Abstract
This paper studies maximum likelihood estimation of autoregressive models of order 1 with a near unit root and Cauchy errors. Autoregressive models with an intercept and with an intercept and a linear time trend are also considered. The maximum likelihood estimator (MLE) for the autoregressive coefficient is \(n^{3/2}\)-consistent with n denoting the sample size and has a mixture-normal distribution in the limit. The MLE for the scale parameter of Cauchy distribution is \(n^{1/2}\)-consistent, and its limiting distribution is normal. The MLEs of the intercept and the linear time trend are \(n^{1/2}\)- and \(n^{3/2}\)-consistent, respectively. It is also shown that the t statistic for the null hypothesis of a unit root based on the MLE has a standard normal distribution in the limit. In addition, finite-sample properties of the MLE are compared with those of the least square estimator (LSE). It is found that the MLE is more efficient than the LSE when the errors have a Cauchy distribution or a distribution which is a mixture of Cauchy and normal distributions. It is also shown that empirical power of the MLE-based t test for a unit root is much higher than that of the Dickey–Fuller t test.
Similar content being viewed by others
References
Ahn, S. K., Fotopoulos, S. B., He, L. (2001). Unit root tests with infinite variance errors. Econometric Reviews, 20(4), 461–483.
Akgiray, V., Booth, G. G. (1988). The stable-law model of stock returns. Journal of Business & Economic Statistics, 6(1), 51–57.
Andrews, B., Calder, M., Davis, R. A. (2009). Maximum likelihood estimation for \(\alpha \)-stable autoregressive processes. The Annals of Statistics, 37(4), 1946–1982.
Arnold, L. (1974). Stochastic differential equations: Theory and applications. New York: Wiley.
Callegari, F., Cappuccio, N., Lubian, D. (2003). Asymptotic inference in time series regressions with a unit root and infinite variance errors. Journal of Statistical Planning and Inference, 116(1), 277–303.
Chan, N. H., Tran, L. T. (1989). On the first-order autoregressive process with infinite variance. Econometric Theory, 5(03), 354–362.
Chan, N. H., Peng, L., Qi, Y. (2006). Quantile inference for near-integrated autoregressive time series with infinite variance. Statistica Sinica, 16(1), 15.
Choi, I. (2015). Almost all about unit roots: Foundations, developments, and applications. Cambridge: Cambridge University Press.
Dickey, D. A., Fuller, W. A. (1979). Distribution of the estimators for autoregressive time series with a unit root. Journal of the American statistical association, 74(366a), 427–431.
DuMouchel, W. H. (1973). On the asymptotic normality of the maximum-likelihood estimate when sampling from a stable distribution. The Annals of Statistics, 1(5), 948–957.
Falk, B., Wang, C. H. (2003). Testing long-run PPP with infinite-variance returns. Journal of Applied Econometrics, 18(4), 471–484.
Fama, E. F. (1965). The behavior of stock-market prices. The Journal of Business, 38(1), 34–105.
Knight, K. (1989). Limit theory for autoregressive-parameter estimates in an infinite-variance random walk. Canadian Journal of Statistics, 17(3), 261–278.
Knight, K. (1991). Limit theory for M-estimates in an integrated infinite variance. Econometric Theory, 7(02), 200–212.
Koedijk, K. G., Kool, C. J. (1992). Tail estimates of East European exchange rates. Journal of Business & Economic Statistics, 10(1), 83–96.
Lau, A. H. L., Lau, H. S., Wingender, J. R. (1990). The distribution of stock returns: New evidence against the stable model. Journal of Business & Economic Statistics, 8(2), 217–223.
Mandelbrot, B. (1963). The Variation of Certain Speculative Prices. The Journal of Business, 36(4), 394–419.
Mandelbrot, B. (1967). The variation of some other speculative prices. The Journal of Business, 40(4), 393–413.
Newey, W. K., McFadden, D. (1994). Large sample estimation and hypothesis testing, Handbook of econometrics (Vol. 4, pp. 2111–2245). Amsterdam: Elsevier.
Park, J. Y., Phillips, P. C. (1988). Statistical inference in regressions with integrated processes: Part 1. Econometric Theory, 4(03), 468–497.
Phillips, P. C. (1990). Time series regression with a unit root and infinite-variance errors. Econometric Theory, 6(01), 44–62.
