\tilde{f}_{0}(z)\neq0$ for almost every $z\in\mathbb{R}^{p}$; • $\tilde{\varrho}\in\mathscr{R}$ is continuously differentiable, with $\tilde{\varrho}(\varepsilon)>0$ for all $\varepsilon\in\mathbb{R}^{p}$, and \begin{align*} \int_{\mathbb{R}^{p}}\tilde{\varrho}(\varepsilon)\ensuremath{\,\ensuremath{\mathrm{d}}}\varepsilon & =1, & \int_{\mathbb{R}^{p}}\varepsilon\tilde{\varrho}(\varepsilon)\ensuremath{\,\ensuremath{\mathrm{d}}}\varepsilon & =0, & \int_{\mathbb{R}^{p}}\varepsilon\varepsilon^{\top}\tilde{\varrho}(\varepsilon)\ensuremath{\,\ensuremath{\mathrm{d}}}\varepsilon & =I_{p}. \end{align*} \end{enumerate}
rem\refstepcounter{subremark} (\roman{subremark}). Local Lipschitzness implies that $\tilde{f}_{0}$ and $\tilde{\b{f}}_{1}$ are differentiable almost everywhere (a.e.). The r.h.s.\ of (ref) is therefore defined at least almost everywhere, which is sufficient to pin down the conditional density $\varphi_{z_{t}\mid\b z_{t-1}}$, since the latter is itself only uniquely defined up to an a.e.\ equivalence. (See (ref) for further details.) Our smoothness conditions and support conditions on the density $\tilde{\varrho}$ (which accord with those of Matz09Ecta) are maintained only for convenience, and could very likely also be relaxed in this same direction. \refstepcounter{subremark} (\roman{subremark}). Since the nonlinear SVAR (ref) is a (dynamic) nonlinear SEM, our work relates closely to the literature on identification in such models: particularly Matz09Ecta,Matz15Ecta and BH18Ecta. Here we have deliberately relaxed the assumption that the functions $\tilde{f}_{0}$ and $\tilde{\b{f}}_{1}$ are (at least once) continuously differentiable, which is standard in that literature, to allow our results to accommodate models that are continuous but merely piecewise differentiable, such as the piecewise affine SVARs introduced in (ref) below. \refstepcounter{subremark} (\roman{subremark}). We naturally require $\tilde{f}_{0}$ to be invertible, which ensures that the model always yields a solution for the endogenous variables $z_{t}$, irrespective of the values of the predetermined variables $\b z_{t-1}$ and the structural shocks $\varepsilon_{t}$. Requiring $\det D\tilde{f}_{0}(z)\neq0$ a.e.\ merely excludes certain `irregular' cases (our assumptions also imply that this quantity must have the same sign a.e.).

Regarding the parameters $(f_{0},\b f_{1},\varrho)$ that generated $\{z_{t}\}$ in (ref), as distinct from the entirety of the model parameter space, we also maintain the following.

{{{{\scalefont{0.76}DGP}}}}

assumption$(f_{0},\b f_{1},\varrho)$ are such that: \begin{enumerate}[label={{{\scalefont{0.76}\arabic*.}}}, ref={{{\scalefont{0.76}\arabic*}}}] • $\b f_{1}:\mathbb{R}^{kp}\ensuremath{\rightarrow}\mathbb{R}^{p}$ is surjective, with $\operatorname{rk} D\b{f}_{1}(\b z)=p$ for almost every $\b z\in\mathbb{R}^{kp}$; • $f_{0}^{-1}$ is locally Lipschitz; and • $\varrho$ has a local maximum at some $\varepsilon^{\ast}\in\mathbb{R}^{p}$, and is twice continuously differentiable in a neighbourhood of $\varepsilon^{\ast}$, with negative definite Hessian there. \end{enumerate}
rem\refstepcounter{subremark} (\roman{subremark}). We interpret (ref){{{\scalefont{0.76}.}}}(ref) as requiring that there be sufficient dependence of the r.h.s.\ of the model (i.e.\ on the conditional mean of $f_{0}(z_{t})$) on the predetermined variables $\b z_{t-1}$, in both a `global' and `local' sense. (Note that this is only a requirement on the $\b{f}_{1}$ that actually generated the data, which need not be satisfied by all members of $\pmb{\mathscr{F}}_{1}$). For a simple illustration of why some such condition cannot be avoided, consider an extreme case in which $\b{f}_{1}(\b z)=0$ for all $\b z\in\mathbb{R}^{kp}$, so that the r.h.s.\ of (ref) does not depend on $\b z_{t-1}$ at all. Then because $z_{t}=f_{0}^{-1}(\varepsilon_{t})$ will be i.i.d.\ and independent of $\b z_{t-1}$, so too will be \[ \tilde{\varepsilon}_{t}\coloneqq\tilde{f}_{0}(z_{t})=\tilde{f}_{0}[f_{0}^{-1}(\varepsilon_{t})] \] for every $\tilde{f}_{0}\in\mathscr{F}_{0}$. Beyond requiring $\tilde{f}_{0}$ to be scale- and location-normalised such that $\ensuremath{\mathbb{E}}\tilde{\varepsilon}_{t}=0$ and $\ensuremath{\mathbb{E}}\tilde{\varepsilon}_{t}\tilde{\varepsilon}_{t}^{\top}=I_{p}$, the model would therefore yield no meaningful identifying restrictions on $\tilde{f}_{0}$. \refstepcounter{subremark} (\roman{subremark}). (ref){{{\scalefont{0.76}.}}}(ref) is a weak regularity condition on the inverse of $f_{0}$, which would e.g.\ be automatically satisfied if $f_{0}$ were continuously differentiable with $\det Df_{0}(z)\neq0$ for all $z\in\mathbb{R}^{p}$. \refstepcounter{subremark} (\roman{subremark}). (ref){{{\scalefont{0.76}.}}}(ref) would clearly be satisfied if $\varepsilon_{t}$ were Gaussian; but note that only a well-behaved local maximum is required for this condition to hold. The main purpose of this assumption is to allow us to deduce that $u=u^{\ast}$ from merely the equality $f_{U}(u)=f_{U}(u^{\ast})$, and further regulate the behaviour of $f_{U}$ in the vicinity of $u^{\ast}$. Though their model and proofs differ significantly from ours -- in particular, because their counterpart of our $\b{f}_{1}$ has the property that each component depends on a variable (an `instrument') that is special to that component -- it is noteworthy that a similar assumption is introduced by BH18Ecta as their Condition M (see also their Corollary 2).

