Abstract
This paper is concerned with the multivariate 3rd moment and its estimation. Mardia (Biometrika 57:519–530, 1970) and Srivastava (Stat Probab Lett 2:263–267, 1984) proposed the multivariate skewness and its estimator, independently. However, these estimators cannot be defined for the case in which the dimension p is larger than the sample size N. In this paper, we treat the multivariate 3rd moment \(\gamma \) which is defined by using Hadamard product of observation vectors, and propose an estimate of \(\gamma \) which is well defined when \(p>N\). Based on the estimator, we propose a new test for multivariate normality. Under the null hypothesis, the test statistic is asymptotically standard normal, which is supported by Monte Carlo simulations. We calculate some empirical powers to see the performance of the test.
Similar content being viewed by others
References
Chen SX, Zhang LX, Zhong PS (2010) Tests for high-dimensional covariance matrices. J Am Stat Assoc 105:810–819
Doornik JA, Hansen H (2008) An omnibus test for univariate and multivariate normality. Oxf Bull Econ Stat 70:927–939
Dudoit S, Fridlyand J, Speed TP (2002) Comparison of discrimination methods for the classification of tumors using gene expression data. J Am Stat Assoc 97:77–87
Fujikoshi Y, Ulyanov VV, Shimizu R (2010) Multivariate statistics high-dimensional and large-sample approximations. Wiley, Hoboken
Henze N (2002) Invariant tests for multivariate normality: a critical review. Stat Pap 43:467–506
Himeno T, Yamada T (2014) Estimations for some functions of covariance matrix in high dimension under non-normality and its applications. J Multivariate Anal 130:27–44
Holgersson HET (2006) A graphical method for assessing multivariate normality. Comput Stat 21:141–149
Holgersson HET, Mansoor R (2013) Assessing normality of high-dimensional data. Commun Stat Simul Comput 42:360–369
Jarque CM, Bera AK (1987) A test for normality of observations and regression residuals. Int Stat Rev 55:163–172
Kankainen A, Taskinen S, Oja H (2007) Tests of multinormality based on location vectors and scatter matrices. Stat Methods Appl 16:357–379
Kubokawa T, Srivastava MS (2008) Estimation of the precision matrix of a singular Wishart distribution and its application in high-dimensional data. J Multivariate Anal 99:1906–1928
Mansoor R (2017) Using principal components to test normality of high-dimensional data. Commun Stat Simul Comput 46:3396–3405
Mardia KV (1970) Measures of multivariate skewness and kurtosis with applications. Biometrika 57:519–530
Mecklin CJ, Mundfrom DJ (2004) An appraisal and bibliography of tests for multivariate normality. Int Stat Rev 72:123–138
Schott JR (2005) Matrix analysis for statistics, 2nd edn. Wiley, Hoboken
Srivastava MS (1984) A measure of skewness and kurtosis and a graphical method for assessing multivariate normality. Stat Probab Lett 2:263–267
Srivastava MS (2002) Methods of multivariate statistics. Wiley, New York
Thulin M (2014) Tests for multivariate normality based on canonical correlations. Stat Methods Appl 23:189–208
Acknowledgements
We would like to thank an associate editor and two referees for constructive comments.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Takayuki Yamada is partially supported by Ministry of Education, Science, Sports, and Culture, a Grant-in-Aid for Scientific Research (Wakate B), 26800088, 2014–2017. Tetsuto Himeno is partially supported by Ministry of Education, Science, Sports, and Culture, a Grant-in-Aid for Scientific Research (Wakate B), 16K16018, 2016–2018.
Appendices
Moment of statistic
In this section, analytic forms of \(\mathrm {Var}(T_1)\), \(\mathrm {Var}(T_2)\) and \(\mathrm {Var}(T_3)\) are proposed. We derive them by using the following lemma (Lemma 1). Proofs of Lemma 1 is tedious but simple, therefore skipped.
Lemma 1
Let \(\varvec{z}\) be a random vector distributed as \(N_p(\varvec{0},\varvec{I}_p)\), and \(\varvec{a}\) and \(\varvec{b}\) be constant vectors. Then,
Here, we write analytic forms of \(\mathrm {Var}(T_1)\), \(\mathrm {Var}(T_2)\) and \(\mathrm {Var}(T_3)\) in the following lemma.
Lemma 2
For \(T_1\), \(T_2\) and \(T_3\), as defined in Sect. 2.1,
Proof
We can express that
where
Since \(\zeta _1,\ldots ,\zeta _N\) are i.i.d., it holds that
where \(\varvec{z} \sim N_p(\varvec{0},\varvec{I}_p)\). We find that
and so
where the third equality follows from Lemma 1.
Since \(E[T_2]=0\), we have
where the fourth equality follows from Lemma 1.
Since \(E[T_3]=0\), we have
where the fourth equality follows from Lemma 1. \(\square \)
Derivation of the unbiasedness for T and V
In this section, we show that T and V, which are defined by (5) and (6), are unbiased estimators of \(E[\{\odot ^3 (\varvec{x}-\varvec{\mu })\}'\varvec{1}_p]\) and \(\varvec{1}_p'(\odot ^3 \varvec{{\varSigma }})\varvec{1}_p\), respectively.