Phillips, P. C. (2014). On confidence intervals for autoregressive roots and predictive regression. Econometrica, 82(3), 1177–1195.
Pollard, D. (1991). Asymptotics for least absolute deviation regression estimators. Econometric Theory, 7(02), 186–199.
Resnick, S., Greenwood, P. (1979). A bivariate stable characterization and domains of attraction. Journal of Multivariate Analysis, 9(2), 206–221.
Samarakoon, D. M., Knight, K. (2009). A note on unit root tests with infinite variance noise. Econometric Reviews, 28(4), 314–334.
Stock, J. H. (1991). Confidence intervals for the largest autoregressive root in U.S. macroeconomic time series. Journal of Monetary Economics, 28(3), 435–459.
Zhang, R. M., Chan, N. H. (2012). Maximum likelihood estimation for nearly non-stationary stable autoregressive processes. Journal of Time Series Analysis, 33(4), 542–553.
Author information
Authors and Affiliations
Corresponding author
Additional information
This paper stemmed from the MA thesis of the first author which was supervised by the second author.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Appendices
Appendix A: Proofs
The following lemma will be used to prove Lemmas 2, 3, 5 which are used for the proofs of all the theorems of this paper.
Lemma 1
Define \(Z_{n} (r) =n^{ -1} \sum _{t =1}^{[n r]}\varepsilon _{t} ,\;B_{n} (r) =n^{ -1/2} \sum _{t =1}^{[n r]}\frac{ \partial \ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}}\;\)and \(Q_{n} (r) =n^{ -1/2} \sum _{t =1}^{[n r]}\frac{ \partial \ln f (\varepsilon _{t} ,\sigma )}{ \partial \sigma }_{ \mid \sigma =\sigma ^{o}}\) for \(0 \le r \le 1.\) Then, \((Z_{n} (r_{1})\text {,}\)\(B_{n} (r_{2})\text {,}\)\(Q_{n} (r_{3}))\overset{d}{ \rightarrow }(Z (r_{1})\text {,}\)\(B (r_{2})\text {,}\)\(Q (r_{3}))\) in \(D [0 ,1]^{3}\text {,}\) where\(\;Z (r_{1})\text {,}\)\(B (r_{2})\,\)and \(Q (r_{3})\) are defined in Theorem 1.
Proof
This lemma is similar to Lemma 1 of Zhang and Chan (2012) except that the limit results for \(B_{n} (r_{2})\) and \(Q_{n} (r_{3})\) under the assumption of Cauchy errors are not dealt with there. Thus, we only need to show marginal weak convergence results for \(B_{n} (r_{2})\) and \(Q_{n} (r_{3})\) under the assumption of Cauchy errors. Because \(E \genfrac(){}{}{ \partial \ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}} =0\) by standard theory and \(E \genfrac(){}{}{ \partial \ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}}^{2} =\frac{1}{2 \sigma ^{o 2}} <\infty \) by Lemma 7, we can apply the classical functional central limit theorem for the sequence of i.i.d. random variables obtaining
Likewise, we have
Since \(E \genfrac(){}{}{ \partial \ln f (\varepsilon _{i} ,\sigma ^{o})}{ \partial \varepsilon _{i}} \left( \frac{ \partial \ln f (\varepsilon _{j} ,\sigma )}{ \partial \sigma }_{ \mid \sigma =\sigma ^{o}}\right) =0\) for every i and \(j ,\;B (r)\) and\(\;Q (s)\) are independent for \(0 \le r \le 1\) and \(0 \le s \le 1.\)\(\square \)
The following lemma will be used to prove Theorem 1.