Remarkably, despite the far greater flexibility afforded by the nonlinear parametrisation of (ref), under the foregoing conditions we obtain the following, effectively identical characterisation of observational equivalence to that of the linear SVAR (ref), the proof of which appears in (ref).

thmSuppose (ref) and (ref) hold, and let $\tilde{f}_{0}\in\mathscr{F}_{0}$ and $\tilde{\b{f}_{1}}\in\pmb{\mathscr{F}}_{1}$. Then there exists a $\tilde{\varrho}\in\mathscr{R}$ such that $(\tilde{f}_{0},\tilde{\b f}_{1},\tilde{\varrho})$ is observationally equivalent to $(f_{0},\b f_{1},\varrho)$, if and only if there exists a $Q\in\mathbb{O}(p)$ such that \begin{align} \tilde{f}_{0}(z) & =Qf_{0}(z),\ \forall z\in\mathbb{R}^{p}, & \tilde{\b f}_{1}(\boldsymbol{z}) & =Q\b f_{1}(\boldsymbol{z}),\ \forall\boldsymbol{z}\in\mathbb{R}^{kp}. \end{align}
rem\refstepcounter{subremark} (\roman{subremark}). Here we are asking whether for given candidate functions $(\tilde{f}_{0},\tilde{\b{f}}_{1})$, it is possible to find a distribution $\tilde{\varrho}\in\mathscr{R}$ for the structural shocks such that \[ \varrho[f_{0}(\xi)-\b{f}_{1}(\b{\xi}_{-1})]\cdot\smlabs{\det Df_{0}(\xi)}=\tilde{\varrho}[\tilde{f}_{0}(\xi)-\tilde{\b{f}}_{1}(\b{\xi}_{-1})]\cdot\smlabs{\det D\tilde{f}_{0}(\xi)} \] holds for almost every $\xi\in\mathbb{R}^{p}$ and $\b{\xi}_{-1}\in\mathbb{R}^{kp}$. The $\tilde{\varrho}$ delivering this equivalence will, for $Q$ as in (ref), be given by the density of \[ \tilde{\varepsilon}_{t}=\tilde{f}_{0}(z_{t})-\tilde{\b{f}}_{1}(\boldsymbol{z}_{t-1})=Q[f_{0}(z_{t})-\b{f}_{1}(\boldsymbol{z}_{t-1})]=Q\varepsilon_{t}, \] which under our assumptions will also lie in $\mathscr{R}$. This implies that the introduction of further (e.g.\ parametric) assumptions on the set $\mathscr{R}$ of allowable densities would not yield any further tightening of our characterisation of observational equivalence, provided that $\mathscr{R}$ remains closed under orthogonal transformations of the variables: as would e.g.\ be the case even if $\mathscr{R}$ were restricted to the set of Gaussian densities on $\mathbb{R}^{p}$ (with mean zero and identity covariance). \refstepcounter{subremark} (\roman{subremark}). The foregoing is a nonparametric identification result, in the sense that neither $(f_{0},\b{f}_{1})$, nor the distribution $\varrho$ of the shocks, are assumed to have any particular (known) parametric form. In practice, however, we would expect the model (ref) to be formulated parametrically, if only because the limited length of the time series available, for most macroeconomic applications, render genuine nonparametric estimation infeasible. In the abstract setting of (ref), these parametric functional form and/or distributional assumptions can be understood as restrictions on the sets $\mathscr{F}_{0}$, $\pmb{\mathscr{F}}_{1}$ and $\mathscr{R}$. The conclusion of the theorem continues to hold in such cases, provided that $\mathscr{R}$ is not so (unusually) constrained that it fails to satisfy the invariance condition noted in the previous remark. See (ref) below for the discussion of a class of parametric models (for $f_{0}$ and $\b{f}_{1}$) for which the conditions required by the theorem may be verified straightforwardly. \refstepcounter{subremark} (\roman{subremark}). As noted above, a consequence of the Markov property of the SVAR is that the notion of observational equivalence appropriate to our setting refers only to the distribution $z_{t}\mid\b z_{t-1}$ of the endogenous variables conditional on the exogenous variables; it therefore coincides exactly with that employed by Matz09Ecta in the context of a (non-dynamic) nonlinear SEM: see her (3.1), in particular. This allows the proof of (ref) to be approached just as if we were analysing identification in a nonlinear SEM, a connection that we draw out more fully in (ref). Relative to the results in the existing SEM literature, we obtain a much tighter characterisation of observational equivalence because of the separability between $z_{t}$ and $\b z_{t-1}$. \refstepcounter{subremark} (\roman{subremark}). Should (ref) fail to hold, then there will be at least some realisations of $\{z_{t}\}$ for which the likelihoods of $(\tilde{f}_{0},\tilde{\b f}_{1},\tilde{\varrho})$ and $(f_{0},\b f_{1},\varrho)$ will be distinct, and so the data will to this extent be informative about these two alternative parametrisations of the model. However, we would not claim, on the basis of this theorem alone, that the parameters of the model are consistently estimable up to an orthogonal transformation. While it seems reasonable to suppose that consistent nonparametric estimation of the model would be possible (under suitable regularity conditions) when $\{z_{t}\}$ is stationary and ergodic, the familiar connection between identification and consistent estimation is attenuated when $\{z_{t}\}$ possesses stochastic (or indeed, deterministic) trends, because of the non-recurrence of those trends in higher dimensions (see Bing01Hdbk; GP13JoE). Consistent estimation of the model parameters (up to $Q$) would in such cases likely require further restrictions on $(f_{0},\b{f}_{1})$, such as those sufficient to ensure that $\{z_{t}\}$ is indeed stationary and ergodic (for a discussion of such conditions in this context, see DMW23stat, and the references cited therein).

Orthogonal reduced-form parametrisation

Analogously (though not identically) to the `orthogonal reduced-form parametrisation' (ARRW18Ecta, Sec.\ 2.3) of the linear SVAR, (ref) suggests the following convenient reparametrisation of the endogenously nonlinear SVAR. Let $z_{0}\in\mathbb{R}^{p}$ be fixed, and a point at which $f_{0}$ is (assumed to be) differentiable, with full rank Jacobian $Df_{0}(z_{0})$. By the QR decomposition, we have $Df_{0}(z_{0})=Q^{\top}L$, where $L$ is lower triangular, and $Q\in\mathbb{O}(p)$; define $(g_{0},\b{g}_{1})\coloneqq(Qf_{0},Q\b{f}_{1})$. Multiplying (ref) through by $Q$, we may reformulate the model as

equation[equation omitted — 85 chars of source]

where now $g_{0}$ is restricted such that $Dg_{0}(z_{0})$ is lower triangular (which need hold only at that chosen $z_{0}$), and $Q\in\mathbb{O}(p)$.

This yields an equivalent parametrisation of the model, in which the parameter spaces for $\b{g}_{1}\in\b{\mathscr{F}}_{1}$ and $\varrho\in\mathscr{R}$ remain as before, but now $\mathscr{F}_{0}$ is additionally restricted (beyond (ref){{{\scalefont{0.76}.}}}(ref){{{\scalefont{0.76}--}}}(ref)) to functions $g_{0}:\mathbb{R}^{p}\ensuremath{\rightarrow}\mathbb{R}^{p}$ for which $Dg_{0}(z_{0})$ is lower triangular (at the nominated $z_{0}\in\mathbb{R}^{p}$); let $\mathscr{F}_{0}^{(z_{0})}$ denote the resulting parameter space for $g_{0}$. To exactly offset this restriction, we now have the additional parameter $Q\in\mathbb{O}(p)$, so that we may equivalently regard the nonlinear SVAR as being parametrised by $(g_{0},\b{g}_{1},Q,\varrho)\in\mathscr{F}_{0}^{(z_{0})}\times\pmb{\mathscr{F}}_{1}\times\mathbb{O}(p)\times\mathscr{R}$. The import of (ref) here is that the parameters $(g_{0},\b{g}_{1})\in\mathscr{F}_{0}^{(z_{0})}\times\pmb{\mathscr{F}}_{1}$ are exactly identified by data on $\{z_{t}\}$, with the non-identified part of the structural parameters being transferred entirely to $Q$. The `nonlinear SVAR identification problem' can thus be framed precisely as one of finding sufficient restrictions to pin down $Q$, from which the structural parameters may then be recovered, via $(f_{0},\b{f}_{1})=(Q^{\top}g_{0},Q^{\top}\b{g}_{1})$.

