Uniform Validity of the Subset Anderson-Rubin Test under Heteroskedasticity and Nonlinearity Thanks: We thank participants of the Econometrics Workshop at Notre Dame 2024 and seminar participants at the Federal Reserve Bank of New York for helpful comments. The views expressed in this paper are the sole responsibility of the authors and to not necessarily reflect the views of the Federal Reserve Bank of San Francisco or the Federal Reserve System. Atsushi Inoue, Òscar Jordà, Guido M. Kuersteiner Abstract We consider the Anderson-Rubin (AR) statistic for a general set of nonlinear moment restrictions. The statistic is based on the criterion function of the continuous updating estimator (CUE) for a subset of parameters not constrained under the Null. We treat the data distribution nonparametrically with parametric moment restrictions imposed under the Null. We show that subset tests and confidence intervals based on the AR statistic are uniformly valid over a wide range of distributions that include moment restrictions with general forms of heteroskedasticity. We show that the AR based tests have correct asymptotic size when parameters are unidentified, partially identified, weakly or strongly identified. We obtain these results by constructing an upper bound that is using a novel perturbation and regularization approach applied to the first order conditions of the CUE. Our theory applies to both cross-sections and time series data and does not assume stationarity in time series settings or homogeneity in cross-sectional settings. JEL classification codes: C11, C12, C22, C32, C44. Keywords: uniformly valid inference, lack of identification, local projection, impulse response. 1 Introduction We analyze Anderson-Rubin (AR) statistics based on the continuous updating estimator (CUE) for parametric non-linear moment conditions and under a null hypothesis that restricts only a subset of the parameters. We extend results about the uniform validity of these statistics by Guggenberger, Kleibergen, Mavroeidis and Chen ( Guggenberger et al. 2012 ) , GKMC12, who considered the case of linear models in iid settings and under homoskedasticity. Our CUE framework is non-linear, allows for heteroskedasticity, general dependence and heterogeneity. In time series environments stationarity is not assumed but we do rule out trending behavior including the presence of unit roots. We make no parametric assumptions about the data generating process. The null hypothesis only imposes a set of possibly non-linear moment restrictions that are parameterized by a finite dimensional parameter. Under our assumptions, estimated parameters may be partially or completely unidentified and there may not be a valid limiting distribution for estimated parameters even in situations where the parameters are identified. The setting also includes classical weak and non-identified scenarios where the limiting distribution of parameter estimates is non-standard and/or depends on nuisance parameters that cannot be estimated. This means that Wald type inference is potentially highly unreliable. Our inference procedure using the Anderson-Rubin statistic is simple to implement and only requires conventional chi-square based critical values. Our conditions center around objects that the researcher observes and controls. These include the data used in the analysis and the moment conditions that are being specified for inference. On the other handwe treat the unobserved data generating process non-parametrically. Our theory contributes to the literature on inference when parameters are unidentified, partially or weakly identified. Phillips 1989 is an early investigation of inference in partially identified models. The weak instrument literature to which Dufour 1997 , Staiger and Stock 1997 , Stock and Wright 2000 , Kleibergen 2002 , Kleibergen 2005 and Moreira 2003 made important early contributions focuses on inference that is robust to a lack of identification. Dufour 1997 and Staiger and Stock 1997 are the first to note that the Anderson-Rubin (AR) statistic, introduced by Anderson and Rubin 1949 , can be used to construct valid confidence intervals in weakly identified settings. Kleibergen 2002 develops a modified LM statistic for linear instrumental variables models under homoskedasticity that reduces to the AR statistic in the just identified case. Stock and Wright 2000 propose a generalized version of the AR statistic based on the CUE. Stock and Wright 2000 and Dufour and Taamouti 2005 consider subvector tests in settings where only parameters that are part of the null hypothesis may be underidentified. Kleibergen 2005 proposes the K-statistic which is based on the first order conditions of CUE to account for conditional heteroskedasticity and serial correlation. Of these papers, Stock and Wright 2000 consider moment conditions and test statistics that are similar to our setting. Their theoretical analysis is limited to weak instrument asymptotic sequences which are not sufficient to establish uniform validity. Kleibergen 2005 considers subvector tests for a general set of moment conditions as we are but imposes strong identification of the parameters not subject to the null hypothesis. His asymptotic analysis includes cases where the parameters restricted under the null are unidentified or weakly identified but is pointwise rather than uniform. As GKMC12, we do not impose any restrictions on the degree of identification of either set of parameters and provide a uniform asymptotic analysis. There is a large subsequent literature on inference robust to identification failure including Andrews, Moreira and Stock (2006), Chaudhuri and Zivot 2011 , Andrews and Cheng 2012 , Andrews and Mikusheva 2016 , Andrews and Guggenberger 2019 , Andrews, Marmer and Yu ( Andrews et al. 2019 ) , Andrews, Cheng and Guggenberger ( Andrews et al. 2020 ) and Guggenberger, Kleibergen and Mavroeidis ( Guggenberger et al. 2024 ) . Our theoretical results are closest to Guggenberger, Kleibergen, Mavroeidis and Chen ( Guggenberger et al. 2012 ) , GKMC12, who show that the AR statistic applied to a subset of the parameters in linear instrumental variables models has uniformly valid size over a class of distributions that include completely unidentified scenarios. Their results are valid for iid data with homoskedastic errors. Guggenberger, Kleibergen and Mavroeidis ( Guggenberger et al. 2019 ) , GKM19, develop a more powerful subvector test under homoskedasticity but require numerically computed critical values. Andrews 2017 proposes a two step subvector test that is robust to heteroskedasticity but requires a data dependent significance level in the second step. Guggenberger, Kleibergen and Mavroeidis ( Guggenberger et al. 2024 ) , GKM24, extend GKM19 to allow for conditional heteroskedasticity by introducing a model selection procedure that selects between the AR statistic for linear models analyzed in GMK19 and the AR/AR test of Andrews 2017 . There does not seem to exist theoretical results that show the uniform validity of a simple one step subvector test robust to non-identification of the untested parameters for general heteroskedasticity, temporal dependence and non-stationarity. We show that a test based on the Anderson-Rubin testing principle formulated for the CUE is uniformly valid. The AR test is based on conventional critical values for the chi-square distribution and thus easy to implement. Our assumptions are high level and do not assume a specific model, nor identification of any of the estimated or tested parameters. Andrews and Guggenberger 2019 consider tests under general identification failure and heteroskedasticity but only consider subvector tests for the case when the parameters not subject to the null are strongly identified. Of the four types of regularity conditions that Andrews and Guggenberger (2019,p. 1709) allow to fail, we account for two failures, namely that the moment functions have multiple solutions and that the Jacobian of the moment functions are column rank deficient. On the other hand, and unlike Andrews and Guggenberger 2019 , we do maintain that the moment functions have full rank covariance matrices and that parameters under the null are in the interior of the parameter space. We also consider valid confidence intervals that are based on inverting the AR statistic. Inverting the AR statistic to construct a confidence interval dates back to Anderson and Rubin 1949 and was more recently investigated in the context of weakly identified or unidentified models by Dufour 1997 , Staiger and Stock 1997 and Stock and Wright 2000 . We illustrate the assumptions of our theory using examples of simultaneous equations models with heteroskedasticity and temporal dependence as well as a time series application with inference for impulse response analysis using the local projections proposed by Jordà 2005 . To the extent that the robust testing literature considered time series settings, stationarity has been maintained to the best of our knowledge. We allow for general forms of non-stationarity for processes with well defined second moments. While this rules out trending forms of non-stationarity such as unit roots and non-stationary fractionally integrated processes we do allow for time-changing relationships between outcomes, policy variables and controls. The paper is organized as follows. Section 2 defines the inference problem and the test statistic we consider. Section 3 discusses high level assumptions and presents the main theoretical result establishing uniform validity. The section also discusses the main steps in our proof strategy. Section 4 presents two examples that illustrate the high level assumptions we impose. Section 5 presents Monte Carlo evidence using the New Keynesian Phillips curve model considered by Kleibergen and Mavroeidis 2009 . Proofs as well as some Lemmas are contained in the Appendix. 2 Test Statistics Consider a probability space \left(\Omega,\mathcal{F},P\right) on which a vector valued double array 1 1 1 A double array does not restrict the index t for a given n, while a triangular array imposes the restriction t\leq n . of random variables \chi_{n,t}\in\mathbb{R}^{d_{\chi}} is defined and where n is an integer value representing sample size and t is the observation index. We do not assume a parametric data generating process but instead consider sequences of induced probability measures P_{\chi_{n,t}} . We are interested in testing a null hypothesis that imposes constraints on P_{\chi_{n,t}} that can be expressed in terms of moment conditions and that are parametrized by finite dimensional parameters \beta and \gamma . More specifically, the null hypothesis imposes the constraint H_{0}:\beta=\beta_{n,0} while \gamma remains unconstrained under the null. Constraints on moments are expressed using moment functions g\left(\chi_{n,t};\beta,\gamma\right) where g\left(.\right):\mathbb{R}^{d_{\chi}}\times\mathbb{R}^{d_{\beta}+d_{\gamma}}\rightarrow\mathbb{R}^{d} . For ease of notation we define g_{t}\left(\beta,\gamma\right)\equiv g\left(\chi_{n,t};\beta,\gamma\right) and suppress the dependence on \chi_{n,t} . We do assume that the moment condition n^{-1}\sum_{t=1}^{n}E\left[g_{t}\left(\beta_{n,0},\gamma_{n,0}\right)\right]=0 (1) holds under H_{0} for at least one value \gamma_{n,0} . In a strongly identified setting \beta_{n,0} and \gamma_{n,0} are the unique solutions to ( 1 ). However, we do allow for partially, weakly and unidentified cases where the values \beta_{n,0} and \gamma_{n,0} may not be the only parameter values that satisfy ( 1 ). Our framework allows for general dependence patters, including time series data. We distinguish between outcome variables y_{n,t} , exogenous or endogenous variables x_{n,t} associated with the parameters \beta in linear models, additional endogenous, exogenous or predetermined variables w_{n,t} associated with auxiliary parameters \gamma and instruments z_{n,t} that contain exogenous components of x_{nt} and w_{n,t} . When moment conditions are based on non-linear functions the association between parameters \beta , \gamma and specific covariates is often less strong and is not required for our results. In linear models where all variables in x_{n,t} and w_{n,t} are exogenous z_{n,t} may be composed only of these variables. The vector \chi_{n,t} contains all the distinct random variables contained in y_{n,t},x_{n,t},w_{n,t} and z_{n,t} . We define X_{n,t}:=\left(x_{n,t}^{\prime},w_{n,t}^{\prime}\right)^{\prime} with X_{n,t}\in\mathbb{R}^{d_{x}} . We assume that we observe a sample \chi_{n,t} for t=1,...,n. The empirical analog to the moment condition is based on sample averages \hat{g}_{n}\left(\beta,\gamma\right)=n^{-1}\sum_{t=1}^{n}g_{t}\left(\beta,\gamma\right). (2) When evaluated at \theta_{n,0}=\left(\beta_{n,0}^{\prime},\gamma_{n,0}^{\prime}\right)^{\prime} we use the shorthand notation \hat{g}_{n,0}=\hat{g}_{n}\left(\beta_{n,0},\gamma_{n,0}\right). The moment conditions ( 1 ) may be general scores of Z -estimators or may be moment functions for a GMM estimator with instruments z_{t} such that g_{t}\left(\beta,\gamma\right)=q\left(y_{n,t},X_{n,t},\beta,\gamma\right)\otimes z_{n,t} (3) and where q\left(.\right):\mathbb{R}^{d_{x}+1}\times\mathbb{R}^{d_{\beta}+d_{\gamma}}\rightarrow\mathbb{R}^{d_{q}} is a d_{q} -dimensional vector of functions. The total number of moment restrictions is d=d_{q}d_{z} . Test statistics for H_{0} are based on the CUE criterion function evaluated under H_{0} . The CUE was proposed by Sargan 1958 and analyzed by Hansen, Heaton and Yaron ( Hansen et al. 1996 ) and more recently Donald and Newey 2000 and Newey and Smith 2004 . Stock and Wright 2000 were the first to consider the CUE criterion function for hypothesis testing in a weakly identified setting involving possibly non-linear moment restrictions. The CUE criterion function depends on an estimate of the variance of \hat{g}_{n}\left(\beta,\gamma\right) . Following the literature on heteroskedasticity and autocorrelation robust standard errors (HAC), see Newey and West (1987, 1994) Newey and West 1987 , Newey and Smith 2004 , Andrews 1991 , Andrews and Monahan 1992 , and in particular de Jong and Davidson 2000 , we define the matrix \Omega_{n}\left(\beta,\gamma\right)=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}E\left[g_{t}\left(\beta,\gamma\right)g_{s}\left(\beta,\gamma\right)^{\prime}\right]. In addition, when evaluated at a sequence of pseudo-true values \beta_{n,0} and \gamma_{n,0} the covariance matrix \Omega_{n,0} is defined as \Omega_{n,0}=\Omega_{n}\left(\beta_{n,0},\gamma_{n,0}\right)=\operatorname{Var}\left(n^{-1/2}\sum_{t=1}^{n}g_{t}\left(\beta_{n,0},\gamma_{n,0}\right)\right) where the last equality holds because the moment conditions n^{-1/2}\sum_{t=1}^{n}E\left[g_{t}\left(\beta_{n,0},\gamma_{n,0}\right)\right]=0 are satisfied for \beta_{n,0} and \gamma_{n,0} . The matrix \Omega_{n,0} can be replaced by a sample analog following constructions in the HAC literature, in particular De Jong and Davidson (2000) who explicitly account for triangular arrays that are central to our theoretical discussion. To define the estimators of the GMM weight matrix and long run variances, we adopt the following definition of the class of kernel functions k\left(.