Lemma 2
-
(i)
\(\;n^{ -3/2} \sum _{t =1}^{n}\frac{ \partial \ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}} Y_{t -1}\overset{d}{ \rightarrow } \smallint _{0}^{1}S (r) \mathrm{d} B (r) ;\)
-
(ii)
\(n^{ -2} \sum _{t =1}^{n}Y_{t -1}\overset{d}{ \rightarrow } \smallint _{0}^{1}S (r) \mathrm{d} r ;\)
-
(iii)
\(n^{ -3} \sum _{t =1}^{n}Y_{t -1}^{2}\overset{d}{ \rightarrow } \smallint _{0}^{1}S^{2} (r) \mathrm{d} r ;\)
-
(iv)
\(n^{ -3} \sum _{t =1}^{n}Y_{t -1}^{2} \left[ \frac{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}^{2}} -E \genfrac(){}{}{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}^{2}}\right] \overset{p}{ \longrightarrow }0 ;\)
-
(v)
\(n^{ -2} \sum _{t =1}^{n}Y_{t -1} \left[ \frac{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma )}{ \partial \varepsilon _{t} \partial \sigma }_{ \mid \sigma =\sigma ^{o}} -E \left( \frac{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma )}{ \partial \varepsilon _{t} \partial \sigma }_{ \mid \sigma =\sigma ^{o}}\right) \right] \overset{p}{ \longrightarrow }0 ;\)
-
(vi)
\(n^{ -1} \sum _{t =1}^{n}\left[ \frac{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma )}{ \partial \sigma ^{2}}_{ \mid \sigma =\sigma ^{o}} -E \left( \frac{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma )}{ \partial \sigma ^{2}}_{ \mid \sigma =\sigma ^{o}}\right) \right] \overset{p}{ \longrightarrow }0.\)
Proof
Because \(S_{n} (r)\overset{d}{ \rightarrow }S (r)\) in \(D [0 ,1]\text {,}\) where \(S_{n} (r) =n^{ -1} Y_{[n r]}\) and S(r) is defined in Theorem 1, as shown in Chan et al. (2006), parts (i), (ii) and (iii) follow as in Lemma 4 of Zhang and Chan (2012). Because the variances of and \(\left\{ \frac{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma )}{ \partial \varepsilon _{t} \partial \sigma }_{ \mid \sigma =\sigma ^{o}}\right\} \) are finite due to Lemma 10, we obtain (iv) and (v) by using the same method as for equation (16) of Zhang and Chan (2012). Part (vi) is trivial. \(\square \)
Proof of Theorem 1
-
(i)
We need to check the four conditions of Newey and McFadden (1994) Theorem 2.5. Conditions (i), (ii) and (iii) are trivially satisfied. To check condition (iv), write
$$\begin{aligned} f_{t} (\rho _{n} ,\sigma )= & {} \frac{\sigma }{\pi \{\sigma ^{2} +(Y_{t} -\rho _{n} Y_{t -1})^{2}\}} \\= & {} \frac{\sigma }{\pi \{\sigma ^{2} +\varepsilon _{t}^{2} +2 (\rho _{n}^{o} -\rho _{n}) \varepsilon _{t} Y_{t -1} +(\rho _{n}^{o} -\rho _{n})^{2} Y_{t -1}^{2}\}}\text {.} \end{aligned}$$Suppose that \(Y_{t -1}\) is given. Then, as in Newey and McFadden (1994; p.2125),
$$\begin{aligned} \left| \ln f_{t} (\rho _{n} ,\sigma )\right|\le & {} \ln \sigma -\ln \pi -\ln \{\sigma ^{2} +\varepsilon _{t}^{2} +2 (\rho _{n}^{o} -\rho _{n}) \varepsilon _{t} Y_{t -1} +(\rho _{n}^{o} -\rho _{n})^{2} Y_{t -1}^{2}\} \\\le & {} C_{1} +\ln (C_{2} +C_{3} \varepsilon _{t}^{2})\end{aligned}$$for some positive constants \(C_{1} ,\;C_{2}\) and \(C_{3} .\;\)Since \(E \left[ \ln (C_{2} +C_{3} \varepsilon _{t}^{2})\right] <\infty \;\)for every \(t ,\;\)we obtain
$$\begin{aligned}E \left( \underset{\rho _{n} ,\sigma }{\sup }\left| \ln f_{t} (\rho _{n} ,\sigma )\right| \right) =E E \left( \underset{\rho _{n} ,\sigma }{\sup }\left| \ln f_{t} (\rho _{n} ,\sigma )\right| \mid Y_{t -1}\right) <\infty \text {,} \end{aligned}$$as desired.