The reparametrisation (ref) provides a convenient perspective from which to import various approaches to identifying $Q$ from the linear SVAR literature. For the most part, these apply directly to the present setting, with little modification required. The following example illustrates how it remains possible to identify impulse responses via external instruments, without requiring any additional assumptions relative to those needed to identify the linear VAR.

example[external instruments] Suppose that $w_{t}$ is a (scalar) `external instrument': an observed (stationary) process that is assumed to be contemporaneously correlated with the first structural shock, but not with any of the others (see e.g.\ SW18EJ, p.\ 931). Defining \begin{equation} u_{t}\coloneqqg_{0}(z_{t})-\b{g}_{1}(z_{t-1})=Q\varepsilon_{t} \end{equation} which by (ref) is identified, we must have \[ \delta\coloneqq\ensuremath{\mathbb{E}} u_{t}w_{t}=Q\ensuremath{\mathbb{E}}\varepsilon_{t}w_{t}=Qe_{1}\alpha=q_{1}\alpha, \] where $q_{1}$ denotes the first column of $Q$, and $\alpha=\ensuremath{\mathbb{E}}\varepsilon_{1t}w_{t}\neq0$. Since $\delta=\ensuremath{\mathbb{E}} u_{t}w_{t}$ is identified, so too is $q_{1}=\delta/\smlnorm{\delta}$, and we can further recover $\varepsilon_{1t}=q_{1}^{\top}u_{t}$. Since the distribution of $u_{t}$ in (ref) is identified, so too is the conditional distribution \[ u_{t}\mid\{\varepsilon_{1t}=\bar{\varepsilon}_{1}\}=_{d} u_{t}\mid\{q_{1}^{\top}u_{t}=\bar{\varepsilon}_{1}\}. \] For given values of $\b z_{t-1}=\bar{\b z}$ and $\bar{\varepsilon}_{1}$, the distribution of the counterfactual quantity \[ z_{t}(\bar{\b z},\bar{\varepsilon}_{1})\mid\{\b z_{t-1}=\bar{\b z},\varepsilon_{1t}=\bar{\varepsilon}_{1}\}=_{d}g_{0}^{-1}(\b{g}_{1}(\bar{\b z})+u_{t})\mid\{q_{1}^{\top}u_{t}=\bar{\varepsilon}_{1}\} \] depends only on $g_{0}$, $\b{g}_{1}$ and the distribution of $u_{t}\mid\{q_{1}^{\top}u_{t}=\bar{\varepsilon}_{1}\}$, all of which are identified. In this way, the impact multipliers of $\varepsilon_{1t}$ may be recovered; the impulse responses at further horizons depend, by the Markov property, additionally only on the conditional distribution of $z_{t}\mid\b z_{t-1}$, which is trivially identified.

Piecewise affine SVARs

Endogenous regime switching

Here we introduce a class of endogenously regime-switching models, in which the conditions required for our results may be verified relatively straightforwardly. Models of this form have been used recently to study monetary policy under an occasionally binding constraint on nominal interest rates: see SM21, AMSV21, ILMZ20, and CCMM25.

Suppose now that the l.h.s.\ of the nonlinear SVAR

equation[equation omitted — 91 chars of source]

is specified as

equation[equation omitted — 147 chars of source]

for $\{\set Z^{(\ell)}\}_{\ell=1}^{L}$ a collection of convex sets that partition $\mathbb{R}^{p}$, $\{\bar{\phi}_{0}^{(\ell)}\}_{\ell=1}^{L}\subset\mathbb{R}^{p}$ and $\{\Phi_{0}^{(\ell)}\}_{\ell=1}^{L}\subset\mathbb{R}^{p\times p}$. When these parameters are such that $f_{0}$ is continuous, we shall say that $f_{0}$ is a piecewise affine function. (We do not consider cases in which $f_{0}$ may be discontinuous, so continuity should always be taken as implied.) The model may then be regarded as consisting of $L$ `regimes' demarcated by the sets $\{\set Z^{(\ell)}\}_{\ell=1}^{L}$. Which of those regimes is operative in period $t$, i.e.\ the value of $\ell_{t}\in\{1,\ldots,L\}$ such that \[ f_{0}(z_{t})=\bar{\phi}_{0}^{(\ell_{t})}+\Phi_{0}^{(\ell_{t})}z_{t} \] is determined jointly with the value of $z_{t}$. For this reason, we say that there is endogenous switching between the $L$ regimes, as distinct from the exogenous regime switching that would result if $\ell_{t}$ were determined prior to the realisation of $z_{t}$. The situation here is thus markedly different from the regime-switching SVARs considered in the previous literature, which as noted in the introduction, can generally be written in the form \[ \Phi_{0}(s_{t-1})z_{t}=c(s_{t-1})+\sum_{i=1}^{k}\Phi_{i}(s_{t-1})z_{t-i}+\varepsilon_{t}, \] where $s_{t-1}$ is determined prior to $\varepsilon_{t}$ and $z_{t}$ (see e.g.\ AG12AEJ; CCCN15EJ; BP24EctJ).

\setcounter{savedexnumber}{\value{examplex}}

{{\ref*{exa:phillips}}}

example[nonlinear Phillips curve; ctd] The nonlinear Phillips curve of BenignoEggertsson2023, in (ref) above, is piecewise affine (and continuous) with a kink at $\log\theta_{t}=0$, which the authors refer to as the `Beveridge threshold'. Their model thus delineates two distinct labour market regimes: a `normal' regime ($\ell_{t}=1$), when the labour market is slack, $\log\theta_{t}\leq0$, and a `labour shortage' regime ($\ell_{t}=2$) in which $\log\theta_{t}>0$. The regime $\ell_{t}$ is entirely driven by the contemporaneous value of the endogenous variable $\log\theta_{t}$, and so the regime-switching is genuinely endogenous. Their Phillips curve (ref) can also be written as \begin{equation} \pi_{t}=\pi+\kappa^{(\ell_{t})}\log\theta_{t}+\eta_{t}. \end{equation} where $\kappa^{(1)}=\kappa$ and $\kappa^{(2)}=\kappa^{\mathrm{tight}}$. Contrast this with an alternative specification in which the slope of the Phillips curve is determined by past values of $\log\theta_{t}$, for example \begin{equation} \pi_{t}=\pi+\kappa^{(\ell_{t-1})}\log\theta_{t}+\eta_{t}. \end{equation} Conditional on the past (i.e.\ on time $t-1$), (ref) is linear in $z_{t}=(\log\theta_{t},\pi_{t})^{\top}$, and so shocks to $\log\theta_{t}$ will have the same proportional effect $\kappa^{(\ell{}_{t-1})}$ on $\pi_{t}$, irrespective of their sign; whereas in (ref) the impact of the shocks will vary additionally (and nonlinearly) depending on the initial (i.e.\ pre-shock) proximity of tightness to the Beveridge threshold.