\right) with bandwidth sequence a_{n} from de Jong and Davidson 2000 , see also Andrews 1991 . Definition 2.1 . Let \mathscr{\mathcal{\mathscr{K}}} be the class of kernel functions k\left(.\right):\mathbb{R}\rightarrow\left[-1,1\right] such that k\left(0\right)=0, k\left(x\right)=k\left(-x\right) for all x\in\mathbb{R}, \int_{-\infty}^{\infty}\left|k\left(x\right)\right|dx<\infty and for \psi\left(\xi\right)=\int_{-\infty}^{\infty}k\left(x\right)e^{i\xi x}dx , \psi\left(\xi\right)\geq 0. Then \hat{\Omega}_{n}\left(\beta,\gamma\right)=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}k\left(\frac{t-s}{a_{n}}\right)g_{t}\left(\beta,\gamma\right)g_{s}\left(\beta,\gamma\right)^{\prime}. (4) The continuous updating estimator for \theta=\left(\beta^{\prime},\gamma^{\prime}\right)^{\prime} is now given by \hat{\theta}_{CUE}=\left(\hat{\beta}^{\prime}_{CUE},\hat{\gamma}^{\prime}_{CUE}\right)^{\prime}=\textrm{argmin}_{\beta,\gamma}n^{-1}Q_{n}\left(\beta,\gamma\right) where Q_{n}\left(\beta,\gamma\right)=n\hat{g}_{n}\left(\beta,\gamma\right)^{\prime}\hat{\Omega}_{n}^{-1}\left(\beta,\gamma\right)\hat{g}_{n}\left(\beta,\gamma\right). The subset AR statistic for the null hypothesis H_{0} based on the CUE is given as \textrm{AR}_{\textrm{C}}\left(\beta_{n,0}\right)=\textrm{min}_{\gamma}Q_{n}\left(\beta_{n,0},\gamma\right). (5) Let d_{\gamma} be the dimension of \gamma and assume that d-d_{\gamma}>0. Under regularity conditions, and when \gamma is strongly identified it is well known from Stock and Wright 2000 , Newey and Smith 2004 and Kleibergen 2005 that \textrm{AR}\left(\beta_{n,0}\right) has a limiting \chi_{d-d_{\gamma}}^{2} distribution. When the AR statistic is formulated based on the LIML estimator for linear models and iid observations with homoskedastic errors GKMC12 show that ( 5 ) is bounded by a random variable with a limiting \chi_{d-d_{\gamma}}^{2} distribution uniformly over sequences of data distributions where \beta and \gamma may not be identified. A nominal size \alpha test of H_{0} then is based on rejecting the null when \textrm{AR}_{\textrm{C}}\left(\beta_{n,0}\right)>c_{1-\alpha,\chi_{d-d_{\gamma}}^{2}} where c_{1-\alpha,\chi_{d-d_{\gamma}}^{2}} is the 1-\alpha quantile of the \chi_{d-d_{\gamma}}^{2} distribution. Based on Anderson and Rubin (1949) we can construct a confidence region with coverage rate of at least 1-\alpha by inverting the AR statistic \textrm{CI}=\left\{\beta\in\mathbb{R}^{d_{\beta}}|\mathit{\textrm{AR}_{\textrm{C}}}\left(\beta\right)\leq c_{1-\alpha,\chi_{d-d_{\gamma}}^{2}}\right\}. The confidence set CI is the set of all values \beta for which the AR statistic does not reject H_{0} . AR based confidence sets were analyzed in the context of weakly identified or non-identified linear IV methods by Dufour 1997 , and for weakly identified instrumental variables by Staiger and Stock 1997 and Stock and Wright 2000 . GKMC12 are the first to show uniform validity of the subset AR test over all sequences of linear models, including weakly identified, partially identified and unidentified cases. A key ingredient to uniform validity of the AR statistic in more general settings than linear models with homoskedastic errors is that the statistic is based on the CUE. The CUE implicitly orthogonalizes the moment functions \hat{g}_{n}\left(\theta\right) from the influence of estimating nuisance parameters \gamma . Orthogonalization lies at the core of the C\left(\alpha\right) test of Neyman 1959 and was shown in Donald and Newey 2000 to be a feature of the first order conditions of the CUE. Let \hat{\gamma}_{n}=\hat{\gamma}_{n}\left(\beta_{n,0}\right)=\operatorname{argmin}_{\gamma}Q_{n}\left(\beta_{n,0},\gamma\right). If \hat{\gamma}_{n} is not unique then any value for \gamma that minimizes Q_{n}\left(\beta_{n,0},\gamma\right) can be chosen. We show that the only effect of estimating \gamma by \hat{\gamma}_{n} on the asymptotic distribution \textrm{AR}_{\textrm{C}}\left(\beta_{n,0}\right)=Q_{n}\left(\beta_{n,0},\hat{\gamma}_{n}\right) is a reduction in the degrees of freedom of the resulting \chi^{2} distribution. This result does not hold when first stage estimators for \gamma such as two stage least squares or GMM are plugged into Q_{n}\left(\beta_{n,0},\gamma\right). 3 Theoretical Results We start by stating detailed regularity conditions we impose. Let \left(\Omega,\mathcal{F\mathit{,P}}\right) be a probability space. Let \chi_{n}=\left\{\chi_{n,t}:t=1,2,...;n=1,2,...\right\} be a double array of random vectors of dimension d_{\chi} . Let y_{n,t} and x_{n,t} be double arrays of random variables with elements from \chi_{n,t}. Let w_{n,t} be a d_{w}\times 1 vector and z_{n,t} a d_{z}\times 1 vector of elements from \chi_{n,t} . Assume that x_{n,t} and w_{n,t} contain different elements, while z_{n,t} may contain x_{n,t} and w_{n,t} as well as additional variables. The exact configuration of z_{n,t} depends on whether x_{n,t} and w_{n,t} are believed to be endogenous or exogenous. Let \lambda_{min}\left(A\right) denote the smallest eigenvalue of the matrix A , let \left\|A\right\|_{2}=\operatorname{tr}\left(A^{\prime}A\right)^{1/2} be the Frobenius norm and \left\|A\right\|=\sup_{\left\|x\right\|=1}\left\|Ax\right\| be the operator norm where \left\|.\right\| applied to a vector in \mathbb{R}^{k} is the usual Euclidean norm. Let \chi_{n}:\Omega\rightarrow\mathbb{R}^{\infty} and denote by \chi_{n} \rightsquigarrow \chi weak convergence of \chi_{n} to a random element \chi in \mathbb{R}^{\infty} , van der Vaart and Wellner 1996 (1996, Definition 1.3.3). Let \mathscr{X} be a collection of measurable processes \chi . Each \chi\in\mathscr{X} induces a probability distribution P_{\chi} . We consider sequences of \chi_{n}=\left(\chi_{n,1},\chi_{n,2},...,\chi_{n,n},...\right) of processes \chi_{n}\in\mathscr{X} that contain subsequences that converge to a limiting random element \chi in \mathbb{R}^{\infty} . This is formalized in Assumption 1 . Assumption 1 . Every sequence \chi_{n}\in\mathscr{X} for n\geq 1 is measurable and asymptotically tight. 2 2 2 A sequence \chi_{n} is asymptotically tight if for every \varepsilon>0 there exists a compact K with \delta -enlargement K^{\delta}=\left\{\eta\in\mathbb{R}^{\infty}|d\left(\eta,K\right)<\delta\right\} for a metric d\left(.,.\right) on \mathbb{R}^{\infty} such that \liminf_{n}P\left(\chi_{n}\in K^{\delta}\right)\geq 1-\varepsilon\text{ } for every \delta>0 , see van der Vaart and Wellner 1996 (1996, p.21). Note that inner probability is replaced by P because we only consider measurable sequences. . Assumption 1 and Prohorov’s theorem (see van der Vaart and Wellner 1996 , 1996, Theorem 1.3.9) imply that for every subsequence n_{k} there is a further subsequence of n_{k} denoted by n_{k_{l}} such that \chi_{n_{k_{l}}} \rightsquigarrow \chi and where the limit \chi may depend on the particular subsequence. In our notation we suppress the dependence of limits \chi on the subsequence. For ease of notation we omit references to specific subsequences and refer to converging subsubsequences n_{k_{l}} simply as converging subsequences n_{k} . Assumption 1 implies that all finite dimensional distributions of \chi_{n_{k}} converge along converging subsequences. To the extent that moments of \chi_{n} exist, moments related to finite dimensional distributions also converge along converging subsequences. Assumption 1 neither implies that ordinary limits (as opposed to subsequential limits) exist, nor does it rule out the existence of such ordinary limits. See Andrews et al. 2020 for a related assumption that makes the dependence of the limit \chi on the subsequence explicit. Assumption 1 is a nonparametric generalization of semiparametric assumptions in GKMC12 and Andrews et al. 2020 and is similar to assumptions imposed in Andrews and Guggenberger 2019 . These papers define classes of data generating processes and sequences taking values in these model classes that converge to limiting data generating processes as the sample size increases. Special cases of such classes can be found in the weak instrument literature that follows the approach of Staiger and Stock 1997 and where only particular sequences leading to so called weak instrument limits are considered. In order to analyze uniformity, the setting in GKMC12 is significantly more general as it accounts for completely unidentified, partially identified as well as fully identified data generating processes. We do not assume that \chi_{n} is generated by a particular parametric or semiparametric model. We only assume that under the null hypothesis a set of finite dimensional moment restrictions parameterized by a set of finite dimensional parameters hold. The null hypothesis is specified in terms of restrictions on a subset of the parameter vector. We do not specifically consider the case where the null hypothesis restricts the entire parameter vector since in that scenario the AR statistic trivially has uniformly correct size under our assumptions and results. The next assumption postulates that the moment conditions ( 1 ) hold along parameter sequences under the null hypothesis. Assumption 2 . Let \theta=\left(\beta^{\prime},\gamma^{\prime}\right)^{\prime} and let \Theta\subseteq\mathbb{R}^{d_{\gamma}+d_{\beta}} be the parameter space. Assume that \Theta is compact. Let \textrm{int}\Theta be the interior of \Theta. There exists a random function g_{t}\left(.\right):\mathbb{R}^{d_{\beta}+d_{\gamma}}\rightarrow\mathbb{R}^{d} , two times continuously differentiable with respect to \gamma and for each n\geq 1 a vector \theta_{n,0}=\left(\beta_{n,0}^{\prime},\gamma_{n,0}^{\prime}\right)^{\prime}\in\textrm{int}\Theta with H_{0}:\beta=\beta_{n,0} such that n^{-1/2}\sum_{t=1}^{n}E\left[g_{t}\left(\beta_{n,0},\gamma_{n,0}\right)\right]=0. (6) Along converging subsequences n_{k} assume that \lim_{k\rightarrow\infty}\theta_{n_{k},0}=\theta_{0} for some \theta_{0} \in\textrm{int}\Theta and where \theta_{0} may depend on the subsequence. In applications the null hypothesis is usually formulated for a fixed constant \beta_{0} such that H_{0}:\beta=\beta_{0} . Our theory covers a fixed null as a special case. Without assuming homogeneity or stationarity \gamma_{n,0} typically varies with n even if \beta_{n,0}=\beta_{0} is fixed. Note that \theta_{n,0} is not necessarily unique under our assumptions and in general there are infinitely many solutions to the moment condition. Our theory is not intended to directly test for violations of Assumption 2 which imposes both functional form restrictions summarized by g_{t}\left(.\right) as well as a restriction on the parameter \beta_{0}. In other words, and in line with existing theory, a test of H_{0} is interpreted as testing \beta=\beta_{0} while taking for granted that the moment condition in ( 6 ) is satisfied for some \theta. However, since our theoretical analysis is only conducted under the null we do not take a stand on whether ( 6 ) holds for some \theta under the alternative. Assumption 3 . Let \hat{\Omega}_{n}\left(\theta\right) be defined in ( 4 ), \Omega_{n}\left(\theta\right)=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}E\left[g_{t}\left(\beta,\gamma\right)g_{s}\left(\beta,\gamma\right)^{\prime}\right] . For all converging subsequences n_{k} , a sequence of bandwidth parameters a_{n} and a sequence \hat{\theta}_{n} such that \hat{\theta}_{n_{k}}\rightarrow_{p}\theta_{0} there exists a symmetric matrix \Omega\left(\theta\right) of continuous functions of \theta with \inf_{\theta\in\Theta}\lambda_{\min}\left(\Omega\left(\theta\right)\right)\geq K_{\Omega}>0 such that i) \Omega_{n_{k}}\left(\theta_{n_{k},0}\right)-\Omega\left(\theta_{0}\right)\rightarrow 0 and \hat{\Omega}_{n_{k}}\left(\theta_{n_{k},0}\right)-\Omega_{n_{k}}\left(\theta_{0}\right)\rightarrow_{p}0 ii) \Omega_{n_{k}}\left(\hat{\theta}_{n_{k}}\right)-\Omega\left(\theta_{0}\right)\rightarrow_{p}0 and \hat{\Omega}_{n_{k}}\left(\hat{\theta}_{n_{k}}\right)-\Omega_{n_{k}}\left(\theta_{0}\right)\rightarrow_{p}0 . . The full rank assumption for \Omega\left(\theta\right) corresponds to similar assumptions in Stock and Wright 2000 and Kleibergen 2005 . The assumption has been relaxed by Andrews and Guggenberger 2019 for a modified version of the AR statistic. The convergence of \hat{\Omega}_{n_{k}}\left(\theta\right)\rightarrow_{p}\Omega\left(\theta\right) was shown for triangular near epoch dependent arrays defined on mixing sequences by de Jong and Davidson 2000 . The limits in Assumption 3 may depend on the specific subsequence. Additional regularity conditions may guarantee that convergence obtains for n_{k}=n , a scenario that is a special case under our assumptions. When g_{t}\left(.\right) is of the form ( 3 ) and the data are stationary a necessary condition for Assumption 3 is that for all a\in\mathbb{R}^{d_{z}} , \left\|a\right\|=1 and all n\geq 1 the variables z_{n,t} are not colinear in the sense that \sup_{t}\Pr\left(z_{n,t}^{\prime}a=0\right)<1. (7) To see this, assume that this condition does not hold. Then, for some a\in\mathbb{R}^{d_{z}} it follows that z_{n,t}^{\prime}a=0\text{ a.s.} for all n and t. But this implies that g_{t}\left(\beta,\gamma\right)\left(I_{q}\otimes a\right)=0\text{ a.s.