Abstract
We investigate the limiting behavior of sample central moments, examining the special cases where the limiting (as the sample size tends to infinity) distribution is degenerate. Parent (non-degenerate) distributions with this property are called singular, and we show in this article that the singular distributions contain at most three supporting points. Moreover, using the delta-method, we show that the (second-order) limiting distribution of sample central moments from a singular distribution is either a multiple, or a difference of two multiples of independent Chi-square random variables with one degree of freedom. Finally, we present a new characterization of normality through the asymptotic independence of the sample mean and all sample central moments.
Similar content being viewed by others
References
Afendras, G. (2013). Moment-based inference for Pearson’s quadratic \(q\) subfamily of distributions. Communications in Statistics: Theory and Methods, 42(12), 2271–2280.
Afendras, G., Markatou, M. (2016). Uniform integrability of the OLS estimators, and the convergence of their moments. Test, 25(4), 775–784.
Billingsley, P. (1995). Probability and measure (3rd ed.). New York: Wiley.
Geary, R. C. (1936). The distribution of “Student’s” ratio for non-normal samples. Journal of the Royal Statistical Society Series B, 3, 178–184.
Gut, A. (1988). Stopped random walks: Limit theorems and applications. New York: Springer.
Haug, S., Klüppelberg, C., Lindner, A., Zapp, M. (2007). Method of moment estimation in the COGARCH(1,1) model. The Econometrics Journal, 10(2), 320–341.
Kagan, A. M., Linnik, Y. V., Rao, C. R. (1973). Characterization problems in mathematical statistics. New York: Wiley.
Kourouklis, S. (2012). A new estimator of the variance based on minimizing mean squared error. The American Statistician, 66(4), 234–236.
Laha, R. G., Lukacs, E., Newman, M. (1960). On the independence of a sample central moment and the sample mean. The Annals of Mathematical Statistics, 31, 1028–1033.
Lehmann, E. L. (1999). Elements of large-sample theory. New York: Springer.
Pewsey, A. (2005). The large-sample distribution of the most fundamental of statistical summaries. Journal of Statistical Planning and Inference, 134, 434–444.
Stefanski, L. A., Boos, D. D. (2002). The calculus of M-estimation. The American Statistician, 56(1), 29–38.
van der Vaart, A. W. (1998). Asymptotic statistics. New York: Cambridge University Press.
Wu, S.-F., Liang, M.-C. (2010). A note on the asymptotic distribution of the process capability index Cpkm. Journal of Statistical Computation and Simulation, 80, 227–235.
Yatracos, Y. (2005). Artificially augmented samples, shrinkage, and mean squared error reduction. Journal of the American Statistical Association, 100, 1168–1175.
Zinger, A. A. (1958). Independence of quasi-polynomial statistics and analytical properties of distributions. Theory of Probability & Its Applications, 3(3), 247–265.
Acknowledgements
We would like to thank H. Papageorgiou for helpful discussions.
Author information
Authors and Affiliations
Corresponding author
Appendix A Proofs
Appendix A Proofs
We shall make use of the following Lemmas. For the proof of Lemma 6 see, e.g., Gut (1988, p. 18); for more general results, see Afendras and Markatou (2016).
Lemma 6
If \(X,X_1,\ldots ,X_n\) are independent and identically distributed with \({\mathbb {E}}(X)=\mu \), \({\textsf {Var}}(X)=\sigma ^2\) and \({\mathbb {E}}|X|^\delta <\infty \) for some \(\delta \ge 2\), then, for any \(\alpha \in (0,\delta ]\),
where \(Z\sim N(0,1)\) and \(\bar{X}_n=(X_1+\cdots +X_n)/n\).
Lemma 7
If \(X,X_1,\ldots ,X_n\) are independent and identically distributed with \({\mathbb {E}}(X)=\mu \) and \({\mathbb {E}}|X|^\nu <\infty \) for some \(\nu \in \{2,3,\ldots \}\), then, for any \(j\in \{2,\ldots ,\nu \}\),
where \(m_{j,n}=n^{-1}\sum _{i=1}^n(X_i-\mu )^j\).
Proof
If \(j=\nu \), then (31) follows by taking expectations to the obvious inequality \(|m_{j,n}|\le \frac{1}{n}\sum _{i=1}^n|X_i-\mu |^j=\frac{1}{n} \sum _{i=1}^n|X_i-\mu |^{\nu }\). If \(j<\nu \) (and thus, \(\nu \ge 3\)), we apply the inequality
(the last inequality is a by-product of Hölder’s inequality) for \(p=\nu /j\) and \(x_i=(X_i-\mu )^j\). Then, we have
\(\square \)
Proof of Proposition 2
(a) Observe that the statement in Proposition 2(a) is equivalent to
Writing
it suffices to verify that
- (i)
\(\surd {n}{\mathbb {E}}(m_{k,n}-\mu _k)=0\),
- (ii)
\(\surd {n}{\mathbb {E}}(m_{1,n}^k)\rightarrow 0\),
- (iii)
\(\surd {n}{\mathbb {E}}(m_{1,n}^{k-j} m_{j,n}) \rightarrow 0\), \(j=2,\ldots ,k-2\) (provided \(k\ge 4\)), and
- (iv)
\(\surd {n}{\mathbb {E}}(m_{1,n} m_{k-1,n}) \rightarrow 0\) (provided \(k\ge 3\)).