\setcounter{examplex}{\value{savedexnumber}} \numberwithin{examplex}{section}

Identification

We would like primitive conditions that ensure $f_{0}$ in (ref) satisfies the requirements (ref){{{\scalefont{0.76}.}}}(ref){{{\scalefont{0.76}-}}}(ref) and (ref){{{\scalefont{0.76}.}}}(ref) of (ref): namely, that both it and its inverse should be locally Lipschitz, and that it should be (globally) invertible, with $\det Df_{0}(z)\neq0$ a.e. Two important special cases of (ref), for which these conditions may be readily verified, are those of:

itemize• a (continuous) piecewise linear function, in which there exists a basis $\{a_{i}\}_{i=1}^{p}$ for $\mathbb{R}^{p}$ such that each $\set Z^{(\ell)}$ can be written as a union of cones of the form \begin{equation} \set C_{{\cal I}}\coloneqq\{z\in\mathbb{R}^{p}\mid a_{i}^{\top}z\geq0,\ \forall i\in{\cal I} and a_{i}^{\top}z<0,\ \forall i\notin{\cal I}\} \end{equation} where ${\cal I}$ ranges over the subsets of $\{1,\ldots,p\}$, and $\bar{\phi}_{0}^{(\ell)}=0$ for all $\ell\in\{1,\ldots,L\}$; and • a (continuous) threshold affine function, in which there exists an $a\in\mathbb{R}^{p}\backslash\{0\}$ and thresholds $\{\tau_{\ell}\}_{\ell=0}^{L}$ with $\tau_{\ell}<\tau_{\ell+1}$, $\tau_{0}=-\infty$ and $\tau_{L}=+\infty$, such that \[ \set Z^{(\ell)}=\{z\in\mathbb{R}^{p}\mid a^{\top}z\in(\tau_{\ell-1},\tau_{\ell}]\}, \] i.e.\ the sets $\{\set Z^{(\ell)}\}$ take the forms of `bands' in $\mathbb{R}^{p}$. (In typical examples, $a=e_{p,i}$, i.e.\ it picks out one `threshold variable' from the elements of $z_{t}$.)

Because the boundaries between the regimes are then affine subspaces (of $\mathbb{R}^{p}$), ensuring the continuity of $f_{0}$ is a straightforward matter of linearly restricting the elements of $\{\bar{\phi}_{0}^{(\ell)}\}_{\ell=1}^{L}$ and $\{\Phi_{0}^{(\ell)}\}_{\ell=1}^{L}$ such that the values prescribed by adjacent regimes agree on those boundaries; see the next appearance of (ref) for an illustration. Regarding our remaining requirements on $f_{0}$, for these it is necessary and sufficient that

equation[equation omitted — 147 chars of source]

See (ref) below; we note that equivalence of the preceding with the invertibility of $f_{0}$ follows directly from Theorems 1 and 4 of GLM80Ecta, and that since $Df_{0}(z)=\sum_{\ell=1}^{L}\ensuremath{\mathbf{1}}\{z\in\set Z^{(\ell)}\}\Phi_{0}^{(\ell)}$ a.e., the Jacobian is then clearly invertible a.e.

\setcounter{savedexnumber}{\value{examplex}}

{{\ref*{exa:phillips}}}

example[nonlinear Phillips curve; ctd] The nonlinear Phillips curve (ref) is piecewise linear, with $z_{t}=(\log\theta_{t},\pi_{t})^{\top}$ and two regimes \begin{align*} \mathscr{Z}^{(1)} & =\{z\in\mathbb{R}^{2}\mid e_{1}^{\top}z\leq0\} & \mathscr{Z}^{(2)} & =\{z\in\mathbb{R}^{2}\mid e_{1}^{\top}z>0\} \end{align*} which can each be written as unions of cones of the form (ref) (e.g.\ by taking $a_{1}=-e_{1}$ and $a_{2}=e_{2}$). (ref) specifies only the first component of the bivariate map $f_{0}(z_{t})$. If the second component is also modelled as piecewise linear, with regimes also determined by the sign of $\log\theta_{t}$ (thus linear on each of the sets $\mathscr{Z}^{(1)}$ and $\mathscr{Z}^{(2)}$), then $f_{0}$ admits the representation (ref). To ensure continuity at the regime boundary where $\log\theta_{t}=0$, we need the equality \[ \bar{\phi}_{0}^{(1)}+\begin{bmatrix}\Phi_{0,1}^{(1)} & \Phi_{0,2}^{(1)}\end{bmatrix}\begin{bmatrix}0\\ \pi_{t} \end{bmatrix}=\bar{\phi}_{0}^{(2)}+\begin{bmatrix}\Phi_{0,1}^{(2)} & \Phi_{0,2}^{(2)}\end{bmatrix}\begin{bmatrix}0\\ \pi_{t} \end{bmatrix} \] to hold for all values of $\pi_{t}\in\mathbb{R}$, where $\Phi_{0}^{(\ell)}=[\Phi_{0,1}^{(\ell)},\Phi_{0,2}^{(\ell)}]$. This entails \begin{align*} \bar{\phi}_{0}^{(1)}-\bar{\phi}_{0}^{(2)} & =0 & \Phi_{0,2}^{(1)}-\Phi_{0,2}^{(2)} & =0, \end{align*} and we may also impose $\bar{\phi}_{0}^{(1)}=0$, for the location normalisation $f_{0}(0)=0$. To put it another way, continuity requires that only the coefficients on the regime-determining variable $\log\theta_{t}$ may change at the threshold, leading to the (non-redundant) specification \begin{equation} \Phi_{0}^{(\ell)}=[\Phi_{0,1}^{(\ell)},\Phi_{0,2}],\ \ell\in\{1,2\} \end{equation} in which the second column of the coefficient matrix is regime-invariant.

\setcounter{examplex}{\value{savedexnumber}} \numberwithin{examplex}{section}

The SVAR specification (ref)--(ref) thus provides a flexible but tractable means of introducing nonlinearity into an SVAR model. This is especially the case if we also specify that the r.h.s.\ should be additively time-separable, and of the same functional form as the l.h.s., so that

equation[equation omitted — 140 chars of source]

where now, for every $i\in\{0,\ldots,k\}$,

equation[equation omitted — 150 chars of source]

is (continuous) piecewise affine. (Note that there is no need to additionally index the regimes $\set Z^{(\ell)}$ by $i$ here, since if the partitions $\{\set Z_{i}^{(\ell)}\}_{\ell=1}^{L_{i}}$ did vary across $i$, we could always find a mutual refinement such that (ref) held for all $i$.) We term this model a piecewise affine SVAR; with piecewise linear and threshold affine SVARs corresponding to those cases where the $f_{i}$'s are either all piecewise linear or all threshold affine functions, respectively.

The conditions (ref){{{\scalefont{0.76}.}}}(ref) and (ref){{{\scalefont{0.76}.}}}(ref) imposed by (ref) on $\b{f}_{1}(\b z_{t-1})=\sum_{i=1}^{k}f_{i}(z_{t-i})$ are rather less taxing than those imposed on $f_{0}$. Under the specification (ref), continuity is readily imposed, and then automatically implies Lipschitz continuity. Moreover, $D\b f_{1}(\b z_{t-1})$ a.e.\ exists and satisfies \[ D\b{f}_{1}(\b z)=

bmatrix[bmatrix omitted — 91 chars of source]

\] for some $\ell_{i}\in\{1,\ldots,L\}$ depending on $\b z$, and so it is easy to verify whether $\operatorname{rk} D\b{f}_{1}(\b z)=p$ a.e.\ (or, since this holds generically, to test the null hypothesis of a deficient rank). In practice, this may be analysed more straightforwardly on the basis of the coefficients on the first lag or two of $z_{t}$, which may themselves be sufficient to satisfy this condition.