} for all n and t which in turn implies that Assumption 3 cannot hold. In addition, ( 7 ) implies that \lambda_{min}\left(E\left[z_{n,t}z_{n,t}^{\prime}\right]\right)>0 and \lambda_{min}\left(n^{-1}\sum_{t=1}^{n}E\left[z_{n,t}z_{n,t}^{\prime}\right]\right)>0. Using the pseudo true parameter value \theta_{n,0} we now define the pseudo residual \varepsilon_{n,t}=g_{t}\left(\beta_{n,0},\gamma_{n,0}\right) (8) where \varepsilon_{n,t} is a d\times 1 vector of random variables and where by Assumption 2 it follows n^{-1/2}\sum_{t=1}^{n}E\left[\varepsilon_{n,t}\right]=0 . Also define d\times d_{\gamma} dimensional matrix of random functions g_{t}^{\gamma}\left(\beta,\gamma\right)=\partial g_{t}\left(\beta,\gamma\right)/\partial\gamma^{\prime} and let \hat{g}_{n,\gamma}\left(\gamma\right)=n^{-1}\sum_{t=1}^{n}g_{t}^{\gamma}\left(\beta_{n,0},\gamma\right) (9) and \hat{g}_{n,\gamma}=\hat{g}_{n,\gamma}\left(\gamma_{n,0}\right). (10) The behavior of the derivative g_{t}^{\gamma}\left(\beta,\gamma\right) determines whether \gamma_{n,0} is identified. Stock and Wright 2000 consider moment functions that additively separate into a component that only depends on \gamma and assume that the derivative of the moment function with respect to that component is full rank. Similarly, Kleibergen 2005 imposes a full column rank assumption on a term equivalent to \hat{g}_{n,\gamma} for the subset version of his test. Instead, we allow for the derivatives \hat{g}_{n,\gamma} to be of reduced column rank in the limit. We Impose the following assumptions on \hat{g}_{n,\gamma}. Assumption 4 . For all converging subsequences n_{k} assume that there exists a sequence of non-stochastic matrices g_{n,\gamma}=E\left[\hat{g}_{n,\gamma}\right] and some matrix g_{\gamma} that depends on n_{k} such that g_{n_{k},\gamma}-g_{\gamma}\rightarrow 0 and \hat{g}_{n_{k},\gamma}-g_{n_{k},\gamma}\rightarrow_{p}0 . In addition, for any \check{\gamma}_{n} such that \check{\gamma}_{n_{k}}-\gamma_{n_{k},0}=O_{p}\left(n_{k}^{-1/2}\right) along converging subsequences n_{k} it holds that \hat{g}_{n_{k},\gamma}\left(\check{\gamma}_{n_{k}}\right)-g_{n_{k},\gamma}\rightarrow_{p}0. Define the d_{\gamma}\times d_{\gamma} non-stochastic matrix \Gamma=g_{\gamma}^{\prime}\Omega_{0}^{-1}g_{\gamma} with eigenvalues \Delta_{1}\geq\Delta_{2}...\geq\Delta_{d_{\gamma}}\geq 0 . Assumption 4 does not restrict the column rank of g_{\gamma} . Thus, we allow the rank r of \Gamma to take values 0\leq r\leq d_{\gamma} and where in general \Delta_{j} and r depend on the subsequence n_{k}. Define \Pi{}_{n,\gamma}=n^{-1}\sum_{t=1}^{n}E\left[g_{t}^{\gamma}\left(\beta_{n,0},\gamma_{n,0}\right)^{\prime}\right] where \Pi_{n,\gamma} is a d\times d_{\gamma} dimensional matrix. Define V_{n,t} as a d\times d_{\gamma} dimensional matrix of pseudo-residuals that can be written as V_{n,t}=g_{t}^{\gamma}\left(\beta_{n,0},\gamma_{n,0}\right)-\Pi_{n,\gamma} (11) such that \sum_{t=1}^{n}E\left[V_{n,t}\right]=0 by construction. As in GKMC12 no other restrictions are imposed on \Pi_{n,\gamma} so that scenarios where \Pi_{n,\gamma} is rank deficient, or even \Pi_{n,\gamma}=0 are allowed. The properties of \Pi_{n,\gamma} depend on the data distribution P_{\chi} . Rank deficiency of \Pi_{n,\gamma} arises when \gamma is not identified, partially or weakly identified. The analysis of the CUE also involves derivatives of \hat{\Omega}_{n_{k}}\left(\theta\right) given by \frac{\partial Q_{n}\left(\theta\right)}{\partial\gamma^{\prime}}=2n\hat{g}_{n}\left(\theta\right)^{\prime}\hat{\Omega}_{n,0}^{-1}\left(\hat{g}_{n,\gamma}\left(\gamma\right)-\left(I_{d}\otimes\hat{g}_{n}\left(\theta\right)^{\prime}\hat{\Omega}_{n,0}^{-1}\right)\hat{\Lambda}_{n}\left(\gamma\right)\right) (12) where \hat{\Lambda}_{n}\left(\gamma\right)=\partial\operatorname{vec}\hat{\Omega}_{n}\left(\beta_{n,0},\gamma\right)/\partial\gamma^{\prime}. (13) To guarantee convergence of \hat{\Lambda}_{n}\left(\gamma\right) define the matrices \hat{\Lambda}_{z,n,0}=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}k\left(\frac{t-s}{a_{n}}\right)\left(I_{d}\otimes\varepsilon_{n,t}\right) (14) and \hat{\Lambda}_{V,n,0}=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}k\left(\frac{t-s}{a_{n}}\right)\left[V_{n,t}\otimes\varepsilon_{n,t}\right] (15) and impose the following restrictions on \hat{\Lambda}_{z,n,0} and \hat{\Lambda}_{V,n,0} . Assumption 5 . Let \hat{\Lambda}_{z,n,0} and \hat{\Lambda}_{V,n,0} be defined in ( 14 ) and ( 15 ), and let \Lambda_{z,n,0}=0 , and \Lambda_{V,n,0}=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}E\left[V_{n,t}\otimes\varepsilon_{n,t}\right]. For all converging subsequences n_{k} as well as a sequence of bandwidth parameters a_{n} it follows that \hat{\Lambda}_{z,n_{k},0}-\Lambda_{z,n_{k},0}\rightarrow_{p}0 and \hat{\Lambda}_{V,n_{k},0}-\Lambda_{V,n_{k},0}\rightarrow_{p}0 . Let \Lambda_{V,0} be matrices with bounded elements that may depend on the subsquence n_{k}. Assume that \Lambda_{V,n_{k},0}\rightarrow\Lambda_{V,0} . . We now consider the process \omega_{n,t}=\left(\varepsilon_{n,t}^{\prime},\operatorname{vec}\left(V_{n,t}\right)^{\prime}\right)^{\prime} where \omega_{n,t} is a d\left(1+d_{\gamma}\right) vector of double arrays of random variables. Let S_{n}=\frac{1}{\sqrt{n}}\sum_{t=1}^{n}\omega_{n,t}. Note that E\left[S_{n}\right]=0 by ( 8 ), ( 11 ) and Assumption 2 , such that, without loss of generality, \omega_{n,t} can be assumed to be mean zero. The next assumption imposes that \nu_{n,t} has a non-degenerate limiting distribution along all converging subsequences. Assumption 6 . For all n\geq 1 : E\left[\text{$\left\|S_{n}\right\|_{2}^{2}$}\right]<\infty , and for \Sigma_{n}=E\left[S_{n}S_{n}^{\prime}\right] , there is some b_{\Sigma} such that \lambda_{min}\left(\Sigma_{n}\right)>b_{\Sigma}>0 for all n\geq 1 . In addition, for all converging subsequences n_{k} it holds that \Sigma_{n_{k}}^{-1/2}S_{n_{k}}\rightarrow_{d}N\left(0,I\right) and \Sigma_{n_{k}}\rightarrow\Sigma where \Sigma is a positive definite symmetric matrix with non-random elements that may depend on the subsequence. Assumptions ( 3 ), ( 4 ), ( 5 ) and ( 6 ) are high level assumptions. They could be formulated more directly by imposing restrictions on the moments and the distribution of \chi_{n} . Such assumptions then could be used to invoke laws of large numbers and central limit theorems implying Assumptions ( 3 )-( 6 ). We do not follow this approach here because low level assumptions tend to be more specific to a particular testing context and obscure the fundamental requirements for our results. However, we consider two specific examples in Section 4 were we discuss a set of low level assumptions that can be imposed to guarantee that Assumptions ( 3 ), ( 4 ), ( 5 ) and ( 6 ) hold. Our main theoretical result is now stated. Recall the definition of the AR statistic \textrm{AR}_{\textrm{C}}\left(\beta_{n,0}\right) in ( 5 ). The asymptotic size of the subset AR test is defined as \textrm{AsySz}_{\alpha}=\limsup_{n\rightarrow\infty}\sup_{\chi_{n}\in\mathscr{X}}P_{\chi_{n}}\left(\text{$\textrm{AR}_{\textrm{C}}\left(\beta_{n,0}\right)$}>c_{1-\alpha,\chi_{\left(d-d_{\gamma}\right)}^{2}}\right) (16) where P_{\chi_{n}} is the induced probability measure of the process \chi_{n} on the underlying probability space and where each \chi_{n} implies a corresponding sequences of parameters \theta_{n,0} under H_{0} . We prove the following result which extends Theorem 1 of GKMC12 to the case of the AR statistic based on the CUE rather than LIML and that allows for nonlinear moment conditions and non-stationary or heterogenous processes with conditional heteroskedasticity. Theorem 3.1 . Assume 0<\alpha<1 and that Assumptions 1 - 6 hold for all \chi_{n}\in\mathscr{X} and all n\geq 1 . Then the asymptotic size of the \textrm{AR}_{\textrm{C}} statistic defined in ( 16 ) satisfies \text{$\textrm{AsySz}_{\alpha}$}=\alpha. The challenge in proving Theorem 3.1 under the maintained assumptions of this paper is that the CUE \hat{\gamma}_{n} of \gamma is not necessarily unique and may not converge to a well defined limit. Then, conventional approximation arguments for the first order conditions of the CUE fail. The proof of Theorem 3.1 is based on a novel approach of regularizing the first order condition. We then perturb the pseudo true value \gamma_{n,0} toward a value \tilde{\gamma}_{n,0} in the direction of the regularized first order conditions. This is done in a way that keeps \tilde{\gamma}_{n,0} close enough to \gamma_{n,0} for the central limit theorem of Assumption ( 6 ) to apply while at the same time approximately satisfying the first order conditions. An immediate consequence of the definition of \textrm{AR}_{\textrm{C}}\left(\beta_{n,0}\right) is that all sequences \gamma_{n} produce upper bounds because \textrm{argmin}_{\gamma}Q\left(\beta_{n,0},\gamma\right)\leq Q\left(\beta_{n,0},\gamma_{n}\right) for any sequence \gamma_{n} . While \gamma_{n,0} leads to one such upper bound, the bound based on \gamma_{n,0} is too large in the sense that Q\left(\beta_{n,0},\gamma_{n,0}\right) has a limiting \chi_{d}^{2} rather than a \chi_{d-d_{\gamma}}^{2} distribution. This occurs because \gamma_{n,0} does not satisfy the first order conditions well enough even in an asymptotic sense. The perturbed point \tilde{\gamma}_{n,0} is constructed in such a way that it approximately solves the CUE moment conditions evaluated at the infeasible sequence \theta_{n,0} under H_{0}:\beta=\beta_{n,0} and that as a result Q\left(\beta_{n,0},\tilde{\gamma}_{n,0}\right) has the desired \chi_{d-d_{\gamma}}^{2} distribution in the limit irrespective of whether \gamma is identified or not. We stress that \tilde{\gamma}_{n,0} is a construct exclusively used in the proofs and not needed to compute the test statistic \textrm{AR}_{\textrm{C}}\left(\beta_{n,0}\right) itself. Donald and Newey 2000 show that the CUE moment conditions remove the highest order bias term of a conventional GMM estimator by showing that the CUE moment conditions are centered at zero. Their insight implies an asymptotic orthogonality condition between the limiting process of the moment function \hat{g}_{n} and the residual of the projection of \hat{g}_{n,\gamma} onto \hat{g}_{n,0} . The same insight underlies Kleibergen 2005 but only for the case where \gamma is strongly identified. To show uniform validity, the asymptotic orthogonality needs to be established for all converging subsequences irrespective of whether \gamma is identified or not. This then allows to characterize the upper bound of the \textrm{AR}_{\textrm{C}}\left(\beta_{n,0}\right) statistic in terms of a \chi_{d-d_{\gamma}}^{2} limiting distribution along all converging subsequences. The limiting \chi_{d-d_{\gamma}}^{2} distribution is obtained by representing the AR-statistic asymptotically as the sum of squares of the residuals of a projection of the moment conditions \hat{g}_{n} onto the column space spanned by the residualized \hat{g}_{n,\gamma} . The rank of this projection residual is d-d_{\gamma} irrespective of whether \gamma is identified or not. We now explain the construction of the sequence \tilde{\gamma}_{n,0} and the proof strategy behind Theorem 3.1 in more detail. Use ( 2 ) to define \hat{g}_{n}\left(\gamma\right)=\hat{g}_{n}\left(\beta_{n,0},\gamma\right) , \hat{g}_{n,0}=\hat{g}_{n}\left(\gamma_{n,0}\right) and \tilde{g}_{n,0}=\hat{g}_{n}\left(\tilde{\gamma}_{n,0}\right) . Recall the definition of \hat{g}_{n,\gamma} in ( 10 ) and use ( 9 ) to define \tilde{g}_{n,\gamma}=\hat{g}_{n,\gamma}\left(\tilde{\gamma}_{n,0}\right) . Also recall the definition of \hat{\Lambda}_{n}\left(\gamma\right) in ( 13 ) where \hat{\Lambda}_{n}\left(\gamma\right) has dimension d^{2}\times d_{\gamma} . Now construct the matrix \hat{A}_{n}=\hat{\Omega}_{n,0}^{-1/2}\left(\hat{g}_{n,\gamma}\left(\gamma_{n,0}\right)-\left(I_{d}\otimes\hat{g}_{n,0}^{\prime}\hat{\Omega}_{n,0}^{-1}\right)\hat{\Lambda}_{n}\left(\gamma_{n,0}\right)\right) (17) and consider the empirical moment conditions 0=\hat{A}_{n}^{\prime}\hat{\Omega}_{n,0}^{-1/2}\hat{g}_{n}\left(\gamma\right). (18) To gain some intuition for the moment condition ( 18 ) we focus on the case where g_{t}\left(.\right) is of the form ( 3 ) with d_{q}=1 such that ( 18 ) can be written as 0=\left(\hat{g}_{n,\gamma}-\hat{g}_{n,0}^{\prime}\hat{\Omega}_{n,0}^{-1}\hat{\Lambda}_{n,0}\right)^{\prime}\hat{\Omega}_{n,0}^{-1}\hat{g}_{n}\left(\gamma\right). As argued in Donald and Newey 2000 , \hat{g}_{n,0}^{\prime}\hat{\Omega}_{n,0}^{-1}\hat{\Lambda}_{n,0} is the projection of \hat{g}_{n,\gamma} onto \hat{g}_{n,0} in the case of independently distributed data. The term \hat{g}_{n,\gamma}-\hat{g}_{n,0}^{\prime}\hat{\Omega}_{n,0}^{-1}\hat{\Lambda}_{n,0} then is the projection residual, which by construction is orthogonal to \hat{g}_{n,0} . As shown by Donald and Newey 2000 , the orthogonality holds exactly in the iid setting and in expectation in a time series framework. The use of long run variance-covariance matrices in the time series case ensures that this interpretation remains valid in the limit in the more general setting of this paper. By constructing our perturbed moment vector \tilde{g}_{n,0} to be close to \hat{g}_{n,0} and approximately orthogonal to \hat{A}_{n} we guarantee that the moment vector \hat{g}_{n,0} is stochastically independent of the column space spanned by \hat{A}_{n} at least in the limit. The construction of \tilde{\gamma}_{n,0} depends on the particular subsequence n_{k} such that \tilde{\gamma}_{n,0} is only defined for n_{k} and may differ for different sequences n_{k} and n_{k^{\prime}} . We start with the mean value expansion of \hat{g}_{n}\left(\gamma\right) around \gamma_{n,0} , \hat{g}_{n}\left(\gamma\right)=\hat{g}_{n,0}+\hat{g}_{n,\gamma}\left(\check{\gamma}_{n,0}\right)\left(\gamma-\gamma_{n,0}\right) (19) with \left\|\check{\gamma}_{n,0}-\gamma_{n,0}\right\|\leq\left\|\gamma-\gamma_{n,0}\right\| which upon substitution into the moment conditions ( 18 ) leads to 0=\hat{A}_{n}^{\prime}\hat{\Omega}_{n,0}^{-1/2}\left(\hat{g}_{n,0}-\hat{g}_{n,\gamma}\left(\check{\gamma}_{n,0}\right)\left(\gamma_{n,0}-\gamma\right)\right). (20) If \hat{A}_{n} and \hat{g}_{n,\gamma} were full column rank at least for large enough samples, then one could solve ( 20 ) for \gamma. Under the sequences considered in this paper, both matrices may have reduced ranks for finite n and in the limit. Solutions based on the Moore-Penrose inverse have delicate convergence properties, see for example Wedin 1973 . In fact, such solutions often do not converge because the operator norm of the MP inverse becomes unbounded as eigenvalues of its argument tend towards zero, see Stewart 1977 (1977, Theorem 3.1). To stabilize these solutions we adopt a regularization scheme called truncated singular value decomposition (TSVD), see the numerical analysis literature Hanson 1971 (1971), Varah 1973 (1973) and Hansen 1987 . Let the singular value decomposition (SVD), which in this case coincides with the spectral representation of \hat{\Gamma}_{n}=\hat{g}_{n,\gamma}^{\prime}\hat{\Omega}_{n,0}^{-1}\hat{g}_{n,\gamma} be equal to \hat{\Gamma}_{n}=\hat{R}_{n}\hat{\Delta}_{n}\hat{R}^{\prime}_{n} and where \hat{\Delta}_{n} is a diagonal matrix of the eigenvalues \hat{\Delta}_{1,n}\geq....