-
(ii)
Let \(\theta _{n} =(\rho _{n} ,\sigma )\text {,}\)\(\hat{\theta }_{n} =(\hat{\rho }_{n} ,\hat{\sigma }_{n})\text {,}\)\(\theta _{n}^{o} =(\rho _{n}^{o} ,\sigma ^{o})\) and \(\theta _{n}^{ *}\) be on the line joining \(\theta _{n}^{o}\) and \(\hat{\theta }_{n} .\;\)Because \(\sum _{t =1}^{n}\frac{ \partial \ln f}{ \partial \theta _{n}}_{ \mid _{\theta _{n} =\widehat{\theta }_{n}}} =0\text {,}\) we obtain by the mean value theorem
$$\begin{aligned}0 =\sum _{t =1}^{n}\frac{ \partial \ln f}{ \partial \theta _{n}}_{ \mid _{\theta _{n} =\theta _{n}^{o}}} +\sum _{t =1}^{n}\frac{ \partial ^{2}\ln f}{ \partial \theta _{n} \partial \theta _{n}^{ \prime }}_{ \mid _{\theta _{n} =\theta _{n}^{ *}}} \left( \begin{array}{l}\hat{\rho }_{n} -\rho _{n}^{o} \\ \hat{\sigma }_{n} -\sigma ^{o}\end{array}\right) \text {,} \end{aligned}$$
which gives
where \(J_{n} =d i a g (n^{3/2} ,n^{1/2}) .\;\)Because \(\theta _{n}^{ *} -\theta _{n}^{o}\overset{p}{ \longrightarrow }0\) as \(n \rightarrow \infty ,\;\)it follows that
Since
parts (iv), (v) and (vi) of Lemma 2 show that \(A_{n \mid _{\theta _{n} =\theta _{n}^{o}}}\) has the same limiting distribution as
which is equal to, due to Lemmas 8 and 9,
Thus, using relation (8) and part (iii) of Lemma 2, we obtain
In addition, Lemmas 1 and 2 yield
The stated result follows, once the continuous mapping theorem is applied to relations (9) and (10). \(\square \)
It is straightforward to show that \(S_{n} (r) =n^{ -1} Y_{[n r]}\overset{d}{ \rightarrow }S (r)\) for model (3). Thus, we continue to use Lemma 2 to prove Theorem 2. In addition to Lemma 2, we need the following lemma to prove Theorem 2.
Lemma 3
-
(i)
\(n^{ -2} \sum _{t =1}^{n}Y_{t -1} \left[ \frac{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}^{2}} -E \genfrac(){}{}{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}^{2}}\right] \overset{p}{ \longrightarrow }0 ;\)
-
(ii)
\(n^{ -1} \sum _{t =1}^{n}\left[ \frac{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma )}{ \partial \varepsilon _{t} \partial \sigma }_{ \mid \sigma =\sigma ^{o}} -E \left( \frac{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma )}{ \partial \varepsilon _{t} \partial \sigma }_{ \mid \sigma =\sigma ^{o}}\right) \right] \overset{p}{ \longrightarrow }0 ;\)
-
(iii)
\(n^{ -1} \sum _{t =1}^{n}\left[ \frac{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}^{2}} -E \genfrac(){}{}{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}^{2}}\right] \overset{p}{ \longrightarrow }0.\)
Proof
Using the same method as for Eq. (16) of Zhang and Chan (2012), we obtain (i). Parts (ii) and (iii) are trivial. \(\square \)
Proof of Theorem 2
-
(i)
Use the same method as for the proof of Theorem 1 (i).