Verifying the (global) surjectivity condition on $\b{f}_{1}$ is a little more challenging, because of the apparent absence of a counterpart to (ref) for this case. In the special case of a model with only one lag, surjectivity of $z_{t-1}\ensuremath{\mapsto}f_{1}(z_{t-1})$ is equivalent to (ref). Though easy to check, this is far more than is necessary for surjectivity when additional lags are present. Alternatively, if some elements of $z_{t}$ enter $f_{i}$ linearly, as will often be the case in practice (as in our next example), then surjectivity holds so long as the coefficient vectors associated with (at least) $p$ of these variables (drawn from across the $k$ lags of $z_{t}$ appearing on the r.h.s.)\ form a rank $p$ matrix.

example[occasionally binding constraint] SM21 proposed the censored and kinked structural VAR (CKSVAR), to model the effects of the zero lower bound (ZLB) constraint on monetary policy: see also AMSV21 and CCMM25. In his setting, $y_{t}$ is a scalar variable whose positive part $y_{t}^{+}\coloneqq\max\{y_{t},0\}$ coincides with the central bank's policy rate (constrained to be non-negative), while its (latent) negative part $y_{t}^{-}\coloneqq\min\{y_{t},0\}$ is the `shadow rate', which summarises the stance of monetary policy desired by the central bank when the ZLB binds, to be engineered via `unconventional' policy, such as asset purchases. The remaining variables in the model are collected in the $(p-1)$-dimensional vector $x_{t}$, in his case the inflation and unemployment rates; we then set $z_{t}=(y_{t},x_{t}^{\top})^{\top}$. To allow for possibility that the ZLB might actually constrain monetary policy, $y_{t}^{+}$ and $y_{t}^{-}$ are permitted to enter the model with different coefficients (in possibly all $p$ equations), \begin{equation} \phi_{0}^{+}y_{t}^{+}+\phi_{0}^{-}y_{t}^{-}+\Phi_{0}^{x}x_{t}=c+\sum_{i=1}^{k}[\phi_{i}^{+}y_{t-i}^{+}+\phi_{i}^{-}y_{t-i}^{-}+\Phi_{i}^{x}x_{t-i}]+u_{t} \end{equation} where $\phi_{i}^{\pm}\in\mathbb{R}^{p}$ and $\Phi_{i}^{x}\in\mathbb{R}^{p\times(p-1)}$, for $i\in\{0,\ldots,k\}$. This may be rendered as an instance of a threshold affine SVAR by defining\begin{subequations} \begin{align} \set Z^{(1)} & \coloneqq\set Z^{-}=\{(y,x)\in\mathbb{R}^{p}\mid y\leq\tau_{1}\} & \Phi_{i}^{(1)} & \coloneqq[\phi_{i}^{-},\Phi_{i}^{x}]\\ \set Z^{(2)} & \coloneqq\set Z^{+}=\{(y,x)\in\mathbb{R}^{p}\mid y>\tau_{1}\} & \Phi_{i}^{(2)} & \coloneqq[\phi_{i}^{+},\Phi_{i}^{x}], \end{align} \end{subequations} with $\tau_{1}=0$, and then setting $f_{i}(z)=\sum_{\ell=1}^{2}\ensuremath{\mathbf{1}}\{z\in\set Z^{(\ell)}\}\Phi_{i}^{(\ell)}z$. (Because there are only two regimes, it may also be equivalently cast as a piecewise linear SVAR.) Here continuity of each $f_{i}$ is guaranteed by the fact that $\Phi_{i}^{(1)}$ and $\Phi_{i}^{(2)}$ only differ by their first column; or equivalently by the linear restrictions $(\Phi_{i}^{(1)}-\Phi_{i}^{(2)})E_{-1}=0$, for $E_{-1}$ the final $p-1$ columns of $I_{p}$. In SM21, identification of the parameters of this model are complicated by the fact that $y_{t}$ is only observed when $y_{t}>0$; it is in effect censored at zero. His results therefore do not fall within the framework of (ref), which implicitly assumes that $z_{t}$ and $\b z_{t-1}$ are observed on the entirety of their supports. However, the model (ref)--(ref) may (of course) also be applied to settings in which $y_{t}$ is observed on both sides of the threshold $\tau_{1}$, which may be treated as an additional unknown parameter to be identified and estimated. From the foregoing discussion, for $f_{0}$ to satisfy the conditions of (ref), we would need only to verify that $\det\Phi_{0}^{(1)}$ and $\det\Phi_{0}^{(2)}$ are both nonzero, and have the same sign. Regarding $\b{f}_{1}$: if $k\geq2$ then it is sufficient to check (or rather, test) whether the $p\times k(p-1)$ matrix $[\Phi_{1}^{x},\ldots,\Phi_{k}^{x}]$, formed from the coefficients on the lags of $x_{t}$, has rank $p$; whereas if $k=1$, then we would need $\{\Phi_{1}^{(\ell)}\}$ to satisfy the same determinantal condition as $\{\Phi_{0}^{(\ell)}\}$ (a condition also sufficient when $k\geq2$).

Smooth transitions

The piecewise affine SVAR (ref)--(ref) may be extended to allow for `smooth transitions' between the $L$ regimes. In the literature on smooth transition (vector) autoregressive models, the conventional approach (e.g.\ HT13) is to replace the indicator functions $\ensuremath{\mathbf{1}}\{z\in\set Z^{(\ell)}\}$ in (ref) by smooth maps $\pi^{(\ell)}(z)$, so that now \[ f_{i}^{\mathrm{ST}}(z)=\sum_{\ell=1}^{L}\pi^{(\ell)}(z)(\bar{\phi}_{i}^{(\ell)}+\Phi_{i}^{(\ell)}z), \] where $\pi^{(\ell)}(z)\in[0,1]$ and $\sum_{\ell=1}^{L}\pi^{(\ell)}(z)=1$ for all $z\in\mathbb{R}^{p}$, so that $f_{i}^{\mathrm{ST}}(z)$ is always a smooth, convex combination of the affine functions $z\ensuremath{\mapsto}\bar{\phi}_{i}^{(\ell)}+\Phi_{i}^{(\ell)}z$, for $\ell\in\{1,\ldots,L\}$. However, the fact that the gradient of $f_{i}^{\mathrm{ST}}$ is not a convex combination of those underlying affine regimes makes it difficult to reduce the high-level conditions of (ref) to a set of verifiable conditions on the underlying regime-specific coefficient matrices, in the manner of (ref). Indeed, as the simple example in (ref) illustrates, it may well be the case that $f_{0}^{\mathrm{ST}}$ is not invertible, even though its unsmoothed counterpart $f_{0}$ is.

As an alternative specification that allows for smooth transitions between regimes, but which also retains the simplicity -- in terms of verifying the conditions for (ref) -- enjoyed by piecewise affine models, consider

equation[equation omitted — 129 chars of source]

where $f_{i}$ is a (continuous) piecewise affine function as in (ref) above, and $K$ is a smooth (kernel) density function with mean zero, with $m\geq1$ continuous derivatives that satisfy the integrability condition

equation[equation omitted — 185 chars of source]

where $\partial_{u_{i}}$ denotes the partial derivative with respect to the $i$th element of $u\in\mathbb{R}^{p}$, for $\alpha_{i}\in\{1,\ldots,p\}$ and $1\leq n\leq m$.

figure[figure omitted — 510 chars of source]

Our next result establishes that $f_{0,K}(z)$ is smooth (with as many continuous derivatives as $K$ has), and moreover invertible if the determinantal condition (ref) is satisfied (its proof appears in (ref)). Recall that a function is said to be bi-Lipschitz if both it and its inverse are Lipschitz continuous.

propSuppose that $f_{0}:\mathbb{R}^{p}\ensuremath{\rightarrow}\mathbb{R}^{p}$ is as in (ref), and is either a piecewise linear or threshold affine function. Then: \begin{enumerate} • $f_{0}$ is invertible and bi-Lipschitz if and only if (ref) holds. \end{enumerate} Suppose that $K:\mathbb{R}^{p}\ensuremath{\rightarrow}\mathbb{R}$ is $m\geq1$ times continuously differentiable and non-negative, satisfying $\int_{\mathbb{R}^{p}}K(u)=1$, $\int_{\mathbb{R}^{p}}uK(u)\ensuremath{\,\ensuremath{\mathrm{d}}} u=0$ and (ref), and that $f_{0,K}$ is formed by convolving $f_{0}$ with $K$, as in (ref). Then if (ref) holds: \begin{enumerate}[resume] • $f_{0,K}$ is invertible, bi-Lipschitz, and $m$ times continuously differentiable. \end{enumerate}

Application: a nonlinear Phillips curve?