\geq\hat{\Delta}_{d_{\gamma},n}\geq 0 of \hat{\Gamma}_{n} . By Assumptions 3 and 4 and for a converging subsequence n_{k} , \hat{\Gamma}_{n_{k}}\rightarrow_{p}\Gamma such that by Theorem A.1 , \hat{\Delta}_{j,n_{k}}\rightarrow_{p}\Delta_{j}\geq 0 . Assume that \Delta_{j}>0 for j\leq r and some 0\leq r\leq d_{\gamma} . Fix \varepsilon arbitrary with \Delta_{r}>\varepsilon>0 . For each converging subsequence n_{k} define \mathring{\Delta}_{n_{k}} as the matrix with diagonal elements \mathring{\Delta}_{n_{k},j} such that for all l\in\left\{1,...,d_{\gamma}\right\}, \text{$\mathring{\Delta}_{n_{k},l}$}=\begin{cases}\hat{\Delta}_{n_{k},l}&\text{\text{if $\hat{\Delta}_{n_{k},l}>\varepsilon$}}\\ 0&\text{otherwise}\end{cases}. The TSVD of \hat{\Gamma}_{n_{k}} is now given by \mathring{\Gamma}_{n_{k}}=\hat{R}_{n_{k}}\mathring{\Delta}_{n_{k}}\hat{R}^{\prime}_{n_{k}}. (21) We construct an infeasible solution \tilde{\gamma}_{n_{k},0} to ( 20 ) with two properties. The solution \tilde{\gamma}_{n,0} is a perturbation of \gamma_{n,0} small enough to converge to \gamma_{n,0} but such that it also satisfies ( 20 ) with sufficient accuracy in large samples. The purpose of the construction is to evaluate the CUE criterion function at a point that allows for a limiting distribution and where the moment conditions of the CUE, and thus orthogonality between \hat{A}_{n_{k}} and \tilde{g}_{n_{k},0} hold approximately. The parameter \tilde{\gamma}_{n_{k},0} is defined as a perturbation of the parameters \gamma_{n_{k},0} in the direction of the first order conditions for the CUE estimator by setting \tilde{\gamma}_{n_{k},0}=\gamma_{n_{k},0}-\mathring{\Gamma}_{n_{k}}^{+}\hat{A}_{n_{k}}^{\prime}\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}. (22) Note that the MP-inverse of \mathring{\Gamma}_{n_{k}} denoted by \mathring{\Gamma}_{n_{k}}^{+} is continuous and therefore converges along converging subsequences. Continuity of \mathring{\Gamma}_{n_{k}}^{+} holds because the eigenvalues of \mathring{\Gamma}_{n_{k}} are bounded away from zero by \varepsilon due to regularization. This is in contrast to the MP-inverse of \hat{\Gamma}_{n_{k}} which may not be continuous and thus not converge along converging subsequences. The following properties of \tilde{\gamma}_{n_{k},0} can now be established. Lemma 3.1 . Let \tilde{\gamma}_{n,0} be defined in ( 22 ). Let Assumptions 1 , 2 , 3 , and 4 hold. Then, (i) for all converging subsequences n_{k} \tilde{\gamma}_{n_{k},0}-\gamma_{n_{k},0}=O_{p}\left(n_{k}^{-1/2}\right). (ii) for all converging subsequences n_{k} it follows that the first order condition in ( 18 ) evaluated at \tilde{\gamma}_{n,0} satisfies \displaystyle\hat{A}_{n_{k}}^{\prime}\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k}}\left(\tilde{\gamma}_{n_{k},0}\right) \displaystyle= \displaystyle o_{p}\left(n_{k}^{-1/2}\right). (iii) for the projection P_{\hat{A}_{n}} onto the column space of \hat{A}_{n} it follows that along converging subsequences n_{k} P_{\hat{A}_{n_{k}}}-\hat{A}_{n_{k}}\mathring{\Gamma}_{n_{k}}^{+}\hat{A}_{n_{k}}^{\prime}=o_{p}\left(1\right). A proof of Lemma 3.1 is given in Appendix A.3 . Let \tilde{\Omega}_{n,0}=\hat{\Omega}_{n}\left(\beta_{n,0},\tilde{\gamma}_{n,0}\right) . Using Lemma 3.1 and denoting by P_{\hat{A}_{n}} the projection onto the column space spanned by \hat{A}_{n} and letting M_{\hat{A}_{n}}=I-P_{\hat{A}_{n}} , the proof of Theorem 3.1 then proceeds to show in Lemma A.1 that \textrm{AR}_{\textrm{C}}\left(\beta_{n,0}\right)\leq n\tilde{g}_{n,0}^{\prime}\tilde{\Omega}_{n,0}^{-1/2}M_{\hat{A}_{n}}\tilde{\Omega}_{n,0}^{-1/2}\tilde{g}_{n,0}+\varpi_{n}, where along converging subsequences n_{k} \varpi_{n_{k}}=n_{k}\tilde{g}_{n_{k}}^{\prime}\tilde{\Omega}_{n_{k}}^{-1/2}P_{\hat{A}_{n_{k}}}\tilde{\Omega}_{n_{k}}^{-1/2}\tilde{g}_{n_{k}}=o_{p}\left(1\right). This result holds because \tilde{g}_{n,0}=\hat{g}_{n}\left(\tilde{\gamma}_{n,0}\right) approximately satisfies the orthogonality condition in Lemma 3.1 (ii). With the help of Lemma 3.1 and ( 22 ) we show in Lemma A.1 that along converging subsequences n_{k}\tilde{g}_{n_{k},0}^{\prime}\tilde{\Omega}_{n_{k},0}^{-1/2}M_{\hat{A}_{n_{k}}}\tilde{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0}=n_{k}\hat{g}_{n_{k},0}^{\prime}\hat{\Omega}_{n_{k},0}^{-1/2}M_{\hat{A}_{n_{k}}}\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}+o_{p}\left(1\right). By Assumption 6 \hat{g}_{n_{k},0} converges to a Gaussian limit variable \omega_{\varepsilon} . We show in Lemma A.2 that along converging subsequences P_{\hat{A}_{n_{k}}}\rightarrow_{d}P_{A} where P_{A} has constant rank d_{\gamma} independent of the subsequential limit and where P_{A} is a possibly random matrix with elements that are stochastically independent of \omega_{\varepsilon}. The last result is essentially due to Lemma 3.1 (ii) which is at the heart of establishing asymptotic uncorrelatedness between the elements of \hat{A}_{n_{k}} and \hat{g}_{n_{k},0} along all converging subsequences. Combining these results, we establish in Lemma A.3 that \textrm{AR}_{\textrm{C}}\left(\beta_{n,0}\right) is bounded above by a random variable that is asymptotically \chi_{d-d_{\gamma}}^{2} along all converging subsequences. 4 Examples Our general framework in Section 2 and our theoretical results in Section 3 do not assume that the data are generated by a parametric model. Nevertheless, parametric models are often the basis for empirical work and serve as useful benchmarks that allow more precise interpretations of the parameters \theta. In this section we consider two specific examples of such parametric models and discuss how they fit into our general framework. 4.1 Linear Instrumental Variables Consider as an example a simple univariate static model where g_{t}\left(.\right) is of the form 3 with d_{q}=1 and q\left(.\right) is linear. The data generating process is given by y_{n,t}=\beta x_{n,t}+w_{n,t}^{\prime}\gamma_{n}+\varepsilon_{n,t} (23) and \displaystyle x_{n,t} \displaystyle=\pi_{n,x}^{\prime}z_{n,t}+V_{n,x,t} (24) \displaystyle w_{n,t} \displaystyle=\pi_{n,w}^{\prime}z_{n,t}+V_{n,w,t} (25) where z_{n,t},\varepsilon_{n,t} and V_{n,t}=\left(V_{x,t}^{\prime},V_{w,t}^{\prime}\right)^{\prime} are the basic inputs with z_{n,t} observed and \varepsilon_{n,t},V_{n,t} unobserved. We assume that \left(z_{n,t},\varepsilon_{n,t},V_{n,t}^{\prime}\right) are independent across t and where w_{n,t} has d_{w} elements and z_{n,t} has d_{z}>d_{w} elements and x_{n,t} is univariate. The parameter vector \pi_{x,n} and the d_{z}\times d_{w} matrix \pi_{w,n} control instrument strength. The model is the same as in GKMC12 but without the assumption of homoskedasticity of \varepsilon_{n,t} and V_{n,t} , nor do we assume that the random variables are identically distributed over t. At the expense of a slightly more restrictive example, one could assume that \left(z_{n,t},\varepsilon_{n,t},V_{n,t}^{\prime}\right) and \gamma_{n} do not depend on n. This scenario would still contain weak instrument limits as considered by Staiger and Stock (1997) by allowing \pi_{x,n} and \pi_{w,n} to be in a 1/\sqrt{n} neighborhood local to zero, but would also include non-identified scenarios where \text{$\pi_{x,n}$}=0 and/or \pi_{w,n}=0. In the more general setting considered here and in GKMC12 one can allow for \operatorname{Cov}\left(\varepsilon_{n,t},V_{n,t}\right) to depend on n , and thus add additional dimensions along which inference may be challenging. If E\left[z_{n,t}\left(\varepsilon_{n,t},V_{n,t}^{\prime}\right)\right]=0 then the moment conditions n^{-1}\sum_{t=1}^{n}E\left[\left(y_{n,t}-X^{\prime}_{n,t}\theta\right)z_{n,t}\right]=0 have at least one solution \theta_{n}=\left(\beta,\gamma_{n}^{\prime}\right)^{\prime} . If n^{-1}\sum_{t=1}^{n}E\left[X_{n,t}z_{n,t}^{\prime}\right] is of reduced column rank then the model is not identified and the average moment condition may have multiple solutions. The CUE for this model is formed from moment functions g_{t}\left(\beta,\gamma\right)=\left(\begin{array}[]{c}y_{n,t}-x_{n,t}\beta-w^{\prime}_{n,t}\gamma\end{array}\right)z_{n,t} (26) and can be evaluated at H_{0}:\beta=\beta_{0} by setting \hat{g}_{n}\left(\beta_{0},\gamma\right)=0 . The weight matrix \hat{\Omega}_{n} is exploiting independence and is given by \hat{\Omega}_{n}\left(\beta,\gamma\right)=n^{-1}\sum_{t=1}^{n}g_{t}\left(\beta,\gamma\right)g_{t}\left(\beta,\gamma\right)^{\prime}. (27) The criterion function Q_{n}\left(\beta,\gamma\right) and the test statistic \textrm{AR}_{\textrm{C}} then are formulated in the same way as in Section 2 using the expressions for \hat{g}_{n} and \hat{\Omega}_{n} given in ( 26 ) and ( 27 ). The derivatives E\left[g_{n,\gamma}\right]=n^{-1}\sum_{t=1}^{n}E\left[z_{n,t}w_{n,t}^{\prime}\right]=n^{-1}\sum_{t=1}^{n}E\left[z_{n,t}z_{n,t}^{\prime}\right]\pi_{n,w} are reduced column rank if \pi_{n,w} is of reduced column rank. Assumption 1 is satisfied in the following way. Assume \eta_{n,t}=\left(z_{n,t}^{\prime},\varepsilon_{n,t},V_{n,x,t},V_{n,w,t}^{\prime}\right)^{\prime} is a double array of independent random variables, where in addition z_{n,t} is independent of \varepsilon_{n,t} , V_{n,x,t} and V_{n,w,t} . Assume that y_{n,t} , x_{n,t} and w_{n,t} are generated by ( 23 ), ( 24 ) and ( 25 ) respectively for parameters \theta, \pi_{n,x} and \pi_{n,w} in some compact parameter space \Theta. When the DGP is linear as in ( 23 )-( 25 ) then Assumptions ( 3 ), ( 4 ) and ( 5 ) can be replaced with the following assumption. Assumption 7 . For all n\geq 1 and t\geq 1: E\left[\left\|\eta_{n,t}\right\|_{2}^{2}\right]<\infty . Let W_{n,\eta\eta}=n^{-1}\sum_{t=1}^{n}\text{$E$$\left[\eta_{n,t}\eta_{n,t}^{\prime}\right]$} , \hat{W}_{n,\eta\eta}=n^{-1}\sum_{t=1}^{n}\text{$\eta_{n,t}\eta_{n,t}$}^{\prime} . For all converging subsequences n_{k} assume that W_{n_{k},\eta\eta}-W_{\eta\eta}=o\left(1\right) where W_{\eta\eta} is a symmetric full rank matrix of non-random constants. In addition, \hat{W}_{n_{k},\eta\eta}-W_{n_{k},\eta\eta}=o_{p}\left(1\right) . Assumption ( 7 ) holds under the following sufficient low level assumptions. Let \eta_{n,t,j} be the j -th element of \eta_{n,t} . Assume \sup_{t}E\left[\left|\eta_{n,t,j}\right|^{2+\delta}\right]<\infty for some 0<\delta<1 and all j and all n\geq 1 . This implies that the elements \chi_{n,t,j} of \chi_{n,t} also have bounded 2+\delta moments. Also assume that there exist subsequences n_{k} such that n_{k}^{-1}\sum_{t=1}^{n_{k}}E\left[\chi_{n_{k},t,j}\chi_{n_{k},t,l}\right] converges to some constant c that may depend on the subsequence n_{k} . If \chi_{n,t} does not depend on n Theorem 1 in Chow and Teicher 1997 (1997, p.124) implies that Assumption 7 holds for all such converging subsequences. A strong law for double arrays of independent heterogenous \chi_{n,t} is given in Teicher 1985 . Assumption 3 holds if for \varepsilon_{n,t}\left(\theta\right)=y_{n,t}-x_{n,t}\beta-w^{\prime}_{n,t}\gamma it follows that \inf_{\theta}\det\left(E\left[\varepsilon_{n,t}\left(\theta\right)^{2}z_{n,t}z_{n,t}^{{}^{\prime}}\right]\right)>0. Sufficient conditions are that \sup_{\theta}\Pr\left(\varepsilon_{n,t}\left(\theta\right)=0\right)<1 which holds if y_{n,t} is not co-linear with x_{n,t} and w_{n,t} and \sup_{a}\Pr\left(z_{n,t}^{\prime}a=0\right)<1 . Assumption ( 6 ) holds if for any \ell\in\mathbb{R}^{\left(2+d_{w}\right)d_{z}} and \nu_{n,t}=\left(\varepsilon_{n,t},V_{n,x,t},V_{n,w,t}^{\prime}\right)^{\prime}\otimes z_{n,t} it follows that E\left[\ell^{\prime}\nu_{n,t}\right]=0 , \sigma_{n,t}^{2}=E\left[\left(\ell^{\prime}\nu_{n,t}\right)^{2}\right]>0 , c_{n}^{2}=\sum_{t=1}^{n}\sigma_{n,t}^{2} and \lim{}_{k\rightarrow\infty}\sum_{t=1}^{n_{k}}E\left[\left(\ell^{\prime}\nu_{n_{k},t}/c_{n_{k}}\right)^{2}1\left\{\left|\ell^{\prime}\nu_{n_{k},t}\right|>\varepsilon c_{n_{k}}\right\}\right]\rightarrow 0 for every converging subsequence n_{k} by Theorem 9.6.1 of Dudley 1989 . 4.2 Local Projections We consider time series data with a non-parametric data generating process and dependence structure. The parameter of interest is the local projection of Jordà 2005 for a univariate outcome variable y_{n,t+h} at horizons h\in\left\{0,...,H\right\} onto a policy or intervention variable x_{n,t} and additional controls w_{n,t} under heterogeneity and nonstationarity. To simplify the discussion we assume that x_{n,t} is univariate. We assume the researcher is specifying a linear model under H_{0} which is formulated as a restriction on the conditional mean function E\left[y_{n,t+h}|x_{n,t},w_{n,t}\right]=x_{n,t}\beta_{t}\left(h\right)+w^{\prime}_{n,t}\gamma_{t}\left(h\right) and where \beta_{t}\left(h\right) for h\in\left\{0,...,H\right\} are the parameters of interest. As in Section 2 we only assume that a linear moment restriction holds under H_{0} but we do not assume that y_{n,t} is necessarily generated by a linear model. The conditioning variables w_{n,t} summarize information \mathcal{I}_{t-1} up to time t-1 , which may also contain instrumental variables necessary to identify structural parameters or to give the impulse coefficient a causal interpretation. When x_{n,t} depends on unobserved components that are potentially correlated with unobserved factors affecting the outcome variable, identification is achieved with an instrumental variable z_{n,t} . In addition, in settings without a recursive structure some of the variables in w_{n,t} may also be endogenous. Note that in the case of selection on observables, z_{n,t} may be identical to x_{n,t} and w_{n,t} . Let d=H+1, Y_{n,t}=\left(y_{n,t},....,y_{n,t+H}\right) , \beta_{n}=\left(\beta_{n}\left(0\right),....,\beta_{n}\left(H\right)\right)^{\prime} and \gamma_{n}=\left(\gamma_{n}\left(0\right)^{\prime},....,\gamma_{n}\left(H\right)^{\prime}\right)^{\prime} . Let X_{n,t}=\left(x_{n,t},w_{n,t}^{\prime}\right)^{\prime} , q\left(Y_{n,t},X_{n,t},\beta,\gamma\right)=\left(Y_{n,t}-\left(I_{d}\otimes x_{n,t}\right)\beta-\left(I_{d}\otimes w^{\prime}_{n,t}\right)\gamma\right) (28) and g_{t}\left(\beta,\gamma\right)=q\left(Y_{n,t},X_{n,t},\beta,\gamma\right)\otimes z_{n,t}. If the underlying process generating the observed data is weakly stationary \beta_{t}\left(h\right) does not depend on t . In the absence of stationarity, we define \beta_{n,0},\gamma_{n,0} as the solution to n^{-1}\sum_{t=1}^{n}E\left[g_{t}\left(\beta_{n,0},\gamma_{n,0}\right)\right]=0 (29) for any fixed n\geq 1 . It is then possible to define pseudo residuals \varepsilon_{n,t}:=Y_{t}-\left(I_{d}\otimes x_{n,t}\right)\beta_{n,0}-\left(I_{d}\otimes w^{\prime}_{n,t}\right)\gamma_{n,0} which inherit the property \sum_{t=1}^{n}E\left[\varepsilon_{n,t}\left(h\right)z_{n,t}\right]=0 from ( 29 ) under H_{0} . With \hat{g}_{n}\left(\beta,\gamma\right) as defined in ( 2 ) and \hat{\Omega}_{n}\left(\beta,\gamma\right) as defined in ( 4 ) the statistic \textrm{A$\textrm{R}_{C}$}\left(\beta_{n,0}\right) is defined in ( 5 ). Assumption ( 1 ) is satisfied for many weakly dependent processes. Assumption 3 can be satisfied by imposing the conditions of de Jong and Davidson 2000 which include L 2 near-epoch dependent triangluar arrays. Assumptions 4 and 5 can be replaced by Assumption 7 due to linearity of 28 . Assumption 7 holds for L_{1} -mixingale arrays and L_{1} -near epoch dependent arrays by results for weak laws of large numbers in Andrews 1988 . De Jong (1996) de Jong 1996 proves strong laws for L_{q} -mixingale arrays for q\geq 1. The multivariate central limit theorem in Assumption 6 can be established with the help of the Cramer-Wold theorem and by verifying regularity conditions for \lambda^{\prime}\Sigma_{n}^{-1/2}\nu_{n,t} for all \lambda with \left\|\lambda\right\|=1 . The assumption that E\left[\text{$\left\|S_{n}\right\|^{2}$}\right]<\infty is not sufficient in itself to guarantee a CLT without additional restrictions on dependence. Low level conditions for univariate double array processes can be found in McLeish 1977 who imposes a weak dependence condition in the form of a mixingale assumption. McLeish’s theory allows for non-stationarity, while trending behavior, including unit roots are ruled out by the assumptions on dependence. Wooldridge and White 1988 prove a related result under near epoch dependence, which implies a mixingale property (see Proposition 2.9 of Wooldridge and White 1988 , 1988). They also prove a multivariate version, directly leading to the condition in Assumption 6 . For non-stationary triangular arrays instead of double arrays, de Jong 1997 proves a central limit theorem directly for heterogenous L_{2} -mixingale arrays. The result is obtained by showing that a blocking scheme can be used to obtain a martingale approximation for which a martingale CLT can be applied. de Jong 1997 and Davidson and de Jong 2000 establish a functional CLT that is based on de Jong 1997 . Rio 1997 establishes a CLT extending Lindeberg’s CLT for iid random variables to double arrays of strongly mixing random variables. An alternative approach is explored by Neumann 2013 and Merlevede, Peligrad and Utev ( Merlevède et al. 2019 ) who impose a maximal correlation measure to control dependence. When \nu_{n,t} is a martingale difference array, McLeish 1974 , see also Hall and Heyde 1980 , provides a central limit theorem. 4.3 Implementation Practitioners are often more interested in confidence intervals for a single parameter than in confidence sets for the entire parameter vector. One can obtain such confidence intervals from our confidence set via the projection method, see e.g., Dufour 1997 and Dufour and Taamouti 2005 . When \beta is a scalar ( d_{\beta}=1 ), which is often the case in empirical applications of LP-IV, a 100(1-\alpha) % confidence interval for \beta is given by [\underline{\beta}_{1-\alpha},\overline{\beta}_{1-\alpha}] , where \underline{\beta}_{1-\alpha}=\min\left\{\beta|AR_{c}(\beta)\leq c_{1-\alpha}\right\} , \overline{\beta}_{1-\alpha}=\max\left\{\beta|AR_{c}(\beta)\leq c_{1-\alpha}\right\} , and c_{1-\alpha} is the 1-\alpha quantile of the \chi^{2} distribution with d-d_{\gamma} degrees of freedom. If the confidence set is convex, this confidence interval has exact (1-\alpha) asymptotic coverage probability. Otherwise, it is conservative. When d_{\beta}>1 , let \tilde{c}_{1-\alpha} be the 1-\alpha quantile of the \chi^{2} distribution with d-d_{\beta}-d_{\gamma}+1 degrees of freedom. A 100(1-\alpha) % confidence interval for \beta_{1} (without of loss of generality) is given by [\underline{\beta}_{1,1-\alpha},\overline{\beta}_{1,1-\alpha}] , where for some \beta_{2:d_{\beta}} \displaystyle\underline{\beta}_{1,1-\alpha} \displaystyle= \displaystyle\min_{\beta_{1}}\left\{\beta_{1}|AR_{c}(\beta_{1},\beta_{2:d_{\beta}})\leq\tilde{c}_{1-\alpha}\right\}, (30) \displaystyle\overline{\beta}_{1,1-\alpha} \displaystyle= \displaystyle\max_{\beta_{1}}\left\{\beta_{1}|AR_{c}(\beta_{1},\beta_{2:d_{\beta}})\leq\tilde{c}_{1-\alpha}\right\}, (31) and \beta_{2:d_{\beta}} denotes the (d_{\beta}-1)\times 1 subvector of \beta excluding \beta_{1} . In practice, the condition, AR_{c}(\beta_{1},\beta_{2:d_{\beta}})\leq\tilde{c}_{1-\alpha} , can be checked by \min_{\beta_{2:d_{\beta}},\gamma}Q_{n}(\beta,\beta_{2:d_{\beta}},\gamma)\leq\tilde{c}_{-1-\alpha} for a given value of \beta_{1} , where the minimum on the left-hand side is achieved by numerical optimization methods. For example, to obtain \underline{\beta}_{1,1-\alpha} [ \overline{\beta}_{1,1-\alpha} ], one minimizes \beta_{1} [resp. - \beta_{1} ] subject to \min_{\beta_{2:d_{\beta}},\gamma}Q_{n}([\beta_{1}\;\beta_{2:d_{\beta}}],\gamma)\leq\tilde{c}_{1-\alpha} . In MATLAB, one can use fminunc to evaluate the constraint function and use fmincon to minimize \beta_{1} [ -\beta_{1} ]. When d_{\beta}>1 , these confidence intervals are conservative in that the asymptotic coverage probability is greater than or equal to 1-\alpha . 5 Monte Carlo Experiments The data generating process is based on the bivariate VAR model of inflation ( \pi_{t} ) and output gap ( x_{t} ) in equations (4) and (7) of Kleibergen and Mavroeidis 2009 : \displaystyle x_{t} \displaystyle= \displaystyle\rho_{1}x_{t-1}+\rho_{2}x_{t-2}+\nu_{t}, (32) \displaystyle\pi_{t} \displaystyle= \displaystyle(\alpha_{0}\rho_{1}+\alpha_{1})x_{t-1}+\alpha_{0}\rho_{2}x_{t-2}+\eta_{t}, (33) where \rho_{1}=\rho(1-\rho_{2})=0.9(1-\rho_{2}) , \alpha_{0}=\lambda/[1-\gamma_{f}(\rho_{1}+\gamma_{f}\rho_{2})] , \alpha_{1}=\lambda\gamma_{f}\rho_{2}/[1-\gamma_{f}(\rho_{1}+\gamma_{f}\rho_{2})] , \lambda=\rho_{\eta\nu}\frac{\sigma_{\eta}}{\sigma_{\nu}}\left(1-\gamma_{f}\left(\rho_{1}+\gamma_{f}\rho_{2}\right)\right) . We introduce conditional heteroskedasticity to allow the conditional variance of the moment conditions to depend on the instruments: \eta_{t}=\exp(0.5h_{1t})\epsilon_{1t}/\kappa , \nu_{t}=\exp(0.5h_{2t})\epsilon_{2t}/\kappa , \epsilon_{1t} and \epsilon_{2t} are iid N(0,1) with correlation \rho_{\eta\nu} , , \kappa is set so that \textrm{Var}(\eta_{t})=\sigma_{\eta}^{2}=1 and \textrm{Var}(\nu_{t})=\sigma_{\nu}^{2}=1 , and h_{1t} and h_{2t} are two independent AR(1) processes: \displaystyle h_{1t} \displaystyle= \displaystyle 0.9h_{1,t-1}+\xi_{1,t}, (34) \displaystyle h_{2t} \displaystyle= \displaystyle 0.9h_{2,t-1}+\xi_{2,t}, (35) and \xi_{1t} and \xi_{2t} are independent iid N(0,0.2) random variables that are independent of \epsilon_{1t} and \epsilon_{2t} . We estimate the Phillips curve, \pi_{t}\;=\;\lambda x_{t}+\gamma_{f}E_{t}(\pi_{t+1})+u_{t}, (36) using the moment conditions, E\left[z_{t}(\pi_{t}-\lambda x_{t}-\gamma_{f}\pi_{t+1})\right]\;=\;0, (37) We consider two choices of instruments: (i) z_{t}=[\pi_{t-1}\;x_{t-1}\;\pi_{t-2}\;x_{t-2}^{\prime}\;\pi_{t-3}\;x_{t-3}]^{\prime} as in Kleibergen and Mavroeidis 2009 and (ii) z_{t}=[x_{t}\;x_{t-1}] which represents a just identified specification. It follows from equation 6 and other equations in Kleibergen and Mavroeidis 2009 that the population reduced-form equations for the RHS endogenous variables can be written as \left[\begin{array}[]{c}\pi_{t+1}\\ x_{t}\end{array}\right]\;=\;\left[\begin{array}[]{cc}\alpha_{0}((\rho_{1}+\rho_{2}\gamma_{f})\rho_{1}+\rho_{2})&\alpha_{0}(\rho_{1}+\rho_{2}\gamma_{f})\rho_{2}\\ \rho_{1}&\rho_{2}\end{array}\right]\left[\begin{array}[]{c}x_{t-1}\\ x_{t-2}\end{array}\right]+\left[\begin{array}[]{c}(\alpha_{1}+\alpha_{0}\rho_{1})\nu_{t}+\eta_{t+1}\\ \nu_{t}\end{array}\right]. (38) There are at least two cases of identification failures. First, when \rho_{2}=0 , the projection coefficient matrix simplifies to \left[\begin{array}[]{cc}\rho^{2}\rho_{\eta\nu}&0\\ \rho&0\end{array}\right]. (39) Thus, \gamma_{f} and \lambda are only partially identified (i.e., their linear combination is identified). Second, when \gamma_{\eta\nu}=0 , the matrix simplifies to \left[\begin{array}[]{cc}0&0\\ \rho(1-\rho_{2})&\rho_{2}\end{array}\right]. (40) In this case, \gamma_{f} is identified but \lambda is not. We consider \rho_{2}\in\{-0.05,-0.65,-0.99\} and \textrm{Corr}(\eta_{t},\nu_{t})\in\{0.2,0.99\} . The number of Monte Carlo simulations is set to 5,000. For HAC estimation, the lag truncation parameter is set to zero, as the moment conditions are serially uncorrelated. Tables 1 and 2 report the finite-sample size of the t test, the subset KLM test and the subset AR_{c} test when testing \gamma_{f}=0.5 at the 5% significance level. Table 1 shows that, when the three lags of \pi_{t} and x_{t} are used as instruments, the conventional t test suffers from size distortions especially when the instruments are weak and the degree of endogeneity is high ( \rho_{2}=-0.05 and \rho_{\epsilon\eta}=0.99 ) However, when the parameters are just identified, the size distortions of the conventional t test are much smaller, as shown in Table 2. In either case, the AR_{c} and KLM tests do not suffer from size distortions as expected. Figures 1 (a)-(f) and Figures 2 (a)-(f) show the power curve of these three tests when z_{t}=[\pi_{t-1}\;x_{t-1}\;\pi_{t-2}\;x_{t-2}\;\pi_{t-3}\;x_{t-3}]^{\prime} and z_{t}=[x_{t-1}\;x_{t-2}]^{\prime} , respectively. The sample size is set to 1,000. When the instruments are weak ( \rho_{2}=-0.05 or \rho_{\epsilon\eta}=0.0 ), the \textrm{AR}_{c} and KLM tests have correct size, although they do not have any power, as shown in Figures 1 (a) and (d) and Figures 2 (a) and (d). 3 3 3 Interestingly, when \rho_{\epsilon\eta}=0.0 , the AR_{c} and KLM tests do not have power when testing \gamma_{f}=0.5 . As the instruments become stronger, the tests become more powerful, as demonstrated in Figures 1 (b),(c) and (d),(f) and Figures 2 (b),(c) and (d),(f). When the parameters are overidentified and are very strongly identified, the KLM test is more powerful than the \textrm{AR}_{c} test, seeFigures 1 (c) and (f). However, when the parameter is mildly strongly identified, the \textrm{AR}_{c} test is more powerful than the KLM test when the true parameter value is close to the null value, see Figures Figures 1 (b) and (d). When the parameters are just identified, their power curves are virtually identical, as seen in Figures Figures 2 (a) through (f). These results are consistent with Moreira 2009 who shows that in the linear Gaussian case the just identified \textrm{AR}_{c} based tests as well as score tests are uniformly most powerful for certain testing problems. 6 Conclusions We obtain a theoretical upper bound for the Anderson Rubin subset statistic based on the continuous updating estimator for general parametric non-linear moment restrictions with heteroskedasticity, dependence and heterogeneity. The upper bound is shown to be valid for all converging subsequences of data distributions irrespective of whether nuisance parameters are identified or not. The upper bound is sharp in the sense that the \textrm{AR}_{c} attains it when nuisance parameters are identified. We establish that the upper bound has a limiting \chi_{d-d_{\gamma}}^{2} distribution along all converging subsequences. These results imply that the \textrm{AR}_{c} statistic has uniformly correct size and can be used to construct confidence intervals that are uniformly valid. Table 1: Finite-sample size (5% nominal size): z_{t}=[\pi_{t-1}\;x_{t-1}\;\pi_{t-2}\;x_{t-2}\;\pi_{t-3}\;x_{t-3}]^{\prime} \rho_{\epsilon\eta}=0.0 \rho_{\epsilon\eta}=0.2 \rho_{\epsilon\eta}=0.99 T \rho_{2} t -test AR_{c} KLM t -test AR_{c} KLM t -test AR_{c} KLM 100 0.0 0.148 0.041 0.022 0.158 0.043 0.023 0.970 0.033 0.075 100 -0.05 0.139 0.041 0.020 0.155 0.043 0.017 0.970 0.034 0.075 100 -0.65 0.099 0.034 0.016 0.102 0.033 0.017 0.172 0.030 0.026 500 0.0 0.124 0.058 0.017 0.131 0.053 0.018 0.978 0.046 0.043 500 -0.05 0.123 0.057 0.016 0.140 0.056 0.022 0.971 0.049 0.047 500 -0.65 0.086 0.038 0.014 0.066 0.041 0.012 0.070 0.050 0.042 Table 2: Finite-sample size (5% nominal size): z_{t}=[x_{t-1}\;x_{t-2}]^{\prime} \rho_{\epsilon\eta}=0.0 \rho_{\epsilon\eta}=0.2 \rho_{\epsilon\eta}=0.99 T \rho_{2} t -test AR_{c} KLM t -test AR_{c} KLM t -test AR_{c} KLM 100 0.0 0.014 0.044 0.044 0.016 0.060 0.060 0.450 0.044 0.044 100 -0.05 0.012 0.043 0.043 0.014 0.054 0.054 0.329 0.050 0.050 100 -0.65 0.001 0.010 0.010 0.005 0.013 0.013 0.056 0.044 0.044 500 0.0 0.010 0.047 0.047 0.013 0.067 0.067 0.453 0.051 0.051 500 -0.05 0.010 0.046 0.046 0.011 0.063 0.063 0.185 0.054 0.054 500 -0.65 0.000 0.007 0.007 0.012 0.011 0.011 0.050 0.050 0.050 (a) \rho_{2}=-0.05 (d) \rho_{\eta\nu}=0.0 (b) \rho_{2}=-0.65 (e) \rho_{\eta\nu}=0.2 (c) \rho_{2}=-0.99 (f) \rho_{\eta\nu}=0.99 Figure 1: (a)-(f): z_{t}=[\pi_{t-1}\;x_{t-1}\;\pi_{t-2}\;x_{t-2}\;\pi_{t-3}\;x_{t-3}]^{\prime} ; (a)-(c): \rho_{\eta\nu}=0.2 , (d)-(f): \rho_{2}=-0.65 (a) \rho_{2}=-0.05 (d) \rho_{\eta\nu}=0.0 (b) \rho_{2}=-0.65 (e) \rho_{\eta\nu}=0.2 (c) \rho_{2}=-0.99 (f) \rho_{\eta\nu}=0.99 Figure 2: (a)-(f): z_{t}=[x_{t-1}\;x_{t-2}]^{\prime} ; (a)-(c): \rho_{\eta\nu}=0.2 , (d)-(f): \rho_{2}=-0.65 Appendix A Proofs A.1 Preliminaries The proof of Theorem 3.1 is based on Lemmas 3.1 , A.1 , A.2 and A.3 . We prove Lemma 3.1 in Section A.3 and we state and prove Lemmas A.1 , A.2 and A.3 in Section A.4 which also contains the proof of Theorem 3.1 . We collect some results on perturbation bounds in Section A.2 . A.2 Perturbation Bounds We state the following result for ease of reference that links the converge in operator norm of matrices to the convergence of the corresponding eigenvalues. For a d_{\gamma}\times d_{\gamma} Hermitian matrices \mathscr{A}_{n} and \mathscr{A} such that \left\|\mathscr{A}_{n}-A\right\|\rightarrow 0 let \lambda_{1}\left(\mathscr{A}\right)\geq...\geq\lambda_{d_{\gamma}}\left(\mathscr{A}\right) and \lambda_{1}\left(\mathscr{A}_{n}\right)\geq...\geq\lambda_{d_{\gamma}}\left(\mathscr{A}_{n}\right) be the eigenvalues of \mathscr{A} and \mathscr{A}_{n} in decreasing order. Theorem A.1 . (Weyl’s perturbation theorem) For \mathscr{A}, \mathscr{A}_{n} and \lambda_{j}\left(.