Lemma 3
Assume that \(\varvec{x}_1\), ..., \(\varvec{x}_N\) are i.i.d., and assume that these random vectors have the multivariate linear model (3). Then T and V are unbiased estimators of \(E[\{\odot ^3 (\varvec{x}-\varvec{\mu })\}'\varvec{1}_p]\) and \(\varvec{1}_p'(\odot ^3 \varvec{{\varSigma }})\varvec{1}_p\), respectively.
Proof
If k, \(\ell \) and \(\alpha \) are all different, then it is easy to see that \(E[t_{k \ell \alpha ,i}^3]=(3/4)E[(x_i-\mu _i)^3]\) for \((t_{k \ell \alpha ,1},\ldots ,t_{k \ell \alpha ,p})'=\varvec{t}_{k \ell \alpha }\), and for \(\begin{pmatrix} x_1 \ldots x_p \end{pmatrix}'=\varvec{x}=\varvec{\mu }+\varvec{{\varSigma }}^{1/2}\varvec{z}\) with \(\varvec{z} \sim F_p(\varvec{0},\varvec{I}_p)\). Adding up over the \(\mathrm {P}_{N,3}\) such triples, multiplying by 4 / 3 and dividing \(\mathrm {P}_{N,3}\) gives an unbiased estimate of \(E[(x_i-\mu _i)^3]\). By adding up it over i, we obtain T. It is also easy to see that \(E[v_{k \ell ,i j} v_{\alpha \beta ,i j} v_{\gamma \delta ,i j} ]=\sigma _{ij}^3\) for \((v_{k \ell ,i j})=\varvec{V}_{k \ell }\), which is defined in (6), and \((\sigma _{ij})=\varvec{{\varSigma }}\) if k, \(\ell \), \(\alpha \), \(\beta \), \(\gamma \) and \(\delta \) are all different. Adding up over the \(\mathrm {P}_{N,6}\) such sextuples and dividing \(\mathrm {P}_{N,6}\) gives an unbiased estimate of \(\sigma _{ij}^3\). By adding up it over (i, j), we obtain V. \(\square \)
Proof of Theorem 2
In this section, we give a proof of Theorem 2. Before proving it, we prepare three lemmas. The first two lemmas (Lemma 4 and Lemma 5) treat formula about multi-sum, and the last lemma (Lemma 6) treats the uniformly boundedness on \(p \in {\mathbb {N}}\) for the expectation, which is used to prove Theorem 2. Since proofs of Lemma 4 and Lemma 5 are tedious but simple, we skipped here.
Lemma 4
The following equations hold:
Lemma 5
The following equations hold:
Lemma 6
Let
where \(\varvec{{\varSigma }}^{1/2}=\varvec{A}=(\varvec{a}_1,\ldots ,\varvec{a}_p)'\), and \(\varvec{z}_k\), \(\varvec{z}_{\ell }\), \(\varvec{z}_{\alpha }\), \(\varvec{z}_{\beta }\), \(\varvec{z}_{\gamma }\) and \(\varvec{z}_{\delta }\) are i.i.d. as \(N_p(\varvec{0},\varvec{I}_p)\). Then, there exist \(M>0\), which does not depend on p, such that
In addition, under the assumption \(\mathrm {C}\),
Proof
Since (19), (20) and (21) can be proved by using the same way to show (18), we only show (18), and omit other three here. It can be expressed that
where the second equality follows from Lemma 1, the third equality from bottom follows from Lemma 4 and the second equality from bottom follows from Lemma 5. Since \(\odot ^2 \varvec{{\varSigma }}\) is positive definite, we have
where the right-hand side of the inequality can be written as \(\{\varvec{1}_p'(\odot ^3 \varvec{{\varSigma }})\varvec{1}_p\}^2\). From this result, it is found that
Note that
From this inequality, we have
Hence under the assumption \(\mathrm {C3}\), there exist \(M>0\), which does not depend on p, such that
and so
\(\square \)
Proof of Theorem 2
Since V is unbiased estimator of \(\varvec{1}_p'(\odot ^3 \varvec{{\varSigma }})\varvec{1}_p\), it is sufficient to show that
as \(N \rightarrow \infty \). Since V is invariant under location shift, without loss of generality, we may assume that \(\varvec{\mu }=\varvec{0}\). Each of A, B, C and D in (16) can be written as
From (16), we can express that
where the inequality follows from Jensen inequality. Thus,
For the first term in the right-hand side of the inequality (23), expanding \(\{.\}^2\), and excluding the term whose expectation becomes 0, we have
where
Using the same calculation method, the second-fourth terms in the right-hand side of the inequality (23) can be written as
where
From Lemma 6, there exist \(M>0\), which does not depend on p, such that
It follows from Cauchy–Schwarz inequality that
From them, we find that there exist \(M>0\), which does not depend on p, such that
Combining these results with (24), (25), (26) and (27), it is found that
and so
which converges to 0 as \(N \rightarrow \infty \). \(\square \)
Rights and permissions
About this article
Cite this article
Yamada, T., Himeno, T. Estimation of multivariate 3rd moment for high-dimensional data and its application for testing multivariate normality. Comput Stat 34, 911–941 (2019). https://doi.org/10.1007/s00180-018-00865-9
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00180-018-00865-9