Formulation as an endogenous regime-switching SVAR

The inflation surge that followed the COVID-19 pandemic reignited academic interest in the possibility of nonlinearity in the transmission of supply shocks to inflation. However, views on the relevance of nonlinearity are divided. On the one hand, BallLeighMishra2022 and BenignoEggertsson2023 find evidence of significant nonlinearity in their formulations of the Phillips curve, and argue that nonlinearity is needed to account for the recent inflation surge. On the other hand, BeaudryHouPortier2025 caution that the evidence on nonlinearity is not robust to functional form assumptions, especially as pertains to the treatment of expectations. Reconsidering this debate, through the lens of an endogenous regime-switching SVAR, provides an illustrative application of the methodology developed in this paper.

Our identification result can be useful in this debate because it highlights the following: since all observationally equivalent structures are identified up to a (linear) orthogonal transformation, then if one finds no (statistically significant) evidence of nonlinearity under one specific identification scheme, then this will remain true irrespective of how the model is identified. Indeed, one can see from the orthogonal reduced-form parametrisation developed in (ref) above, that the structural parameters $f_{0}$ (and $\b{f}_{1}$) will be nonlinear if and only if their normalised (and exactly identified) counterparts $g_{0}$ (and $\b{g}_{1}$) are also nonlinear, as will be the case for $Qf_{0}$ (and $Q\b{f}_{1}$) for any $Q\in\mathbb{O}(p)$. Thus the presence of nonlinearity can be tested for in a way that is robust to the identifying scheme employed. To be clear, this is a consequence of modelling the joint determination of $z_{t}=(\log\theta_{t},\pi_{t})^{\top}$ in its entirety; the argument does not carry over to the methodology employed in the aforementioned papers, because these provide only a single-equation analysis of the Phillips curve, and so their findings are potentially contingent on the assumptions made in order to identify that equation.

Building on the development already given to this point in (ref), inspired by the recent work of BenignoEggertsson2023 we consider the following endogenous regime-switching SVAR for $z_{t}=(\log\theta_{t},\pi_{t})^{\top}$,

align[align omitted — 213 chars of source]

where $\theta_{t}=v_{t}/u_{t}$ is the vacancy--unemployment ratio, $\pi_{t}$ is consumer price inflation, $\varepsilon_{t}$ are the structural shocks, and

equation[equation omitted — 224 chars of source]

where $z_{1t}=\log\theta_{t}$. This model thus has two regimes, determined by the sign of $z_{1t}$. Following the arguments that led to (ref) above, to ensure continuity of the model in both $z_{t}$ and its lags, we parametrise the regime-dependent coefficient matrices non-redundantly as

equation[equation omitted — 298 chars of source]

so that only the coefficients of the regime-determining variable, $z_{1t}$, are permitted to vary across the two regimes. The model is then guaranteed to yield a solution for $z_{t}$, for every possible value of the r.h.s.\ of (ref), provided that $\det\Phi_{0}^{(1)}\cdot\det\Phi_{0}^{(2)}>0$.

To obtain a just-identified specification, by (ref) it suffices to impose $p(p-1)/2=1$ restrictions on the model parameters (see also the discussion in (ref) above). For some identifying schemes, this may involve imposing a restriction on only one of the two regimes. However, the identifying assumption in BenignoEggertsson2023 corresponds to the `recursive' or `Cholesky' restriction under which (a shock to) inflation $\pi_{t}=z_{2t}$ has no contemporaneous effect on tightness $\log\theta_{t}=z_{1t}$, and thus that the matrix $\Phi_{0}^{(\ell)}$ is lower triangular for $\ell\in\{1,2\}$. In view of (ref), this in fact constitutes only a single restriction on the model parameters, that $\Phi_{0,12}=0$, and so is exactly identifying rather than over-identifying. The second equation of the nonlinear SVAR (ref) can in this case be estimated by nonlinear regression (with $\pi_{t}$ as the dependent variable), as was done by BenignoEggertsson2023.

Testing for linearity in the Phillips curve

Let $\{\Gamma_{i}^{(\ell)}\}$ momentarily denote the SVAR parameters corresponding to the recursive identification scheme of BenignoEggertsson2023. In light of (ref), because of the lower-triangular structure imposed on the Jacobian $\Phi_{0}^{(1)}$ of $f_{0}$ (at some nominated point $z_{0}$ in the `normal' regime), these are the coefficients associated with the orthogonal reduced-form parametrisation (ref) of the SVAR, when the $g_{j}$ are modelled as piecewise linear. (ref), together with a sign-normalisation of the shocks, then implies that all observationally equivalent models can be obtained by a common rotation of the recursively identified model, i.e.\ $\Phi_{i}^{(\ell)}=Q\Gamma_{i}^{(\ell)}$ for $\ell\in\{1,2\}$ and $i\in\{0,\ldots k\}$, where $Q\in\mathbb{O}(p)$ with $\det Q>0$.

Because $Q$ is not regime dependent, every observationally equivalent parametrisation of the model obtained in this way will exhibit regime dependence if, and only if, this is also true of the parameters $\{\Gamma_{i}^{(\ell)}\}$ obtained under the BenignoEggertsson2023 identification scheme. The presence of some regime dependence in $\{\Gamma_{i}^{(\ell)}\}$ is thus a necessary condition for the existence of a nonlinear Phillips curve under any identification scheme. Since the null hypothesis of no regime dependence in $\{\Gamma_{i}^{(\ell)}\}$ is testable, a failure to reject it would provide evidence, in favour of a linear Phillips curve, that is robust to all possible identifying schemes. (In this respect, our imposition of the BenignoEggertsson2023, restrictions merely provides a convenient way to normalise the system, in the manner of (ref)).

Observe that the specification of (ref) allows for nonlinearities in all lags of the SVAR. This permits the dynamic response of inflation to labour market tightness shocks to be nonlinear, even if the impact responses are linear, i.e., even if $\Phi_{0}^{(\ell)}$ is regime-invariant. We therefore consider two separate tests of linearity. The first tests

equation[equation omitted — 185 chars of source]

The null hypothesis $H_{0}^{\mathrm{NS}}$ can be interpreted as saying that there is no endogenous regime switching, and implies that the impact effect of labour tightness shocks on inflation does not depend on the state of the labour market.

However, $H_{0}^{\mathrm{NS}}$ does not exclude the possibility that $\Phi_{i}^{(1)}\neq\Phi_{i}^{(2)}$ for some $i\in\{1,\ldots,k\}$, in which case the dynamic effects of tightness shocks may still be regime dependent, at longer horizons. This motivates our second, more restrictive hypothesis:

equation[equation omitted — 211 chars of source]

which under the null entails a linear SVAR. Failure to reject $H_{0}^{\mathrm{lin}}$ would suggest that a linear SVAR provides an adequate description of the dynamic causal effects (modulo the usual invertibility caveats), and thus that the Phillips curve is linear, in a very strong sense, under any identification scheme.