-
(ii)
Let \(\theta _{n} =(\rho _{n} ,\mu ,\sigma )\text {,}\)\(\hat{\theta }_{n} =(\hat{\rho }_{n} ,\hat{\mu }_{n} ,\hat{\sigma }_{n})\text {,}\)\(\theta _{n}^{o} =(\rho _{n}^{o} ,\mu ^{o} ,\sigma ^{o})\) and \(\theta _{n}^{ *}\) be on the line joining \(\theta _{n}^{o}\) and \(\hat{\theta }_{n}\text {.}\) Because \(\sum _{t =1}^{n}\frac{ \partial \ln f}{ \partial \theta _{n}}_{ \mid _{\theta _{n} =\widehat{\theta }_{n}}} =0\text {,}\) the mean value theorem yields
$$\begin{aligned}0 =\sum _{t =1}^{n}\frac{ \partial \ln f}{ \partial \theta _{n}}_{ \mid _{\theta _{n} =\theta _{n}^{o}}} +\sum _{t =1}^{n}\frac{ \partial ^{2}\ln f}{ \partial \theta _{n} \partial \theta _{n}^{ \prime }}_{ \mid _{\theta _{n} =\theta _{n}^{ *}}} \left( \begin{array}{l}\hat{\rho }_{n} -\rho _{n}^{o} \\ \hat{\mu }_{n} -\mu ^{o} \\ \hat{\sigma }_{n} -\sigma ^{o}\end{array}\right) \text {,} \end{aligned}$$
which gives
where \(K_{n} =d i a g (n^{3/2} ,n^{1/2} ,n^{1/2}) .\;\)Because \(\theta _{n}^{ *} -\theta _{n}^{o}\overset{p}{ \longrightarrow }0\) as \(n \rightarrow \infty ,\;\)it follows that
Since
parts (iv), (v) and (vi) of Lemma 2 and parts (i), (ii) and (iii) of Lemma 3 show that \(E_{n \mid _{\theta _{n} =\theta _{n}^{o}}}\) has the same limiting distribution as
which is equal to, due to Lemmas 8 and 9,
Thus, using relation (11) and parts (ii) and (iii) of Lemma 2, we obtain
In addition, Lemmas 1 and 2 yield
The stated result is obtained by applying the continuous mapping theorem to relations (12) and (13). \(\square \)
The following two lemmas will be used to prove Theorem 3.
Lemma 4
For model (5), we have \(S_{n} (r) =n^{ -1} Y_{[n r]}\overset{d}{ \rightarrow }U (r) =S (r) +b_{0} r\) in \(D [0 ,1]\text {.}\)
Proof
Because \(X_{t} =\rho _{n} X_{t -1} +\varepsilon _{t}\text {,}\)\(n^{ -1} X_{[n r]}\overset{d}{ \rightarrow }S (r)\) in D[0, 1], we obtain
as stated. \(\square \)
Lemma 5
-
(i)
\(\;n^{ -3/2} \sum _{t =1}^{n}\frac{ \partial \ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}} Y_{t -1}\overset{d}{ \rightarrow } \smallint _{0}^{1}U (r) \mathrm{d} B (r) ;\)
-
(ii)
\(n^{ -2} \sum _{t =1}^{n}Y_{t -1}\overset{d}{ \rightarrow } \smallint _{0}^{1}U (r) \mathrm{d} r ;\)
-
(iii)
\(n^{ -3} \sum _{t =1}^{n}Y_{t -1}^{2}\overset{d}{ \rightarrow } \smallint _{0}^{1}U^{2} (r) \mathrm{d} r ;\)
-
(iv)
\(n^{ -3} \sum _{t =1}^{n}t Y_{t -1}\overset{d}{ \rightarrow } \smallint _{0}^{1}r U (r) \mathrm{d} r ;\)
-
(v)
\(n^{ -3/2} \sum _{t =1}^{n}t \frac{ \partial \ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}}\overset{d}{ \rightarrow } \smallint _{0}^{1}r \mathrm{d} B (r) ;\)
-
(vi)
\(n^{ -3} \sum _{t =1}^{n}t Y_{t -1} \left[ \frac{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}^{2}} -E \genfrac(){}{}{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}^{2}}\right] \overset{p}{ \longrightarrow }0 ;\)
-
(vii)
\(n^{ -2} \sum _{t =1}^{n}t \left[ \frac{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma )}{ \partial \varepsilon _{t} \partial \sigma }_{ \mid \sigma =\sigma ^{o}} -E \left( \frac{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma )}{ \partial \varepsilon _{t} \partial \sigma }_{ \mid \sigma =\sigma ^{o}}\right) \right] \overset{p}{ \longrightarrow }0 ;\)
-
(viii)
\(n^{ -2} \sum _{t =1}^{n}t \left[ \frac{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}^{2}} -E \genfrac(){}{}{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}^{2}}\right] \overset{p}{ \longrightarrow }0 ;\)
-
(ix)
\(n^{ -3} \sum _{t =1}^{n}t^{2} \left[ \frac{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}^{2}} -E \genfrac(){}{}{ \partial ^{2}\ln f (\varepsilon _{t} ,\sigma ^{o})}{ \partial \varepsilon _{t}^{2}}\right] \overset{p}{ \longrightarrow }0.\)
Proof
Because \(S_{n} (r)\overset{d}{ \rightarrow }U (r)\) in D[0, 1] by Lemma 4, parts (i), (ii), (iii) and (iv) follow as in Lemma 4 of Zhang and Chan (2012). Part (v) follows from the central limit theorem. Using the same method as for Eq. (16) of Zhang and Chan (2012), we obtain (vi), (vii), (viii) and (ix). \(\square \)
Proof of Theorem 3
-
(i)
Using the same method as for the proof of Theorem 1 (i), it is straightforward to prove this.
