Bernoulli

  • Bernoulli
  • Volume 20, Number 3 (2014), 1234-1259.

Limiting spectral distribution of sample autocovariance matrices

Anirban Basak, Arup Bose, and Sanchayan Sen

Full-text: Open access

Abstract

We show that the empirical spectral distribution (ESD) of the sample autocovariance matrix (ACVM) converges as the dimension increases, when the time series is a linear process with reasonable restriction on the coefficients. The limit does not depend on the distribution of the underlying driving i.i.d. sequence and its support is unbounded. This limit does not coincide with the spectral distribution of the theoretical ACVM. However, it does so if we consider a suitably tapered version of the sample ACVM. For banded sample ACVM the limit has unbounded support as long as the number of non-zero diagonals in proportion to the dimension of the matrix is bounded away from zero. If this ratio tends to zero, then the limit exists and again coincides with the spectral distribution of the theoretical ACVM. Finally, we also study the LSD of a naturally modified version of the ACVM which is not non-negative definite.

Article information

Source
Bernoulli, Volume 20, Number 3 (2014), 1234-1259.

Dates
First available in Project Euclid: 11 June 2014

Permanent link to this document
https://projecteuclid.org/euclid.bj/1402488939

Digital Object Identifier
doi:10.3150/13-BEJ520

Mathematical Reviews number (MathSciNet)
MR3217443

Zentralblatt MATH identifier
1327.60023

Keywords
autocovariance function autocovariance matrix banded and tapered autocovariance matrix linear process spectral distribution stationary process Toeplitz matrix

Citation

Basak, Anirban; Bose, Arup; Sen, Sanchayan. Limiting spectral distribution of sample autocovariance matrices. Bernoulli 20 (2014), no. 3, 1234--1259. doi:10.3150/13-BEJ520. https://projecteuclid.org/euclid.bj/1402488939


Export citation

References

  • [1] Arcones, M.A. (2000). Distributional limit theorems over a stationary Gaussian sequence of random vectors. Stochastic Process. Appl. 88 135–159.
  • [2] Bai, Z. and Zhou, W. (2008). Large sample covariance matrices without independence structures in columns. Statist. Sinica 18 425–442.
  • [3] Bai, Z.D. (1999). Methodologies in spectral analysis of large-dimensional random matrices, a review (with discussions). Statist. Sinica 9 611–677.
  • [4] Basak, A. (2009). Large dimensional random matrices. M. Stat. Project report, May 2009. Indian Statistical Institute.
  • [5] Basak, A., Bose, A. and Sen, S. (2011). Limiting spectral distribution of sample autocovariance matrices. Technical Report R11 2011. Stat-Math Unit, Indian Statistical Institute. Available at http://arxiv.org/pdf/1108.3147v1.pdf.
  • [6] Basak, A., Bose, A. and Sen, S. (2013). Supplement to “Limiting spectral distribution of sample autocovariance matrices.” DOI:10.3150/13-BEJ520SUPP.
  • [7] Bose, A., Gangopadhyay, S. and Sen, A. (2010). Limiting spectral distribution of $XX'$ matrices. Ann. Inst. Henri Poincaré Probab. Stat. 46 677–707.
  • [8] Bose, A. and Sen, A. (2008). Another look at the moment method for large dimensional random matrices. Electron. J. Probab. 13 588–628.
  • [9] Böttcher, A. and Silbermann, B. (1999). Introduction to Large Truncated Toeplitz Matrices. Universitext. New York: Springer.
  • [10] Bryc, W., Dembo, A. and Jiang, T. (2006). Spectral measure of large random Hankel, Markov and Toeplitz matrices. Ann. Probab. 34 1–38.
  • [11] Dudley, R.M. (2002). Real Analysis and Probability. Cambridge Studies in Advanced Mathematics 74. Cambridge: Cambridge Univ. Press. Revised reprint of the 1989 original.
  • [12] Giraitis, L., Robinson, P.M. and Surgailis, D. (2000). A model for long memory conditional heteroscedasticity. Ann. Appl. Probab. 10 1002–1024.
  • [13] Hammond, C. and Miller, S.J. (2005). Distribution of eigenvalues for the ensemble of real symmetric Toeplitz matrices. J. Theoret. Probab. 18 537–566.
  • [14] McMurry, T.L. and Politis, D.N. (2010). Banded and tapered estimates for autocovariance matrices and the linear process bootstrap. J. Time Series Anal. 31 471–482.
  • [15] Sen, A. (2006). Large dimensional random matrices. M. Stat. Project report, May 2006. Indian Statistical Institute.
  • [16] Sen, S. (2010). Limiting spectral distribution of random matrices. M. Stat. Project report, July 2010. Indian Statistical Institute.
  • [17] Wu, W.B. and Pourahmadi, M. (2009). Banding sample autocovariance matrices of stationary processes. Statist. Sinica 19 1755–1768.
  • [18] Xiao, H. and Wu, W.B. (2012). Covariance matrix estimation for stationary time series. Ann. Statist. 40 466–493.
  • [19] Yao, J. (2012). A note on a Marčenko–Pastur type theorem for time series. Statist. Probab. Lett. 82 22–28.

Supplemental materials

  • Supplementary material: Simulations. Recall that none of the LSDs have a nice description. Following the suggestion of one of the Referees, we have collected some simulation results in a supplementary file Basak, Bose and Sen [6]. The simulations are for the AR(1) and MA(1) models. These simulations provide evidence that the limits are indeed universal and exhibit some mass on the negative axis for the ESD (and hence the LSD) of $\Gamma_{n}^{*}(X)$. They also show how the LSD of type I banded $\Gamma_{n}(X)$ changes with the model as well as the value of the parameter $\alpha$. The unbounded nature of the LSD is also evident from these simulations. For the banded matrices, the simulations demonstrate that for small values of $\alpha$, the LSD of $\Sigma_{n}(X)$ and $\Gamma_{n}(X)$ are virtually indistinguishable for large $n$, confirming that thinly banded ACVMs are consistent for $\Sigma_{n}(X)$. As the value of $\alpha$ increases, the right tail of the LSD thickens, and the probability of being near zero decreases. In general, there may be considerable amount of mass in the negative axis. This mass reduces as the value of $\alpha$ decreases. The LSD of $\Gamma_{n}(X)$ varies as the parameter of the models change. For both AR(1) and MA(1) models, as $\theta$ increases from $0$, the tail thickens, and the mass near zero decreases. For the AR(1) model, when $\theta$ approaches $1$, that is, when the process is near non-stationary the LSD becomes very flat, and its tail becomes huge.