\right) defined above it follows that \text{max}_{j}\left|\lambda_{j}\left(\mathscr{A}\right)-\lambda_{j}\left(\mathscr{A}_{n}\right)\right|\leq\left\|\mathscr{A}-\mathscr{A}_{n}\right\|\rightarrow 0. (41) For a proof of the theorem see Bhatia 1996 (1996, Corollary III.2.6). A related result links the convergence of eigenvectors to the converge in operator norm of matrices. The convergence of the eigenvectors is in terms of the supspaces they span. The following result is slightly adapted from Bhatia 1996 (1996, Theorem VII.3.2). Theorem A.2 . Let \mathscr{L}_{1} and \mathscr{L}_{2} be two subsets of \mathbb{R} such that \textrm{dist}\left(\mathscr{L}_{1},\mathscr{L}_{2}\right)=\delta>0 for some \delta. Let \mathscr{A}_{n} and \mathscr{A} be Hermitian matrices with eigenvectors R_{n}, and R respectively. Let \hat{R}_{1,n}=P_{\mathscr{A}_{n}}\left(\mathscr{L}_{1}\right) be the projection matrix projecting onto the subspace spanned by the eigenvectors R_{n} of \mathscr{A}_{n} associated with eigenvalues who are elements of the set \mathscr{L}_{1}. Similarly, define \hat{R_{2}}=P_{\mathscr{A}}\left(\mathscr{L}_{2}\right) . Then, \left\|\hat{R}_{1,n}\hat{R_{2}}\right\|\leq\frac{\pi}{2\delta}\left\|\mathscr{A}_{n}-\mathscr{A}\right\|. (42) For a proof of the theorem see Bhatia 1996 (1996, p.212). A.3 Proof of Lemma 3.1 Proof of Lemma 3.1 . First introduce some notation. Recall the definition of \hat{\Omega}_{n}\left(\theta\right) in ( 4 ) and \Omega_{n}\left(\theta\right) in Assumption 3 which implies \det\left(\Omega_{n}\left(\theta\right)\right)\geq K_{\Omega}>0 uniformly in \theta and define \hat{\Omega}_{n,0}=\hat{\Omega}_{n}\left(\theta_{n,0}\right). For \theta_{1}=\left(\beta_{1},\gamma_{1}\right) and \theta_{2}=\left(\beta_{2},\gamma_{2}\right) define \hat{\Omega}_{n}\left(\theta_{1},\theta_{2}\right)=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}k\left(\frac{t-s}{a_{n}}\right)g_{t}\left(\beta_{1},\gamma_{1}\right)g_{s}\left(\beta_{2},\gamma_{2}\right)^{\prime}. Next consider \hat{g}_{n}\left(\beta,\gamma\right)=n^{-1}\sum_{t=1}^{n}g_{t}\left(\beta,\gamma\right) and define \hat{g}_{n}\left(\gamma\right)=\hat{g}_{n}\left(\beta_{n,0},\gamma\right), (43) \hat{g}_{n,0}=\hat{g}_{n}\left(\beta_{n,0},\gamma_{n,0}\right), (44) g_{t}^{\gamma}\left(\gamma\right)=\partial g_{t}\left(\beta_{n,0},\gamma\right)/\partial\gamma^{\prime}, (45) where g_{t}^{\gamma}\left(\gamma\right) is a d\times d_{\gamma} matrix of functions. Use the short hand notation \hat{g}_{n,\gamma}=\hat{g}_{n,\gamma}\left(\gamma_{n,0}\right) for \hat{g}_{n,\gamma}\left(\gamma\right) defined in ( 9 ). In the same way, it follows from \varepsilon_{n,t}=g_{t}\left(\beta_{n,0},\gamma_{n,0}\right) in ( 8 ) that \displaystyle\hat{\Omega}_{n}\left(\theta_{n,0},\left(\beta_{n,0},\gamma\right)\right) \displaystyle=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}k\left(\frac{t-s}{a_{n}}\right)g_{t}\left(\beta_{n,0},\gamma_{n,0}\right)g_{s}\left(\beta_{n,0},\gamma\right)^{\prime} \displaystyle=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}k\left(\frac{t-s}{a_{n}}\right)\varepsilon_{n,t}g_{s}\left(\beta_{n,0},\gamma\right)^{\prime}. To show (i) consider an arbitrary converging subsequence n_{k} with limit \Gamma=g_{\gamma}^{\prime}\Omega_{0}^{-1}g_{\gamma} and eigenvalues \Delta_{j} in Assumption 4 . Fix \varepsilon such that \Delta_{r}>\varepsilon>0, and recall that \left\|.\right\| is the operator norm such that \left\|\mathring{\Gamma}_{n}^{+}\right\|\leq\frac{1}{\min_{j\text{ s.t.$\hat{\lambda}_{n,j}$$>\varepsilon$ }}\hat{\lambda}_{n,j}}\leq\varepsilon^{-1}<\infty where \mathring{\Gamma}_{n} is defined in ( 21 ). This shows that \left\|\mathring{\Gamma}_{n_{k}}^{+}\right\|=O_{p}\left(1\right). The result in (i) then follows from \hat{A}_{n_{k}}=O_{p}\left(1\right) which holds by ( 17 ) and Assumptions 4 , 5 and 6 , \hat{\Omega}_{n_{k},0}^{-1/2}=O_{p}\left(1\right) which holds by Assumption 3 and \hat{g}_{n_{k},0}=O_{p}\left(n_{k}^{-1/2}\right) which holds by ( 8 ) and Assumption 6 . For the second result in (ii) consider any of the converging subsequences n_{k}. Note that by Assumptions 1 and 4 it follows that \hat{g}_{n_{k},\gamma}\rightarrow_{p}g_{\gamma} for some d\times d_{\gamma} matrix g_{\gamma} that may depend on n_{k} and that \hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},\gamma}=\Omega_{0}^{-1/2}g_{\gamma}+o_{p}\left(1\right). Then, define \bar{A}=\Omega_{0}^{-1/2}g_{\gamma} (46) and \Gamma=g_{\gamma}^{\prime}\Omega_{0}^{-1}g_{\gamma}=\bar{A}^{\prime}\bar{A} (47) and denote \operatorname{rank}\left(\Gamma\right)=r . First consider the case where r\geq 1. By Theorem A.1 and Assumptions 1 and 4 it follows that \hat{\Delta}_{n_{k},r}\rightarrow_{p}\Delta_{r}>0 . Now set \varepsilon=\Delta_{r}/2>0. When r\geq 1 Hansen 1987 (1987, Theorem 3.2) applies because \Delta_{r}-\Delta_{r+1}=\Delta_{r}>\varepsilon and \left\|\Gamma-\hat{\Gamma}_{n_{k}}\right\|<\varepsilon with probability approaching one by Assumptions 3 and 4 . Then, for n_{k} large enough Hansen 1987 (1987, Theorem 3.2) 4 4 4 Note that Hansen 1987 (1987, p.534) uses \left\|.\right\|_{2} for the spectral norm, which is the same as the operator norm \left\|.\right\| defined in this paper, see Horn and Johnson 1985 (1985, Section 5.6, p.295). provides an upper bound for the difference between the regularized and limiting MP inverse of \Gamma where it follows that \left\|\Gamma^{+}-\mathring{\Gamma}_{n_{k}}^{+}\right\|\leq\frac{3}{\Delta_{r}^{2}\left(1-\eta_{r}\right)}\left\|\Gamma-\hat{\Gamma}_{n_{k}}\right\|=o_{p}\left(1\right) (48) and where \eta_{r}=\frac{\left\|\Gamma-\hat{\Gamma}_{n_{k}}\right\|}{\Delta_{r}}=o_{p}\left(1\right). For the case when r=0 we have \Gamma^{+}=0 and \left\|\mathring{\Gamma}_{n_{k}}^{+}\right\|=o_{p}\left(1\right) because \hat{\Delta}_{n,j}\rightarrow_{p}\Delta_{j}=0 for all j=1,...,d . Now define \bar{A}_{n}=\hat{\Omega}_{n,0}^{-1/2}\hat{g}_{n,\gamma} such that \bar{A}_{n}^{\prime}\bar{A}_{n}=\hat{g}_{n,\gamma}^{\prime}\hat{\Omega}_{n,0}^{-1}\hat{g}_{n,\gamma}=\hat{\Gamma}_{n} (49) and \hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}=O_{p}\left(n_{k}^{-1/2}\right). (50) Then, for \hat{A}_{n} defined in ( 17 ), it follows \hat{A}_{n_{k}}=\bar{A}_{n_{k}}+O_{p}\left(n_{k}^{-1/2}\right) (51) Next, note that by evaluating ( 18 ) at \tilde{\gamma}_{n,0} defined in ( 22 ), setting \tilde{g}_{n_{k},0}=\hat{g}_{n}\left(\tilde{\gamma}_{n,0}\right) and using ( 19 ) and ( 22 ) leads to \displaystyle\hat{A}_{n_{k}}^{\prime}\hat{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0} \displaystyle= \displaystyle\hat{A}_{n_{k}}^{\prime}\hat{\Omega}_{n_{k},0}^{-1/2}\left(\hat{g}_{n_{k},0}+\hat{g}_{n_{k},\gamma}\left(\check{\gamma}_{n_{k},0}\right)\left(\tilde{\gamma}_{n_{k},0}-\gamma_{n_{k},0}\right)\right) (52) \displaystyle= \displaystyle\left(\hat{A}_{n_{k}}^{\prime}-\hat{A}_{n_{k}}^{\prime}\bar{A}_{n_{k}}\mathring{\Gamma}_{n_{k}}^{+}\hat{A}_{n_{k}}^{\prime}\right)\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}+o_{p}\left(n_{k}^{-1/2}\right) where \bar{A} is defined in ( 46 ). Next consider the RHS of ( 52 ) where by ( 49 ) and ( 51 ) \left(\hat{A}_{n_{k}}^{\prime}-\hat{A}_{n_{k}}^{\prime}\bar{A}_{n_{k}}\mathring{\Gamma}_{n}^{+}\hat{A}_{n_{k}}^{\prime}\right)\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}=\left(\bar{A}_{n_{k}}^{\prime}-\hat{\Gamma}_{n_{k}}\mathring{\Gamma}_{n_{k}}^{+}\bar{A}_{n_{k}}^{\prime}\right)\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}+o_{p}\left(n_{k}^{-1/2}\right). (53) The first term on the RHS of ( 53 ) can be approximated as \left(\bar{A}_{n_{k}}^{\prime}-\hat{\Gamma}_{n_{k}}\mathring{\Gamma}_{n_{k}}^{+}\bar{A}_{n_{k}}^{\prime}\right)=\left(\bar{A}^{\prime}-\Gamma\mathring{\Gamma}_{n_{k}}^{+}\bar{A}^{\prime}\right)+o_{p}\left(1\right) (54) where \Gamma is defined in ( 47 ). Then, ( 48 ) implies that \bar{A}\mathring{\Gamma}_{n_{k}}^{+}\bar{A}^{\prime}=P_{\bar{A}}+o_{p}\left(1\right) (55) where P_{\bar{A}}=\bar{A}\left(\bar{A}^{\prime}\bar{A}\right)^{+}\bar{A}^{\prime} is the projection onto the column space spanned by \bar{A} . Note \bar{A}^{\prime}\bar{A}=\Gamma and that \mathring{\Gamma}_{n_{k}}-\Gamma=o_{p}\left(1\right) along converging subsequences, such that by ( 48 ) it follows that \mathring{\Gamma}_{n_{k}}^{+}-\Gamma^{+}=o_{p}\left(1\right). Plugging back ( 55 ) in ( 54 ), and using properties of projections, shows that \bar{A}^{\prime}-\bar{A}^{\prime}\bar{A}\mathring{\Gamma}_{n_{k}}^{+}\bar{A}^{\prime}=\bar{A}^{\prime}-\bar{A}^{\prime}P_{\bar{A}}+o_{p}\left(1\right)=o_{p}\left(1\right). (56) Now, combine ( 52 ), ( 53 ), ( 54 ) and ( 56 ) and use ( 50 ) to establish the claim in (ii). For the last claim in (iii), note that for P_{\hat{A}_{n}}=\hat{A}_{n}\left(\hat{A}_{n}^{\prime}\hat{A}_{n}\right)^{+}\hat{A}_{n}^{\prime} it follows that \displaystyle\hat{A}_{n_{k}}^{\prime}\left(P_{\hat{A}_{n_{k}}}-\hat{A}_{n_{k}}\mathring{\Gamma}_{n_{k}}^{+}\hat{A}_{n_{k}}^{\prime}\right) \displaystyle= \displaystyle\hat{A}_{n_{k}}^{\prime}-\hat{A}_{n_{k}}^{\prime}\hat{A}_{n_{k}}\mathring{\Gamma}_{n_{k}}^{+}\hat{A}_{n_{k}}^{\prime} (57) \displaystyle= \displaystyle\bar{A}^{\prime}-\bar{A}^{\prime}\bar{A}\mathring{\Gamma}_{n_{k}}^{+}\bar{A}^{\prime}+o_{p}\left(1\right)=o_{p}\left(1\right) where the first equality follows from the fact that P_{\hat{A}_{n_{k}}} is a projection, the second equality follows from ( 51 ) and the third equality follows from ( 56 ). Similarly, let \hat{A}_{n_{k}}^{\bot} be the matrix with columns orthogonal to the column space of \hat{A}_{n_{k}}. Then, it follows immediately that \left(\hat{A}_{n_{k}}^{\bot}\right)^{\prime}\left(P_{\hat{A}_{n_{k}}}-\hat{A}_{n_{k}}\mathring{\Gamma}_{n_{k}}^{+}\hat{A}_{n_{k}}^{\prime}\right)=0. (58) Combining ( 57 ) and ( 58 ) and using the fact that projections are uniquely defined by the space they project on, establishes the claim. ∎ A.4 Proof of Theorem 3.1 Lemma A.1 . Let \textrm{AR}_{\textrm{C}} be defined in ( 5 ). Let \theta_{n,0}=\left(\beta_{n,0},\gamma_{n,0}\right) be as in Assumption 2 and for any converging subsequence n_{k} let \tilde{\gamma}_{n,0} be as defined in ( 22 ). Recall the definitions of \tilde{\Omega}_{n,0}=\hat{\Omega}_{n}\left(\beta_{n,0},\tilde{\gamma}_{n,0}\right) , \widetilde{g}_{n,0}=\hat{g}_{n}\left(\beta_{n,0},\tilde{\gamma}_{n,0}\right), \tilde{g}_{n,\gamma}=\partial\hat{g}_{n}\left(\beta_{n,0},\tilde{\gamma}_{n,0}\right)/\partial\gamma^{\prime} , and let \hat{\Lambda}_{n,0}=\hat{\Lambda}_{n}\left(\gamma_{n,0}\right) be as defined in ( 13 ), \hat{A}_{n} as in ( 17 ) and P_{\hat{A}_{n}} the projection onto the column space spanned by \hat{A}_{n} . Let M_{\hat{A}_{n}}=I-P_{\hat{A}_{n}}. Then, along all converging subsequences n_{k}, (i) \textrm{AR}_{\textrm{C}}\left(\beta_{n_{k},0}\right)\leq n_{k}\tilde{g}_{n_{k},0}^{\prime}\tilde{\Omega}_{n_{k},0}^{-1/2}M_{\hat{A}_{n_{k}}}\tilde{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0}+\varpi_{n_{k}}, (59) (ii) \varpi_{n_{k}}=n_{k}\tilde{g}_{n_{k}}^{\prime}\tilde{\Omega}_{n_{k}}^{-1/2}P_{\hat{A}_{n_{k}}}\tilde{\Omega}_{n_{k}}^{-1/2}\tilde{g}_{n_{k}}=o_{p}\left(1\right), (60) (iii) n_{k}\tilde{g}_{n_{k},0}^{\prime}\tilde{\Omega}_{n_{k},0}^{-1/2}M_{\hat{A}_{n_{k}}}\tilde{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0}=n_{k}\hat{g}_{n_{k},0}^{\prime}\hat{\Omega}_{n_{k},0}^{-1/2}M_{\hat{A}_{n_{k}}}\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}+o_{p}\left(1\right). (61) Proof. By Assumption 3 \hat{\Omega}_{n}\left(\beta,\gamma\right) is full rank for all \beta,\gamma w.p.a.1. Thus, w.p.a.1, \hat{\Omega}_{n}^{-1}\left(\beta,\gamma\right) exists. Recall the definitions \hat{g}_{n,0}=\hat{g}_{n}\left(\beta_{n,0},\gamma_{n,0}\right), \hat{\Omega}_{n,0}=\hat{\Omega}_{n}\left(\beta_{n,0},\gamma_{n,0}\right) and \hat{g}_{n,\gamma}=\partial\hat{g}_{n}\left(\beta_{n,0},\gamma_{n,0}\right)/\partial\gamma^{\prime} . We start with the proof of (i). We have the following inequality for all converging subsequences n_{k} : \textrm{min}_{\gamma}\hat{g}_{n_{k}}\left(\beta_{n_{k},0},\gamma\right) ’ \hat{\Omega}_{n_{k}}^{-1}\left(\beta_{n_{k},0},\gamma\right)\hat{g}_{n_{k}}\left(\beta_{n_{k},0},\gamma\right) \displaystyle\leq\tilde{g}_{n_{k},0}^{\prime}\tilde{\Omega}_{n_{k},0}^{-1}\tilde{g}_{n_{k},0} (62) which holds because \tilde{\gamma}_{n,0} does not necessarily minimize the expression on the LHS. Then, \tilde{g}_{n_{k},0}^{\prime}\tilde{\Omega}_{n_{k},0}^{-1}\tilde{g}_{n_{k},0}=\tilde{g}_{n_{k},0}^{\prime}\tilde{\Omega}_{n_{k},0}^{-1/2}M_{\hat{A}_{n_{k}}}\tilde{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0}+\varpi_{n_{k}} since M_{\hat{A}_{n}}+P_{\hat{A}_{n}}=I. This establishes part (i). We now turn to part (iii) first. We mimic standard calculations for GMM, see Hansen 1982 . Recall the mean value expansion of \tilde{g}_{n,0} in 19 which gives \displaystyle\tilde{g}_{n,0} \displaystyle=\hat{g}_{n,0}+\hat{g}_{n,\gamma}\left(\check{\gamma}_{n,0}\right)\left(\tilde{\gamma}{}_{n,0}-\gamma_{n,0}\right). (63) Then premultiply ( 63 ) by \hat{\Omega}_{n,0}^{-1/2} such that \displaystyle\hat{\Omega}_{n,0}^{-1/2}\tilde{g}_{n,0} \displaystyle= \displaystyle\hat{\Omega}_{n,0}^{-1/2}\hat{g}_{n,0}+\hat{\Omega}_{n,0}^{-1/2}\hat{g}_{n,\gamma}\left(\tilde{\gamma}{}_{n,0}-\gamma_{n,0}\right) (64) \displaystyle+\hat{\Omega}_{n,0}^{-1/2}\left(\hat{g}_{n,\gamma}-\hat{g}_{n,\gamma}\left(\check{\gamma}_{n,0}\right)\right)\left(\tilde{\gamma}{}_{n,0}-\gamma_{n,0}\right). By Lemma 3.1 and Assumptions 2 and 4 , it follows that \hat{g}_{n_{k},\gamma}-\hat{g}_{n_{k},\gamma}\left(\check{\gamma}_{n_{k},0}\right)=o_{p}\left(1\right) such that \hat{\Omega}_{n_{k},0}^{-1/2}\left(\hat{g}_{n_{k},\gamma}-\hat{g}_{n_{k},\gamma}\left(\check{\gamma}_{n_{k},0}\right)\right)\left(\tilde{\gamma}{}_{n_{k},0}-\gamma_{n_{k},0}\right)=o_{p}\left(n_{k}^{-1/2}\right). (65) By substituting ( 17 ) in ( 64 ) together with ( 65 ) it follows that \displaystyle\hat{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0} \displaystyle= \displaystyle\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}+\hat{A}_{n}\left(\tilde{\gamma}_{n_{k},0}-\gamma_{n_{k},0}\right) (66) \displaystyle+\hat{\Omega}_{n_{k},0}^{-1/2}\left(I_{d}\otimes\hat{g}_{n_{k},0}^{\prime}\hat{\Omega}_{n_{k},0}^{-1}\right)\hat{\Lambda}_{n,0}\left(\tilde{\gamma}{}_{n_{k},0}-\gamma_{n_{k},0}\right)+o_{p}\left(n_{k}^{-1/2}\right). The last term on the RHS is o_{p}\left(n_{k}^{-1/2}\right) along all converging subsequences n_{k} by Lemma 3.1 (i) and Assumptions 3 and 6 . Now substitute ( 22 ) in ( 66 ) such that \displaystyle\hat{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0} \displaystyle= \displaystyle\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}+\hat{A}_{n_{k}}\left(\tilde{\gamma}{}_{n_{k},0}-\gamma_{n_{k},0}\right)+o_{p}\left(n_{k}^{-1/2}\right) (67) \displaystyle= \displaystyle\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}-\hat{A}_{n_{k}}\mathring{\Gamma}_{n_{k}}^{+}\hat{A}_{n_{k}}^{\prime}\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}+o_{p}\left(n_{k}^{-1/2}\right) \displaystyle= \displaystyle\left(I-P_{\hat{A}_{n_{k}}}\right)\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}+\left(P_{\hat{A}_{n_{k}}}-\hat{A}_{n_{k}}\mathring{\Gamma}_{n_{k}}^{+}\hat{A}_{n_{k}}^{\prime}\right)\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}+o_{p}\left(n_{k}^{-1/2}\right) \displaystyle= \displaystyle\left(I-P_{\hat{A}_{n_{k}}}\right)\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}+o_{p}\left(n_{k}^{-1/2}\right) where