-
(ii)
Let \(\theta _{n} =(\rho _{n} ,\mu ,\beta ,\sigma )\text {,}\)\(\hat{\theta }_{n} =(\hat{\rho }_{n} ,\hat{\mu }_{n} ,\hat{\beta }_{n\text {,}} \hat{\sigma }_{n})\text {,}\)\(\theta _{n}^{o} =(\rho _{n}^{o} ,\mu ^{o} ,\beta ^{o} ,\sigma ^{o})\) and \(\theta _{n}^{ *}\) be on the line joining \(\theta _{n}^{o}\) and \(\hat{\theta }_{n}\text {.}\) Because \(\sum _{t =1}^{n}\frac{ \partial \ln f}{ \partial \theta _{n}}_{ \mid _{\theta _{n} =\widehat{\theta }_{n}}} =0\text {,}\) we obtain by the mean value theorem
$$\begin{aligned}0 =\sum _{t =1}^{n}\frac{ \partial \ln f}{ \partial \theta _{n}}_{ \mid _{\theta _{n} =\theta _{n}^{o}}} +\sum _{t =1}^{n}\frac{ \partial ^{2}\ln f}{ \partial \theta _{n} \partial \theta _{n}^{ \prime }}_{ \mid _{\theta _{n} =\theta _{n}^{ *}}} \left( \begin{array}{l}\hat{\rho }_{n} -\rho _{n}^{o} \\ \hat{\mu }_{n} -\mu ^{o} \\ \hat{\beta }_{n} -\beta ^{o} \\ \hat{\sigma }_{n} -\sigma ^{o}\end{array}\right) \text {.} \end{aligned}$$
Thus, we have
where \(L_{n} =d i a g (n^{3/2} ,n^{1/2} ,n^{3/2} ,n^{1/2}) .\;\)Because \(\theta _{n}^{ *} -\theta _{n}^{o}\overset{p}{ \longrightarrow }0\) as \(n \rightarrow \infty ,\;\)it follows that
Since
Lemmas 2, 3 and 5 show that \(C_{n \mid _{\theta _{n} =\theta _{n}^{o}}}\) has the same limiting distribution as
where \(G =E \genfrac[]{}{}{ \partial ^{2}\ln f}{ \partial \varepsilon _{t}^{2}}\text {,}\)\(H =E \genfrac[]{}{}{ \partial ^{2}\ln f}{ \partial \varepsilon _{t} \partial \sigma }\) and \(J =E \genfrac[]{}{}{ \partial ^{2}\ln f}{ \partial \sigma ^{2}}\text {.}\) Due to Lemmas 8 and 9, the above matrix is equal to
Thus, using relation (14) and parts (ii), (iii) and (iv) of Lemma 5, we obtain
In addition, Lemmas 1 and 5 yield
The stated result follows from relations (15) and (16). \(\square \)
Proof of Theorem 4
This follows straightforwardly from the mixture normality results (2), (4) and (6), and the block-diagonal structure of the information matrices in the limit. \(\square \)
Appendix B: Auxiliary lemmas
Lemma 6
Proof
The first result follows because \( \smallint _{\mathbb {R}}\frac{1}{x +\varepsilon _{1}^{2}} \mathrm{d} \varepsilon _{1} =\pi x^{ -1/2} .\;\)The rest are obtained by successively differentiating both sides of this equation with respect to x and setting \(x =\sigma ^{o 2}\text {.}\) Differentiating within the integral sign is allowed, because for \(k \in \mathbb {N}\) and \(x \in \mathbb {R}^{ +}\text {,}\)\(\left| \frac{ \partial \left( x +\varepsilon _{1}^{2}\right) ^{ -k}}{ \partial x} \right| =\left| \frac{k}{\left( x +\varepsilon _{1}^{2}\right) ^{k +1}} \right| \le \left| \frac{k}{\varepsilon _{1}^{2 (k +1)}}\right| \) and \(\left| \frac{k}{\varepsilon _{1}^{2 (k +1)}}\right| \) is integrable. \(\square \)
Lemma 7
-
(i)