Results

We use the data from the 2025 version of BenignoEggertsson2023, available on the authors' websites. Specifically, inflation $\pi_{t}$ is the quarterly, annualised core CPI inflation (excluding food and energy), constructed from monthly CPI data averaged to quarterly frequency and sourced from the BLS via FRED. The vacancy-to-unemployment ratio $\theta_{t}=v_{t}/u_{t}$ is the ratio of job vacancies to unemployed workers, using the Barnichon2010 vacancy series (as updated by the author), also averaged from a monthly to a quarterly frequency. We estimate the piecewise linear SVAR (ref) with two lags ($k=2$) over the sample periods 1960Q1-2024Q4 and 2008Q1-2024Q4, to mirror the analysis of BenignoEggertsson2023.

Testing linearity

table[table omitted — 989 chars of source]

(ref) reports likelihood ratio (LR) tests of our two linearity hypotheses: $H_{0}^{\mathrm{NS}}$ (no endogenous regime switching) and $H_{0}^{\mathrm{lin}}$ (fully linear SVAR). The results clearly reject the linearity hypothesis, both in its weak and strong forms. The apparent deterioration in fit of the linear models is even stronger in the shorter, more recent, sample.

Even though failure to reject would have been conclusive evidence against nonlinearity, these results are not enough to conclude that the Phillips curve itself, being only one equation in our bivariate system, is nonlinear. They imply that impulse responses to identified structural inflation and tightness shocks will be significantly state-dependent under any identification scheme, but it remains to be seen what this state dependence looks like in the Phillips curve that emerges from any specific identification scheme. We turn to this question next.

Phillips curve slope

figure[figure omitted — 916 chars of source]

Further evidence on the nonlinearity of the Phillips curve is obtained by computing estimates of its slope under both regimes. We do this in two different ways. First, we produce a kinked Phillips curve plot (equivalent to Figure 6(b) of BenignoEggertsson2023). This is shown in the left panel of (ref). The scatterplot shows inflation after removing the effects of all explanatory variables from the supply equation in model (ref) except $\log\theta_{t}$. The solid lines trace out the estimated Phillips curve in the $(\log\theta_{t},\pi_{t})$ space. In particular, the slope coefficient under each regime is computed as $-\Phi_{0,21}^{\left(\ell\right)}/\Phi_{0,22}$, which is given by the equation in the bottom row of (ref), solved for $z_{2,t}=\pi_{t}$, and using the fact that $\Phi_{0,22}$ is regime-independent, as per (ref). For the 2008Q1--2024Q4 sample, the estimated slopes are $\hat{\beta}^{(1)}=3.82$ ($\log\theta_{t}\leq0$ regime) and $\hat{\beta}^{(2)}=16.92$ ($\log\theta_{t}>0$ regime).

The right panel of (ref) shows a dynamic Phillips curve multiplier under each regime, computed from the state-dependent IRFs. We choose two starting points that are representative of the two regimes: 2009Q3 ($\log\theta_{t}=-1.84$, the Great Recession trough) for the $\log\theta_{t}\leq0$ regime, and 2022Q2 ($\log\theta_{t}=0.68$, the post-COVID peak) for the $\log\theta_{t}>0$ regime. The multiplier is the ratio of the cumulative inflation IRF (at horizon $h$) to the cumulative tightness IRF following a market tightness shock which raises $\log\theta_{t}$ by 1 unit over the next $h$ periods: \[ \text{Slope}_{h}^{PC}=\frac{\sum_{s=0}^{h}\frac{\partial\pi_{t+s}}{\partial\varepsilon_{\theta,t}}}{\sum_{s=0}^{h}\frac{\partial\log\theta_{t+s}}{\partial\varepsilon_{\theta,t}}}. \]

Both approaches show a substantially steeper Phillips curve in the tight labour market regime $\log\theta_{t}>0$ compared to the loose labour market regime $\log\theta_{t}\leq0$. The results are qualitatively and quantitatively consistent with BenignoEggertsson2023, which is not surprising given that we used the same identifying assumption as them.

Extensions

The appearance of a nonlinear transformation on the l.h.s.\ of the (endogenously) nonlinear SVAR

equation[equation omitted — 98 chars of source]

entails that the model automatically accommodates certain forms of regime-dependent heteroskedasticity. This can be readily seen, for example, when $f_{0}$ has the piecewise linear form \[ f_{0}(z_{t})=\sum_{\ell=1}^{L}\ensuremath{\mathbf{1}}\{z_{t}\in\set Z^{(\ell)}\}\Phi_{0}^{(\ell)}z_{t}. \] In this case, whenever the r.h.s.\ of the model is such that $z_{t}\in\set Z^{(\ell_{t})}$, the model behaves locally like a linear SVAR, with reduced form \[ z_{t}=(\Phi_{0}^{(\ell_{t})})^{-1}\b{f}_{1}(\boldsymbol{z}_{t-1})+(\Phi_{0}^{(\ell_{t})})^{-1}\varepsilon_{t}, \] for all $\varepsilon_{t}$ such that $z_{t}$ continues to lie in $\set Z^{(\ell_{t})}$. (Note that unlike in a model with exogenous regimes, $\ell_{t}$ depends on $\varepsilon_{t}$, and so the preceding does not hold for all $\varepsilon_{t}$.)

Nonetheless, in some situations it may be desirable to augment the model to allow for ARCH-type conditional heteroskedasticity, in which the variances of the structural shocks depend on certain (observed) predetermined variables. To that end, consider the following extension of (ref), to

equation[equation omitted — 179 chars of source]

where $\b z_{t-1}^{(1)}$ and $\b z_{t-1}^{(2)}$ partition (into vectors of dimension $d_{(1)}+d_{(2)}=kp$) the elements of $\b z_{t-1}$, while $\{v_{t}\}$ is strictly exogenous in the sense of being independent of $(\b z_{0},\{\varepsilon_{t}\})$, and takes values in the (possibly discrete) set $\mathcal{V}\subset\mathbb{R}^{d_{v}}$. (Rather than requiring $\{v_{t}\}$ to be stationary, we suppose that there is a measure $\mu_{v}$ on $\mathcal{V}$ to which the distribution of $v_{t}$ is equivalent, for every $t\geq0$.)

The skedastic function, $\sigma(\cdot)$, allows the volatilities of the structural shocks \[ w_{t}\coloneqq\sigma(\boldsymbol{z}_{t-1}^{(2)},v_{t-1})\varepsilon_{t} \] to depend on $(\boldsymbol{z}_{t-1}^{(2)},v_{t-1})$; we require $\sigma(\cdot)$ to be a diagonal matrix (with strictly positive entries), so that the structural shocks $w_{t}$ remain mutually uncorrelated (cf.\ Section 14.2 of KL17book). By introducing $\{v_{t}\}$, we also extend the model so as to permit the r.h.s.\ to depend on processes that are exogenous to the SVAR (such as deterministic processes). We continue to maintain that $\{\varepsilon_{t}\}$ is i.i.d.\ with mean zero and variance $I_{G}$, and moreover that $\varepsilon_{t+1}$ is independent of $(\b z_{0},\{\varepsilon_{s},v_{s}\}_{s\leq t})$, for all $t\geq0$.