the second equality uses ( 22 ). The last equality follows from Lemma 3.1 (iii) and \hat{g}_{n_{k},0}=O_{p}\left(n_{k}^{-1/2}\right) . Next note that by Assumption 4 \hat{g}_{n_{k},\gamma}\left(\check{\gamma}_{n_{k},0}\right)=O_{p}\left(1\right) . By Lemma 3.1 (i) and ( 19 ) it then follows that \tilde{g}_{n_{k},0}-\hat{g}_{n_{k},0}=O_{p}\left(n_{k}^{-1/2}\right) . By Assumption 3 and Lemma 3.1 (i) it follows that \hat{\Omega}_{n_{k},0}^{-1/2}-\tilde{\Omega}_{n_{k},0}^{-1/2}=o_{p}\left(1\right) . Then, \displaystyle\hat{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0} \displaystyle= \displaystyle\tilde{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0}+\left(\hat{\Omega}_{n_{k},0}^{-1/2}-\tilde{\Omega}_{n_{k},0}^{-1/2}\right)\tilde{g}_{n_{k},0} (68) \displaystyle= \displaystyle\tilde{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0}+\left(\hat{\Omega}_{n_{k},0}^{-1/2}-\tilde{\Omega}_{n_{k},0}^{-1/2}\right)\left(\hat{g}_{n,0}+\hat{g}_{n,\gamma}\left(\check{\gamma}_{n,0}\right)\left(\tilde{\gamma}{}_{n,0}-\gamma_{n,0}\right)\right) \displaystyle= \displaystyle\tilde{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0}+o_{p}\left(n_{k}^{-1/2}\right). By substituting the LHS in ( 67 ) for the RHS in ( 68 ) one obtains \tilde{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0}=\left(I-P_{\hat{A}_{n}}\right)\hat{\Omega}_{n,0}^{-1/2}\hat{g}_{n,0}+o_{p}\left(n_{k}^{-1/2}\right). (69) This establishes part (iii) and ( 61 ). The claim in (ii) and ( 60 ) follows from Lemma 3.1 (ii) because \hat{A}_{n_{k}}^{\prime}\tilde{\Omega}_{n_{k}}^{-1/2}\tilde{g}_{n_{k},0}=o_{p}\left(n_{k}^{-1/2}\right) which implies that P_{\hat{A}_{n_{k}}}^{\prime}\tilde{\Omega}_{n_{k}}^{-1/2}\tilde{g}_{n_{k},0}=o_{p}\left(n_{k}^{-1/2}\right). Alternatively, one can argue that P_{\hat{A}_{n_{k}}}^{\prime}\tilde{\Omega}_{n_{k}}^{-1/2}\tilde{g}_{n_{k},0}=P_{\hat{A}_{n_{k}}}\left(I-P_{\hat{A}_{n_{k}}}\right)\hat{\Omega}_{n,0}^{-1/2}\hat{g}_{n,0}+o_{p}\left(n_{k}^{-1/2}\right)=o_{p}\left(n_{k}^{-1/2}\right) where the first equality uses ( 69 ) and the second equality uses the fact that P_{\hat{A}_{n}} is a projection. ∎ We employ a device that was developed by GKMC12 to control for possible rank deficiency of \Pi_{n,w} along converging subsequences, see in particular p.2662 of that paper. For a converging subsequence n_{k} define Q_{n_{k}}=n_{k}^{1/2}\Pi_{n_{k},\gamma} (70) and let the singular value decomposition of Q_{n_{k}} be (see Bhatia, 1996, p. 6) Q_{n_{k}}=R_{n_{k}}\Lambda_{n_{k}}T_{n_{k}}^{\prime} (71) where R_{n_{k}} is a d\times d unitary matrix, T_{n_{k}} is a d_{\gamma}\times d_{\gamma} unitary matrix and \Lambda_{n_{k}} is a d\times d_{\gamma} matrix with diagonal elements \lambda_{n_{k}j} equal to the positive square roots of the eigenvalues of Q_{n_{k}}^{\prime}Q_{n_{k}} . Consider \tilde{Q}_{n_{k}}=\Pi_{n_{k},\gamma}=R_{n_{k}}\tilde{\Lambda}_{n_{k}}T_{n_{k}}^{\prime} with \tilde{\Lambda}_{n_{k}}=n_{k}^{-1/2}\Lambda_{n_{k}} and elements \text{$\tilde{\lambda}_{n_{k}j}$}=n_{k}^{-1/2}\lambda_{n_{k}j}. By continuity, along converging subsequences n_{k} , it follows that \Pi_{n_{k},\gamma} converges to some matrix \Pi_{\gamma} that may depend on the subsequence n_{k} . Then, \tilde{Q}_{n_{k}} converges, to a matrix with possibly reduced rank \tilde{Q}=\Pi_{\gamma}=R\tilde{\Lambda}T^{\prime} . The singular value decomposition of \tilde{Q}_{n_{k}}^{\prime}\tilde{Q}_{n_{k}} is T_{n_{k}}\tilde{\Lambda}_{n_{k}}^{\prime}\tilde{\Lambda}_{n_{k}}T_{n_{k}}^{\prime} with eigenvalues \tilde{\lambda}_{n_{k}j}^{2} . By continuity of the eigenvalues, see ( 41 ), \tilde{\lambda}_{n_{k}j}^{2}\rightarrow\tilde{\lambda}_{j}^{2} where all \tilde{\lambda}_{j}^{2}\geq 0 are bounded and some may be zero. By consequence, \lambda_{n_{k}j}\rightarrow\text{$\lambda_{j}$} with some \lambda_{j}=\infty. Collect all eigenvalues \lambda_{j} in a corresponding matrix \Lambda. Assume without loss of generality, as in GKMC12, that for j\leq p and 0\leq p\leq d , the j -th diagonal element of \Lambda, \lambda_{j}<\infty, while \lambda_{j}=\infty for j>p. Again as in GKMC12, define a full rank diagonal d_{\gamma}\times d_{\gamma} matrix L_{n} with diagonal element \left[L_{n}\right]_{jj}=l_{n,j}=\lambda_{n,j}^{-1} where \left[L_{n}\right]_{jj}=\left\{\begin{array}[]{cc}\lambda_{n,j}^{-1}&\text{ if $j>p$ }\\ 1&\text{if $j\leq p$}\end{array}\right.. (72) The left and right singular vectors R_{n_{k}} and T_{n_{k}} of \tilde{Q}_{n_{k}} converge to R and T in the following sense. Let \mathscr{A}_{n}=\tilde{Q}_{n}\tilde{Q}_{n}^{\prime} , \mathscr{A}=\tilde{Q}\tilde{Q}^{\prime}, \mathscr{B}_{n}=\tilde{Q}_{n}^{\prime}\tilde{Q}_{n}, and \mathscr{B}=\tilde{Q}^{\prime}\tilde{Q}. Note that \mathscr{A}_{n},\mathscr{A},\mathscr{B}_{n} and \mathscr{B} are Hermitian and therefore normal matrices with eigenvectors R_{n}, R , T_{n} and T respectively. Let \hat{R}_{1,n}=P_{\mathscr{A}_{n}}\left(\mathscr{L}_{1}\right) be the projection matrix projecting onto the subspace spanned by the eigenvectors R_{n} of \mathscr{A}_{n} associated with eigenvalues who are elements of the set \mathscr{L}_{1}. Similarly, define \hat{R_{2}}=P_{\mathscr{A}}\left(\mathscr{L}_{2}\right) , \hat{T}_{1,n}=P_{\mathscr{B}_{n}}\left(\mathscr{L}_{1}\right) and \hat{T}_{2}=P_{\mathscr{B}}\left(\mathscr{L}_{2}\right). For n_{k} large enough, distinct eigenvalues of \tilde{Q}_{n_{k}} and \tilde{Q} can be placed into separated subsets. Then it follows from Theorem A.2 that eigenvectors R_{n_{k}} associated with a distinct eigenvalue \tilde{\lambda}_{n_{k}j} converge to the eigenvector in R associated with eigenvalue \tilde{\lambda}_{j} . This holds because in the limit \hat{R}_{1,n_{k}} is orthogonal to all other eigenvectors of R by ( 42 ). Another way to understand this result is to note that the composition of projections projecting onto orthogonal subspaces is zero, for example P_{\mathscr{A}}\left(\mathscr{L}_{1}\right)P_{\mathscr{A}}\left(\mathscr{L}_{2}\right)=0 . As \mathscr{A}_{n} approaches \mathscr{A} the same relationship holds approximately for \hat{R}_{1,n} and \hat{R_{2}} . For eigenvectors related to eigenvalues with multiplicities, the result shows that these eigenvectors span a space that is orthogonal to the space spanned by all eigenvectors in R not associated with that eigenvalue. The limits of these eigenvectors with multiple eigenvalues can be chosen to equal the corresponding eigenvectors in R. The next Lemma characterizes the upper bound in ( 59 ) and establishes that n^{1/2}\hat{A}_{n}T_{n}L_{n} , where T_{n} and L_{n} are defined in ( 71 ) and ( 72 ), converges to a full column rank matrix along converging subsequences n_{k} . Lemma A.2 . Let \hat{A}_{n} be defined as in ( 17 ). Then, for any converging subsequence n_{k} , there exist sequences of full rank matrices T_{n_{k}} and L_{n_{k}} of dimension d_{\gamma}\times d_{\gamma} and a matrix A with possibly random elements such that n_{k}^{1/2}\hat{A}_{n_{k}}T_{n_{k}}L_{n_{k}}\rightarrow_{d}A and where A has full column rank w.p.1. Proof. Note that statements made with subscript n hold for all n=1,... while statements made with subscript n_{k} only hold along converging subsequences n_{k} . Let \omega_{n,V}=n^{-1/2}\sum_{t=1}^{n}V_{n,t} and S_{n,V}=\operatorname{vec}\left(\omega_{n,V}\right) . Now use the definition of V_{n,t} in ( 11 ) and write n^{1/2}\operatorname{vec}\left(\hat{g}_{n,\gamma}\right)=n^{-1/2}\sum_{t=1}^{n}\operatorname{vec}\left(V_{n,t}+\Pi_{n,\gamma}\right)=\operatorname{vec}\left(n^{1/2}\Pi_{n,\gamma}+\omega_{n,V}\right). We then consider n_{k}^{1/2}\hat{A}_{n_{k}}T_{n_{k}}L_{n_{k}}=\hat{\Omega}_{n,0}^{-1/2}\left(\left(n^{1/2}\Pi_{n,\gamma}+\omega_{n,V}\right)-\left(I_{d}\otimes\hat{g}_{n,0}^{\prime}\hat{\Omega}_{n,0}^{-1}\right)\hat{\Lambda}_{n}\left(\gamma_{n,0}\right)\right)T_{n_{k}}L_{n_{k}}. (73) Also define S_{n,\varepsilon}=n^{-1/2}\sum_{t=1}^{n}\varepsilon_{n,t} and partition the limiting process S_{n}\rightarrow_{d}\omega as \omega=\left(\omega_{\varepsilon}^{\prime},\operatorname{vec}\omega_{V}^{\prime}\right)^{\prime} where \omega_{\varepsilon} is a d\times 1 vector of Gaussian random variables, \operatorname{vec}\omega_{V} is a dd_{\gamma} dimensional vector of Gaussian random variables and where \omega\sim N\left(0,\Sigma\right) by Assumption 6 . Write \Sigma=\left[\begin{array}[]{cc}\Sigma_{\varepsilon}&\Sigma_{\varepsilon V}\\ \Sigma_{V\varepsilon}&\Sigma_{V}\end{array}\right] (74) where the sub-matrices \Sigma_{\varepsilon}, \Sigma_{V\varepsilon}=\Sigma_{\varepsilon V}^{\prime} and \Sigma_{V} are partitioned according to \omega_{\varepsilon} and \omega_{V}. Using the definitions of T_{n_{k}} and L_{n_{k}} in ( 71 ) and ( 72 ) and for Q_{n_{k}} defined in ( 70 ) it follows that the first part of the RHS in ( 73 ) can be written as n_{k}^{1/2}\text{$\hat{g}_{n_{k},\gamma}$}T_{n_{k}}L_{n_{k}}=Q_{n_{k}}T_{n_{k}}L_{n_{k}}+\omega_{Vn_{k}}T_{n_{k}}L_{n_{k}}. (75) Then, along converging subsequences it follows from Assumption 4 , Theorem ( A.1 ), ( 11 ) and ( 42 ) that \text{$Q_{n_{k}}$}T_{n_{k}}L_{n_{k}}=R_{n_{k}}\Lambda_{n_{k}}L_{n_{k}}\rightarrow R\bar{L} where \bar{L} is a d_{\gamma}\times d_{\gamma} matrix with diagonal elements equal to \lambda_{j} for j\leq p and equal to one for j>p. By Assumption 6 , \omega_{Vn_{k}}T_{n_{k}}L_{n_{k}}\rightarrow_{d}\left(\omega_{V}T_{1},....,\omega_{V}T_{p},0...,0\right)=\bar{\omega}_{V} (76) where T_{j} is the j -th column of T and \operatorname{vec}\left(\omega_{V}\right)\sim N\left(0,\Sigma_{V}\right) such that \operatorname{vec}\left(\omega_{V}T_{1},....,\omega_{V}T_{p}\right)\sim N\left(0,\Sigma_{V,pp}\right) where \Sigma_{V,pp} is the upper corner p\times p block of \Sigma_{V} because T_{j} are orthonormal vectors. It now follows that n_{k}^{1/2}\text{$\hat{g}_{n_{k},\gamma}$}T_{n_{k}}L_{n_{k}}\rightarrow_{d}R\bar{L}+\bar{\omega}_{V} (77) where the RHS is full column rank with probability 1 as in GKMC12. Next consider the limit of \hat{\Lambda}_{n,0}=\hat{\Lambda}_{n}\left(\gamma\right) defined in ( 13 ). Recall that \hat{\Lambda}_{n,0}=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}k\left(\frac{t-s}{a_{n}}\right)\left(g_{s}^{\gamma}\left(\gamma_{n,0}\right)\otimes\varepsilon_{n,t}\right) where g_{t}^{\gamma}\left(\gamma_{n,0}\right)=V_{n,t}+\Pi_{n,\gamma} from ( 11 ). Note that \left(\Pi_{n,\gamma}\otimes\varepsilon_{n,t}\right)=\left(I_{d}\otimes\varepsilon_{n,t}\right)\Pi_{n,\gamma} such that \hat{\Lambda}_{n,0} can be decomposed into \hat{\Lambda}_{n,0}=\hat{\Lambda}_{\gamma,n,0}\Pi_{n,\gamma}+\hat{\Lambda}_{V,n,0} (78) with \hat{\Lambda}_{\gamma,n,0}=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}k\left(\frac{t-s}{a_{n}}\right)\left(I_{d}\otimes\varepsilon_{n,t}\right) (79) and \hat{\Lambda}_{V,n,0}=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}k\left(\frac{t-s}{a_{n}}\right)\left(V_{n,s}\otimes\varepsilon_{n,t}\right). (80) Now consider \displaystyle n_{k}^{1/2}\left(I_{d}\otimes\hat{g}_{n_{k},0}^{\prime}\hat{\Omega}_{n_{k},0}^{-1}\right)\hat{\Lambda}_{\gamma,n_{k},0}\Pi_{n_{k},\gamma}T_{n_{k}}L_{n_{k}} \displaystyle=\left[\left(I_{d}\otimes\hat{g}_{n_{k},0}^{\prime}\hat{\Omega}_{n_{k},0}^{-1}\right)\hat{\Lambda}_{\gamma,n,0}\right]\text{$Q_{n_{k},zz}$}T_{n_{k}}L_{n_{k}} (81) \displaystyle=o_{p}\left(1\right). where the n_{k}^{1/2} term on the LHS of ( 81 ) is absorbed in Q_{n_{k},zz} which was defined in ( 70 ). To see that the RHS of ( 81 ) is o_{p}\left(1\right) , note that \hat{\Omega}_{n_{k},0}^{-1}\rightarrow_{p}\Omega_{0}^{-1} where \Omega_{0} is full rank by Assumption 3 , and by Assumption 5 , \hat{\Lambda}_{\gamma,n_{k},0}\rightarrow_{p}\Lambda_{\gamma,0} (82) for some matrix \Lambda_{\gamma,0} with fixed and finite elements, and \hat{g}_{n_{k},0}^{\prime}\hat{\Omega}_{n_{k},0}^{-1}=o_{p}\left(1\right) by Assumption 6 . In addition Q_{n_{k},zz}T_{n_{k}}L_{n_{k}}=O_{p}\left(1\right) by a previous argument. This establishes that ( 81 ) is o_{p}\left(1\right). For \Sigma_{\varepsilon V} and \Sigma_{\varepsilon} defined in ( 74 ), note that \Sigma_{\varepsilon V}=\lim_{k\rightarrow\infty}n_{k}^{-1}\sum_{s=1}^{n_{k}}\sum_{t=1}^{n_{k}}E\left[\varepsilon_{n_{k},t}\operatorname{vec}\left(V_{n_{k},s}\right)^{\prime}\right] such that \operatorname{vec}\left(\Sigma_{\varepsilon V}\right)=\lim_{k\rightarrow\infty}n_{k}^{-1}\sum_{s=1}^{n_{k}}\sum_{t=1}^{n_{k}}E\left[\left(\operatorname{vec}\left(V_{n_{k},s}\right)\otimes\varepsilon_{n_{k},t}\right)\right]. Similarly, note that \operatorname{vec}\left(V_{n,s}\otimes\varepsilon_{n,t}\right)=\left(\operatorname{vec}\left(V_{n,s}\right)\otimes\varepsilon_{n,t}\right) and \Lambda_{V,0}=\lim_{k\rightarrow\infty}n_{k}^{-1}\sum_{s=1}^{n_{k}}\sum_{t=1}^{n_{k}}E\left[V_{n,s}\otimes\varepsilon_{n,t}\right]. Write the elements of \Lambda_{V,0} as \Lambda_{V,0}=\left(\begin{array}[]{ccc}\Lambda_{V,0}^{11}&\cdots&\Lambda_{V,0}^{1d_{\gamma}}\\ \vdots&\ddots&\vdots\\ \Lambda_{V,0}^{d1}&\cdots&\Lambda_{V,0}^{dd_{\gamma}}\end{array}\right) (83) where \Lambda_{V,0}^{ij} are vectors of dimension d. This implies that the term \hat{\Lambda}_{V,n_{k},0} in ( 80 ) satisfies \operatorname{vec}\left(\hat{\Lambda}_{V,n_{k},0}\right)\rightarrow_{p}\operatorname{vec}\left(\Lambda_{V,0}\right)=\operatorname{vec}\left(\Sigma_{\varepsilon V}\right) (84) and where the convergence follows by Assumptions 5 and 6 . Also note that Assumptions 3 and 6 lead to n_{k}^{1/2}\hat{g}_{n_{k},0}^{\prime}\hat{\Omega}_{n_{k},0}^{-1}\rightarrow_{d}\omega_{\varepsilon}^{\prime}\Sigma_{\varepsilon}^{-1}. (85) Now consider \displaystyle n_{k}^{1/2}\left(I_{d}\otimes\hat{g}_{n_{k},0}^{\prime}\hat{\Omega}_{n_{k},0}^{-1}\right)\hat{\Lambda}_{V,n_{k},0} \displaystyle\rightarrow_{d}\left(I_{d}\otimes\omega_{\varepsilon}^{\prime}\Sigma_{\varepsilon}^{-1}\right)\Lambda_{V,0}. (86) To represent the limit ( 86 ) let \omega_{V\varepsilon}=\left(I_{d}\otimes\omega_{\varepsilon}^{\prime}\Sigma_{\varepsilon}^{-1}\right)\Lambda_{V,0}. Then, using ( 83 ) \left(I_{d}\otimes\omega_{\varepsilon}^{\prime}\Sigma_{\varepsilon}^{-1}\right)\Lambda_{V,0}=\left(\begin{array}[]{ccc}\omega_{\varepsilon}^{\prime}\Sigma_{\varepsilon}^{-1}\Lambda_{V,0}^{11}&\cdots&\omega_{\varepsilon}^{\prime}\Sigma_{\varepsilon}^{-1}\Lambda_{V,0}^{1d_{\gamma}}\\ \vdots&\ddots&\vdots\\ \omega_{\varepsilon}^{\prime}\Sigma_{\varepsilon}^{-1}\Lambda_{Vz,0}^{d1}&\cdots&\omega_{\varepsilon}^{\prime}\Sigma_{\varepsilon}^{-1}\Lambda_{V,0}^{dd_{\gamma}}\end{array}\right) where \omega_{\varepsilon}^{\prime}\Sigma_{\varepsilon}^{-1}\Lambda_{V,0}^{ij} are scalars such that E\left[\omega_{\varepsilon}\operatorname{vec}\left(\omega_{V\varepsilon}\right)^{\prime}\right]=\left(E\left[\omega_{\varepsilon}\omega_{\varepsilon}^{\prime}\Sigma_{\varepsilon}^{-1}\Lambda_{V,0}^{11}\right],...,E\left[\omega_{\varepsilon}\omega_{\varepsilon}^{\prime}\Sigma_{\varepsilon}^{-1}\Lambda_{V,0}^{d1}\right],....,E\left[\omega_{\varepsilon}\omega_{\varepsilon}^{\prime}\Sigma_{\varepsilon}^{-1}\Lambda_{V,0}^{dd_{\gamma}}\right]\right)=\Sigma_{\varepsilon V} (87) Collect terms in 73 using( 75 ), ( 78 ) and ( 81 ), such that \displaystyle n_{k}^{1/2}\hat{A}_{n_{k}}T_{n_{k}}L_{n_{k}} \displaystyle=\hat{\Omega}_{n_{k},0}^{-1/2}\left(\left(n_{k}^{1/2}\Pi_{n_{k},\gamma}+\omega_{n_{k},V}\right)-\left(I_{d}\otimes n_{k}^{1/2}\hat{g}_{n_{k},0}^{\prime}\hat{\Omega}_{n_{k},0}^{-1}\right)\hat{\Lambda}_{V,n_{k},0}\right)T_{n_{k}}L_{n_{k}}+o_{p}\left(1\right). It now follows from ( 76 ) and ( 86 ) that \left(\omega_{n_{k},V}-\left(I_{d}\otimes n_{n_{k}}^{1/2}\hat{g}_{n,0}^{\prime}\hat{\Omega}_{n,0}^{-1}\right)\hat{\Lambda}_{V,n_{k},0}\right)T_{n_{k}}L_{n_{k}}\rightarrow_{d}\bar{\omega}_{V}-\overline{\omega_{V\varepsilon}} (88) and where \overline{\omega_{V\varepsilon}}=\left(\omega_{V\varepsilon}T_{1},...,\omega_{V\varepsilon}T_{p},0,...0\right) . The column vectors T_{1},....,T_{p} are as in ( 76 ). Introduce the notation \omega_{V.