\(\;E \left[ \genfrac(){}{}{ \partial \ln f (\varepsilon _{1} ,\sigma ^{o})}{ \partial \varepsilon _{1}}^{2}\right] =\frac{1}{2 \sigma ^{o 2}} ;\)
-
(ii)
\(E \left[ \left( \frac{ \partial \ln f (\varepsilon _{1} ,\sigma )}{ \partial \sigma }_{ \mid \sigma =\sigma ^{o}}\right) ^{2}\right] =\frac{1}{2 \sigma ^{o 2}}\text {.}\)
Proof
Using Lemma 6, we obtain
where \(D_{2}\) and \(D_{3}\) are defined in Lemma 6. In the same manner, we have
\(\square \)
Lemma 8
\(E \left[ \frac{ \partial ^{2}\ln f (\varepsilon _{1} ,\sigma )}{ \partial \varepsilon _{1} \partial \sigma }_{ \mid \sigma =\sigma ^{o}}\right] =0.\)
Proof
Because
we have
Because the integrand is an odd function, the stated result follows. \(\square \)
Lemma 9
\(E \left[ \frac{ \partial ^{2}\ln f (\varepsilon _{1} ,\sigma )}{ \partial \sigma ^{2}}_{ \mid \sigma =\sigma ^{o}}\right] = -\frac{1}{2 \sigma ^{o 2}}\) and \(E \genfrac[]{}{}{ \partial ^{2}\ln f (\varepsilon _{1} ,\sigma ^{o})}{ \partial \varepsilon _{1}^{2}} = -\frac{1}{2 \sigma ^{o 2}}\text {.}\)
Proof
We begin with the identity
By Leibniz’s rule, taking the derivative of the both sides of the above equation with respect to \(\sigma \) results in
Differentiating this again, we obtain by Leibniz’s rule
Hence, by Lemma 7,
To prove the second result, consider the relation
which gives
as stated. \(\square \)
Lemma 10
\(E \left[ \genfrac(){}{}{ \partial ^{2}\ln f (\varepsilon _{1} ,\sigma ^{o})}{ \partial \varepsilon _{1}^{2}}^{2}\right] \text {,}\)\(E \left[ \genfrac(){}{}{ \partial ^{2}\ln f (\varepsilon _{1} ,\sigma )}{ \partial \sigma ^{2}}_{ \mid \sigma =\sigma ^{o}}^{2}\right] \) and \(E \left[ \genfrac(){}{}{ \partial ^{2}\ln f (\varepsilon _{1} ,\sigma )}{ \partial \varepsilon _{1} \partial \sigma }_{ \mid \sigma =\sigma ^{o}}^{2}\right] \) are finite.
Proof
First, we show that \(E \left[ \genfrac(){}{}{ \partial ^{2}\ln f (\varepsilon _{1} ,\sigma ^{o})}{ \partial \varepsilon _{1} \partial \sigma }_{ \mid \sigma =\sigma ^{o}}^{2}\right] =\frac{5}{8} \sigma ^{o -4}\text {.}\) By Lemma 6,
Second, \(E \left[ \genfrac(){}{}{ \partial ^{2}\ln f (\varepsilon _{1} ,\sigma ^{o})}{ \partial \varepsilon _{1}^{2}}^{2}\right] =\frac{7}{8} \sigma ^{o -4}\text {.}\) By Lemma 6,
To prove \(E \left[ \genfrac(){}{}{ \partial ^{2}\ln f (\varepsilon _{1} ,\sigma )}{ \partial \sigma ^{2}}_{ \mid \sigma =\sigma ^{o}}^{2}\right] =\frac{7}{8} \sigma ^{o -4}\text {,}\) we should note the following equations
By using these equations, we have
\(\square \)
About this article
Cite this article
Choi, J., Choi, I. Maximum likelihood estimation of autoregressive models with a near unit root and Cauchy errors. Ann Inst Stat Math 71, 1121–1142 (2019). https://doi.org/10.1007/s10463-018-0671-z
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10463-018-0671-z