Under the assumptions given below, the augmented model (ref) yields the following (time-invariant) density for $z_{t}$ conditional on $(\b z_{t-1},v_{t-1})$, \[ \varphi_{z_{t}\mid\b z_{t-1},v_{t-1}}(\xi\mid\b{\xi}_{-1},\upsilon)=\varrho\{\sigma(\b{\xi}_{-1}^{(2)},\upsilon)^{-1}[f_{0}(\xi)-\b{f}_{1}(\b{\xi}_{-1}^{(1)},\b{\xi}_{-1}^{(2)},\upsilon)]\}\cdot\smlabs{\det Df_{0}(\xi)}, \] where $\b{\xi}_{-1}\in\mathbb{R}^{kp}$ is partitioned into $(\b{\xi}_{-1}^{(1)},\b{\xi}_{-1}^{(2)})$ conformably with that of $\b z_{t-1}$ into $(\b z_{t-1}^{(1)},\b z_{t-1}^{(2)})$. Since the likelihood for $\{z_{t}\}_{t=1}^{n}$ conditional on $(\b z_{0},\{v_{t}\}_{t=0}^{n-1})$ can be expressed entirely in terms of these conditional densities, we continue to regard two alternative parametrisations as being observationally equivalent if they yield the same $\varphi_{z_{t}\mid\b z_{t-1},v_{t-1}}$ (up to the usual a.e.\ equivalences), similarly to (ref) above.

The parameters $(f_{0},\b{f}_{1},\sigma)$ of the model (ref) are, in a quite trivial sense, indistinguishable from $(\Lambdaf_{0},\Lambda\b{f}_{1},\Lambda\sigma)$, if $\Lambda$ is a diagonal matrix with strictly positive entries. Such a rescaling has no effect on the (scale-normalised) impulse responses implied by the model parameters, and is merely a consequence of the lack of a scale normalisation in (ref) -- something that was previously delivered, in the context of (ref), by the requirement that $\ensuremath{\mathbb{E}}\varepsilon_{t}\varepsilon_{t}^{\top}=I_{p}$. Letting $\set S\ni\sigma$ denote the parameter space for the skedastic function, we may fix the overall scale of the model by requiring every $\tilde{\sigma}\in\set S$ to satisfy

equation[equation omitted — 82 chars of source]

at some (user specified) value of $(\b z^{(2)\ast},v^{\ast})\in\mathbb{R}^{d_{(2)}}\times\mathcal{V}$. (To prevent this from being satisfied simply by a modification of $\tilde{\sigma}$ on a null set, we further maintain that $\tilde{\sigma}$ is continuous at $(\b z^{(2)\ast},v^{\ast})$, and that $\mu_{v}$ has strictly positive measure in every neighbourhood of $v^{\ast}$.)

Here we shall also relax the requirement that $\b f_{1}$ be continuous in all of its arguments: in fact we only require continuity of $\b z^{(1)}\ensuremath{\mapsto}\b{f}_{1}(\boldsymbol{z}^{(1)},\boldsymbol{z}^{(2)},v)$, at the cost of a strengthening of the surjectivity condition given in (ref){{{\scalefont{0.76}.}}}(ref) above. This reflects the crucial role that the variables $\boldsymbol{z}_{t-1}^{(1)}$, which are excluded from the skedastic function, now play in delivering the identification of the model parameters.

{{{{\scalefont{0.76}EXT}}}}

assumption(ref) and (ref) hold with only the following modifications, which apply for every $(\b z^{(2)},v)\in\mathbb{R}^{d_{(2)}}\times\mathcal{V}$: \begin{enumerate}[leftmargin=1.5cm] • for every $\tilde{f}_{0}\in\mathscr{F}_{0}$ and $\tilde{\b{f}}\in\pmb{\mathscr{F}}_{1}$: $\tilde{f}_{0}$ and $\b z^{(1)}\ensuremath{\mapsto}\tilde{\b{f}}_{1}(\b z^{(1)},\b z^{(2)},v)$ are locally Lipschitz; • $\b z^{(1)}\ensuremath{\mapsto}\b{f}_{1}(\b z^{(1)},\b z^{(2)},v)$ is surjective (onto $\mathbb{R}^{p}$), with $\operatorname{rk} D_{\b z^{(1)}}\b{f}(\b z^{(1)},\b z^{(2)},v)=p$ for almost every $\b z^{(1)}\in\mathbb{R}^{d_{(1)}}$. \end{enumerate} Moreover, for every $\tilde{\sigma}\in\set S$: $\tilde{\sigma}(\b z^{(2)},v)$ is a $(p\times p)$ diagonal matrix with strictly positive entries, for every $(\b z^{(2)},v)\in\mathbb{R}^{d_{(2)}}\times\mathcal{V}$; and the scale normalisation (ref) holds.

We may thus state the main result of this section, which extends (ref) above by allowing for: (i) ARCH-type heteroskedasticity; (ii) dependence of the r.h.s.\ of the model on an exogenous process $\{v_{t}\}$, and (iii) $\b{f}_{1}$ to be discontinuous in some arguments.

thmSuppose that (ref) holds. Then there exists a $\tilde{\sigma}\in\set S$ and a $\tilde{\varrho}\in\mathscr{R}$ such that $(\tilde{f}_{0},\tilde{\b{f}}_{1},\tilde{\sigma},\tilde{\varrho})$ is observationally equivalent to $(f_{0},\b{f}_{1},\sigma,\varrho)$, if and only if there exists a $Q\in\mathbb{O}(p)$ such that, for almost every $\b z^{(2)}\in\mathbb{R}^{d_{(2)}}$ and $\mu_{v}$-almost every $v\in\mathcal{V}$: \begin{align*} \tilde{f}_{0}(z) & =Qf_{0}(z),\ \forall z\in\mathbb{R}^{p}, & \tilde{\b f}_{1}(\boldsymbol{z}^{(1)},\boldsymbol{z}^{(2)},v) & =Q\b f_{1}(\boldsymbol{z}^{(1)},\boldsymbol{z}^{(2)},v),\ \forall\boldsymbol{z}^{(1)}\in\mathbb{R}^{d_{(1)}}, \end{align*} and \begin{equation} Q\sigma^{2}(\boldsymbol{z}^{(2)},v)Q^{\top} \end{equation} is a diagonal matrix; in which case $\tilde{\sigma}^{2}(\boldsymbol{z}^{(2)},v)=Q\sigma^{2}(\boldsymbol{z}^{(2)},v)Q^{\top}$.

Since the skedastic function must be a diagonal matrix, (ref) may provide further restrictions on $Q$; the extent of these will depend on the properties of the actual skedastic function $\sigma$. On the one hand, suppose that $\sigma(\boldsymbol{z}^{(2)},v)=\lambda(\boldsymbol{z}^{(2)},v)I_{p}$ is always a rescaling of the identity matrix. Then (ref) yields a diagonal matrix for every $Q\in\mathbb{O}(p)$, and no further restrictions on $Q$ are implied. On the other hand, suppose that $\sigma(\boldsymbol{z}^{(2)},v)$ varies in such a way that it is not always proportional to the identity matrix, so that the variances of some of the structural shocks may differ from each other, at least for certain values of $(\boldsymbol{z}^{(2)},v)$. In particular, if there exists a $(\boldsymbol{z}^{(2)\dagger},v^{\dagger})\in\mathbb{R}^{d_{(2)}}\times\mathcal{V}$ such that all the (diagonal) entries of $\sigma(\boldsymbol{z}^{(2)\dagger},v^{\dagger})$ are distinct, then $Q$ must be a signed permutation matrix (as follows from Theorem 2.5.4 in HJ13book; cf.\ Proposition 1 in LLM10JEDC), in which case the structural impulse response functions are identified, up to a signing and economic `labelling' of the shocks. In this way, we here obtain exactly the same kinds of restrictions that are familiar from the linear SVAR literature on `identification by heteroskedasticity' (see e.g.\ the discussion in Sections 14.2--14.3 of KL17book).

{\singlespacing

}