\varepsilon}=\omega_{V}-\omega_{V\varepsilon} and \overline{\omega_{V.\varepsilon}}=\bar{\omega}_{V}-\overline{\omega_{V\varepsilon}} . It follows from Assumption 6 that \omega_{\varepsilon} and \operatorname{vec}\left(\omega_{V}\right) are jointly Gaussian, and therefore that \omega_{\varepsilon} and \operatorname{vec}\left(\omega_{V.\varepsilon}\right) are joint Gaussian. It follows from ( 87 ) that E\left[\omega_{\varepsilon}\operatorname{vec}\left(\omega_{V.\varepsilon}\right)^{\prime}\right]=E\left[\omega_{\varepsilon}\operatorname{vec}\left(\omega_{V}-\omega_{V\varepsilon}\right)^{\prime}\right]=\Sigma_{\varepsilon V}-E\left[\omega_{\varepsilon}\operatorname{vec}\left(\omega_{V\varepsilon}\right)^{\prime}\right]=0. (89) This establishes that \omega_{V.\varepsilon} and \omega_{\varepsilon} are independent. Combining ( 77 ), ( 82 ), ( 84 ), ( 85 ) and ( 88 ) then leads to n_{k}^{1/2}\hat{A}_{n_{k}}T_{n_{k}}L_{n_{k}}\rightarrow_{d}A=\Omega_{0}^{-1/2}\left(R\bar{L}+\overline{\omega_{V.\varepsilon}}\right) (90) where A has full rank with probability 1. To see this, note that \bar{L} is a diagonal matrix with the first p elements equal to the eigenvalues \lambda_{j} and the remaining elements equal to one. Thus, the space spanned by R\bar{L} consists of the last d_{\gamma}-p columns of R and possibly some of the first p. Since R is unitary, this space has rank d_{\gamma}-p+p^{*} where 0\leq p^{*}\leq p is the number of non-zero eigenvalues \lambda_{j} for 0\leq j\leq p. In addition, \omega_{V.\varepsilon} is a set of p Gaussian vectors of dimension d_{\gamma} . Because \Sigma_{V.\varepsilon}=E\left[\operatorname{vec}\omega_{V.\varepsilon}\operatorname{vec}\left(\omega_{V.\varepsilon}\right)^{\prime}\right] is non-singular, \omega_{V.\varepsilon}T_{1},...,\omega_{V.\varepsilon}T_{p} is a full column rank matrix w.p.1. Then, \left(R\bar{L}+\overline{\omega_{V.\varepsilon}}\right)=\left(\omega_{V.\varepsilon}T_{1}+R_{1}\lambda_{11},...,\omega_{V.\varepsilon}T_{p}+R_{p}\lambda_{pp},R_{p+1},...,R_{d_{\gamma}}\right) (91) which is full column rank w.p.1. ∎ Lemma A.3 . Let \hat{A}_{n} be defined as in 17 . Then, for any converging subsequence n_{k} it follows that n_{k}\hat{g}_{n_{k},0}^{\prime}\hat{\Omega}_{n_{k},0}^{-1/2}M_{\hat{A}_{n_{k}}}\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}\rightarrow_{d}\chi_{d-d_{\gamma}}^{2}. Proof. First note that by Lemma A.1 (iii) along converging subsequences n_{k}\tilde{g}_{n_{k},0}^{\prime}\tilde{\Omega}_{n_{k},0}^{-1/2}M_{\hat{A}_{n_{k}}}\tilde{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0}=n_{k}\hat{g}_{n_{k},0}^{\prime}\hat{\Omega}_{n_{k},0}^{-1/2}M_{\hat{A}_{n_{k}}}\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}+o_{p}\left(1\right). (92) Note that by Assumption 6 and for S_{n} partitioned as S_{n}=\left[S^{\prime}_{n,\varepsilon},S^{\prime}_{n,V}\right]^{\prime} where S_{n,\varepsilon}=n^{-1/2}\sum_{i=1}^{n}\varepsilon_{n,t}=\hat{g}_{n,0} and S_{n,V}=n^{-1/2}\sum_{i=1}^{n}\operatorname{vec}\left(V_{n,t}\right) it follows that \Sigma_{n_{k}}^{-1/2}S_{n}\rightarrow_{d}\Sigma^{-1/2}\omega\sim N\left(0,I\right) . The limiting random variable \omega can be partitioned conformingly as \omega=\left(\omega_{\varepsilon}^{\prime},\operatorname{vec}\omega_{V}^{\prime}\right)^{\prime} . By Assumption 6 the matrix \Sigma_{n_{k}} converges to a limit, denoted by \Sigma which can be partitioned conformingly with \omega_{\varepsilon} as in ( 74 ) such that \Sigma_{\varepsilon} is equal to \Omega_{0} . As in GKMC12, we note that P_{\hat{A}_{n_{k}}} is invariant to the scaling of \hat{A}_{n_{k}} by n_{k}^{1/2} as well as to the rotation of the column space of \hat{A}_{n_{k}} by a full rank matrix T_{n_{k}}L_{n_{k}}. By ( 90 ) it follows that P_{\hat{A}_{n_{k}}}\rightarrow_{d}P_{A} by the continuous mapping theorem. To see this note that P_{A} is continuous in A as long as A has full column rank, Stewart 1977 (1977, Theorem 2.2). Since A is full column rank with probability one, van der Vaart and Wellner 1996 (1996, Theorem 1.3.6) can be applied. This argument shows that M_{\hat{A}_{n_{k}}}=I-P_{\hat{A}_{n_{k}}}\rightarrow_{d}I-P_{A}=M_{A}. In addition, \omega_{\varepsilon} is independent of A by ( 88 ), ( 89 ), ( 90 ) and ( 91 ). It now follows from ( 92 ), Assumption 6 and the continuous mapping theorem that n_{k}\hat{g}_{n_{k},0}^{\prime}\hat{\Omega}_{n_{k},0}^{-1/2}M_{\hat{A}_{n_{k}}}\hat{\Omega}_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}\rightarrow_{d}\omega_{\varepsilon}^{\prime}\Sigma_{\varepsilon}^{-1/2}M_{A}\Sigma_{\varepsilon}^{-1/2}\omega_{\varepsilon} where, conditionally on A, \Sigma_{\varepsilon}^{-1/2}\omega_{\varepsilon}\sim N\left(0,I_{d}\right). Since this distribution does not depend on A, it follows by the same arguments as in GKMC12 and the fact that M_{A} is projection matrix of rank d-d_{\gamma} that the distribution is \omega_{\varepsilon}^{\prime}\Sigma_{\varepsilon}^{-1/2}M_{A}\Sigma_{\varepsilon}^{-1/2}\omega_{\varepsilon}\sim\chi_{d-d_{\gamma}}^{2}. ∎ We now present the proof of the main theorem. Proof of Theorem 3.1 . The proof of the Theorem follows arguments in GKMC12 and Andrews et al. 2020 with the necessary adjustments. Let \chi_{n}^{*} be a double array that solves \sup_{\chi_{n}\in\mathscr{X}}P\left(\text{$\textrm{AR}_{\textrm{C}}\left(\beta_{n,0}\right)$}>c_{1-\alpha,\chi_{d-d_{\gamma}}^{2}}\right) (93) and denote by P_{\chi_{n}^{*}} the measure induced by \chi_{n}^{*} . Then, \displaystyle\textrm{AsySz}_{\alpha} \displaystyle= \displaystyle\limsup_{n\rightarrow\infty}P_{\chi_{n}^{*}}\left(\text{$\textrm{AR}_{\textrm{C}}\left(\beta_{n,0}\right)$}>c_{1-\alpha,\chi_{d-d_{\gamma}}^{2}}\right). (94) Now let n_{k} be a converging subsequence such that \limsup_{n\rightarrow\infty}P_{\chi_{n}^{*}}\left(\text{$\textrm{AR}_{\textrm{C}}\left(\beta_{n,0}\right)$}>c_{1-\alpha,\chi_{d-d_{\gamma}}^{2}}\right)=\lim_{k\rightarrow\infty}P_{\chi_{n_{k}}^{*}}\left(\text{$\textrm{AR}_{\textrm{C}}\left(\beta_{n_{k},0}\right)$}>c_{1-\alpha,\chi_{d-d_{\gamma}}^{2}}\right). (95) Such a subsequence n_{k} exists by the properties of ’ \limsup ’ and Assumption 1 . By Lemma A.1 (i) and ( 59 ) the following inequality holds for all converging subsequences n_{k} \textrm{AR}_{\textrm{C}}\left(\beta_{n_{k},0}\right)\leq n_{k}\tilde{g}_{n_{k},0}^{\prime}\tilde{\Omega}_{n_{k},0}^{-1/2}M_{\hat{A}_{n_{k}}}\tilde{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0}+\varpi_{n_{k}}. (96) Combining ( 94 ), ( 95 ) and ( 96 ) gives \textrm{AsySz}_{\alpha}\leq\lim_{k\rightarrow\infty}P_{\chi_{n_{k}}^{*}}\left(n_{k}\tilde{g}_{n_{k},0}^{\prime}\tilde{\Omega}_{n_{k},0}^{-1/2}M_{\hat{A}_{n_{k}}}\tilde{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0}+\varpi_{n_{k}}>c_{1-\alpha,\chi_{d-d_{\gamma}}^{2}}\right). (97) By Lemmas A.1 , A.2 and A.3 and for any converging subsequence n_{k} , it follows that n\tilde{g}_{n_{k},0}^{\prime}\tilde{\Omega}_{n_{k},0}^{-1/2}M_{\hat{A}_{n_{k}}}\tilde{\Omega}_{n_{k},0}^{-1/2}\tilde{g}_{n_{k},0}+\varpi_{n_{k}}\rightarrow_{d}\chi_{d-d_{\gamma}}^{2}. (98) Then, ( 97 ) and ( 98 ) imply that \textrm{AsySz}_{\alpha}\leq\alpha. We show strict equality by following the argument in GKMC12. When the CUE estimator \hat{\gamma}_{n} is identified it follows that \hat{\gamma}_{n_{k}}-\gamma_{n_{k},0}=O_{p}\left(n_{k}^{-1/2}\right). Then, the arguments in Lemmas A.1 , A.2 and A.3 imply that \textrm{AR}_{\textrm{C}}\left(\beta_{n_{k},0}\right)=n_{k}\hat{g}_{n_{k},0}^{\prime}\hat{\Omega}_{n_{k},0}^{-1/2}M_{\hat{A}_{n_{k}}}\Omega_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}+o_{p}\left(1\right) and n\hat{g}_{n_{k},0}^{\prime}\hat{\Omega}_{n_{k},0}^{-1/2}M_{\hat{A}_{n_{k}}}\Omega_{n_{k},0}^{-1/2}\hat{g}_{n_{k},0}\rightarrow_{d}\chi_{d-d_{\gamma}}^{2} along all converging subsequences where \gamma is identified. This implies that the size of \textrm{AR}_{\textrm{C}}\left(\beta_{n_{k},0}\right) is equal to \alpha in the identified case and it follows that \textrm{AsySz}_{\alpha}=\alpha. ∎ A.5 Derivation of CUE First Order Condition Consider the criterion Q_{n}\left(\beta,\gamma\right)=n\hat{g}_{n}\left(\beta,\gamma\right)^{\prime}\hat{\Omega}_{n}^{-1}\left(\beta,\gamma\right)\hat{g}_{n}\left(\beta,\gamma\right) evaluated at Q_{n}\left(\beta_{n,0},\gamma\right) . We evaluate the first order condition for the minimization problem \min_{\gamma}Q_{n}\left(\beta_{n,0},\gamma\right) . Following Magnus and Neudecker 1988 , p. 183 as well as the discussion in Section 13, we start by computing total derivatives of Q_{n}\left(\beta_{n,0},\gamma\right). We use the shorthand notation \hat{g}_{n}^{\prime}\Omega_{n}^{-1}\hat{g}_{n} when no confusion arises. Using d as the differential operator We have \displaystyle d\left(\hat{g}_{n}^{\prime}\Omega_{n}^{-1}\hat{g}_{n}\right) \displaystyle=\left(d\hat{g}_{n}\right)^{\prime}\Omega_{n}^{-1}\hat{g}_{n}+\hat{g}_{n}^{\prime}\Omega_{n}^{-1}\left(d\hat{g}_{n}\right)-\hat{g}_{n}^{\prime}\Omega_{n}^{-1}d\Omega_{n}\Omega_{n}^{-1}\hat{g}_{n} \displaystyle=2\hat{g}_{n}^{\prime}\Omega_{n}^{-1}\left(d\hat{g}_{n}\right)-\hat{g}_{n}^{\prime}\Omega_{n}^{-1}d\Omega_{n}\Omega_{n}^{-1}\hat{g}_{n}. Since the LHS is a scalar it follows that \displaystyle d\left(\hat{g}_{n}^{\prime}\Omega_{n}^{-1}\hat{g}_{n}\right) \displaystyle=\left(2d\hat{g}_{n}-\hat{g}_{n}^{\prime}\Omega_{n}^{-1}d\Omega_{n}\right)^{\prime}\Omega_{n}^{-1}\hat{g}_{n} \displaystyle=\left(\operatorname{vec}\left(2d\hat{g}_{n}-\hat{g}_{n}^{\prime}\Omega_{n}^{-1}d\Omega_{n}\right)\right)^{\prime}\Omega_{n}^{-1}\hat{g}_{n} where the second quality follows because d\hat{g}_{n} , \hat{g}_{n}^{\prime}\Omega_{n}^{-1}d\Omega_{n} , and \Omega_{n}^{-1}\hat{g}_{n} are vectors. Then, it is easy to see that d\left(\hat{g}_{n}^{\prime}\Omega_{n}^{-1}\hat{g}_{n}\right)=\hat{g}_{n}^{\prime}\Omega_{n}^{-1}\left(2d\hat{g}_{n}-\left(I_{d}\otimes\hat{g}_{n}^{\prime}\Omega_{n}^{-1}\right)d\operatorname{vec}\left(\Omega_{n}\right)\right). Following Magnus and Neudecker 1988 , p. 173, Definition 1, we obtain the vector of partial derivatives as \frac{\partial\hat{g}_{n}^{\prime}\Omega_{n}^{-1}\hat{g}_{n}}{\partial\gamma^{\prime}}=\hat{g}_{n}^{\prime}\Omega_{n}^{-1}\left(2\frac{\partial\hat{g}_{n}}{\partial\gamma^{\prime}}-\left(I_{d}\otimes\hat{g}_{n}^{\prime}\Omega_{n}^{-1}\right)\frac{\partial\operatorname{vec}\left(\Omega_{n}\right)}{\partial\gamma^{\prime}}\right) where \partial\hat{g}_{n}/\partial\gamma^{\prime} is a d\times d_{\gamma} dimensional matrix and \partial\operatorname{vec}\left(\Omega_{n}\right)/\partial\gamma^{\prime} is a d^{2}\times d_{\gamma} dimensional matrix. Finally, consider the term d\Omega_{n} which can be evaluated as \displaystyle d\operatorname{vec}\Omega_{n} \displaystyle=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}k\left(\frac{t-s}{a_{n}}\right)\left\{g_{s}\left(\beta,\gamma\right)\otimes\left(dg_{t}\left(\beta,\gamma\right)\right)+\left(dg_{s}\left(\beta,\gamma\right)\right)\otimes g_{t}\left(\beta,\gamma\right)\right\} \displaystyle=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}k\left(\frac{t-s}{a_{n}}\right)\left\{\left(g_{s}\left(\beta,\gamma\right)\otimes I_{d}\right)dg_{t}\left(\beta,\gamma\right)+\left(I_{d}\otimes g_{t}\left(\beta,\gamma\right)\right)dg_{s}\left(\beta,\gamma\right)\right\} where the second equality follows from Magnus and Neudecker 1988 , p. 185, (11) and therefore the derivative can be written as \displaystyle\frac{\partial\operatorname{vec}\Omega_{n}}{\partial\gamma^{\prime}} \displaystyle=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}k\left(\frac{t-s}{a_{n}}\right)\left\{\left(g_{s}\left(\beta,\gamma\right)\otimes I_{d}\right)\frac{\partial g_{t}\left(\beta,\gamma\right)}{\partial\gamma^{\prime}}+\left(I_{d}\otimes g_{t}\left(\beta,\gamma\right)\right)\frac{\partial g_{s}\left(\beta,\gamma\right)}{\partial\gamma^{\prime}}\right\} \displaystyle:=\hat{\Lambda}_{n}^{1}+\hat{\Lambda}_{n}^{2} However, note that the derivative in \hat{g}_{n}^{\prime}\Omega_{n}^{-1}d\Omega_{n}\Omega_{n}^{-1}\hat{g}_{n} satisfies the property \displaystyle d\Omega_{n} \displaystyle=n^{-1}\sum_{t=1}^{n}\sum_{s=1}^{n}k\left(\frac{t-s}{a_{n}}\right)\left(\left(dg_{s}\left(\beta,\gamma\right)\right)g_{t}\left(\beta,\gamma\right)^{\prime}+g_{s}\left(\beta,\gamma\right)\left(dg_{t}\left(\beta,\gamma\right)\right)^{\prime}\right) \displaystyle:=\tilde{d\Omega_{n}}+\tilde{d\Omega_{n}}^{\prime} and \displaystyle\hat{g}_{n}^{\prime}\Omega_{n}^{-1}d\Omega_{n}\Omega_{n}^{-1}\hat{g}_{n} \displaystyle=\hat{g}_{n}^{\prime}\Omega_{n}^{-1}\left(\tilde{d\Omega_{n}}+\tilde{d\Omega_{n}}^{\prime}\right)\Omega_{n}^{-1}\hat{g}_{n} \displaystyle=2\hat{g}_{n}^{\prime}\Omega_{n}^{-1}\tilde{d\Omega_{n}}\Omega_{n}^{-1}\hat{g}_{n} because the criterion function is a scalar which is invariant to transposes. The implication is that we can focus \frac{\partial\hat{g}_{n}^{\prime}\Omega_{n}^{-1}\hat{g}_{n}}{\partial\gamma^{\prime}}=2\hat{g}_{n}^{\prime}\Omega_{n}^{-1}\left(\frac{\partial\hat{g}_{n}}{\partial\gamma^{\prime}}-\left(I_{d}\otimes\hat{g}_{n}^{\prime}\Omega_{n}^{-1}\right)\hat{\Lambda}_{n}\right) and the factor 2 can be dropped when evaluating first order conditions. 10 12 12 26