Now, (i) is obvious (since \({\mathbb {E}}(m_{k,n})=\mu _k\)), (iv) follows from \({\mathbb {E}}(m_{1,n} m_{k-1,n})=\mu _k/n\) and (ii) can be seen by using Lemma 6 with \(\alpha =\delta =k\), which shows that
and thus, \(|\surd {n}{\mathbb {E}}(m_{1,n}^k)|\le n^{-(k-1)/2}{\mathbb {E}}|\surd {n}(\bar{X}_n-\mu )|^k\rightarrow 0\). To show (iii), we assume that \(k\ge 4\) and \(2\le j\le k-2\), and we use Hölder’s inequality with \(p=k/(k-j)>1\), Lemma 7 with \(\nu =k\) and Lemma 6 with \(\alpha =\delta =k\) to obtain
because \({\mathbb {E}}|\surd {n}(\bar{X}_n-\mu )|^k\rightarrow \sigma ^{k}{\mathbb {E}}|Z|^k<\infty \).
(b) Observe that the statement in Proposition 2(b) is equivalent to
and since \({\mathbb {E}}(\bar{X}_n-\mu )=0\), it suffices to verify that
If \(k=2\), then \(n{\mathbb {E}}[m_{1,n}(M_{2,n}-\mu _2)]=n{\mathbb {E}}[(\bar{X}_n-\mu )(m_{2,n}-\mu _2)] -n{\mathbb {E}}(\bar{X}_n-\mu )^3=\mu _3-n{\mathbb {E}}(\bar{X}_n-\mu )^3\), and it easily seen, by Lemma 6 with \(\alpha =\delta =3\), that \(|n{\mathbb {E}}(\bar{X}_n-\mu )^3| \le n^{-1/2} {\mathbb {E}}|\surd {n}(\bar{X}_n-\mu )|^3\rightarrow 0\); thus, \(n{\mathbb {E}}[m_{1,n}(M_{2,n}-\mu _2)]\rightarrow \mu _3\). Since \(\mu _1=0\), (35) is satisfied for \(k=2\).
If \(k=3\), \(n{\mathbb {E}}[m_{1,n}(M_{3,n}-\mu _3)]=n{\mathbb {E}}[(\bar{X}_n-\mu )(m_{3,n}-\mu _3)] +2n{\mathbb {E}}(\bar{X}_n-\mu )^4-3n {\mathbb {E}}[m_{2,n}(\bar{X}_n-\mu )^2]\), and it is easy to see that \(n{\mathbb {E}}[(\bar{X}_n-\mu )(m_{3,n}-\mu _3)]=\mu _4\). Also, by Lemma 6 with \(\alpha =\delta =4\), \(2n{\mathbb {E}}(\bar{X}_n-\mu )^4\rightarrow 0\). Finally, \(-\,3n{\mathbb {E}}[m_{2,n}(\bar{X}_n-\mu )^2]=-\,3[\mu _4+(n-1)\mu _2^2]/n\rightarrow -3 \mu _2^2=-\,3\sigma ^4\), which verifies (35) for \(k=3\).
In the general case when \(k\ge 4\), we write \(M_{k,n}-\mu _k\) as in (33) and we observe that for (35) to hold it suffices to verify that
- (i)
\(n {\mathbb {E}}[m_{1,n}(m_{k,n}-\mu _k)]=\mu _{k+1}\),
- (ii)
\(n{\mathbb {E}}(m_{1,n}^{k+1}) \rightarrow 0\),
- (iii)
\(n {\mathbb {E}}(m_{1,n}^{k+1-j} m_{j,n}) \rightarrow 0\), \(j=2,\ldots ,k-2\), and
- (iv)
\(n{\mathbb {E}}(m_{1,n}^2 m_{k-1,n}) \rightarrow \sigma ^2 \mu _{k-1}\).
Calculating \({\mathbb {E}}[m_{1,n}(m_{k,n}-\mu _k)]={\mathbb {E}}[(\bar{X}_n-\mu )(m_{k,n}-\mu _k)] ={\mathbb {E}}[(\bar{X}_n-\mu )m_{k,n}] = n^{-2}\sum _{i_1=1}^n\sum _{i_2=1}^n {\mathbb {E}}[(X_{i_1}-\mu )(X_{i_2}-\mu )^k] ={\mu _{k+1}}/{n}\), we conclude (i), while (ii) follows by using Lemma 6 with \(\alpha =\delta =k+1\). Also,
which shows that (iv) is satisfied, and it remains to verify (iii). To this end, we use Hölder’s inequality with \(p=(k+1)/(k+1-j)>1\) and Lemma 7 with \(\nu =k+1\) to obtain
because \(n^{-(k-1-j)/2}\rightarrow 0\); and, by Lemma 6 with \(\alpha =\delta =k+1\), \({\mathbb {E}}|\surd {n}(\bar{X}_n-\mu )|^{k+1}\rightarrow \sigma ^{k+1}{\mathbb {E}}|Z|^{k+1} <\infty \).
(c) Without loss of generality assume that \(2\le r\le k\) and observe that the first statement of Proposition 2(c) is equivalent to
Since \({\mathbb {E}}|X|^{r+k}<\infty \), (32) shows that \({\mathbb {E}}[\surd {n}(M_{k,n}-\mu _k)]\rightarrow 0\) and \({\mathbb {E}}[\surd {n}(\bar{X}_n-\mu )]\rightarrow 0\), and it suffices to verify that
The proof can be deduced by showing that (37) holds for each one of the cases \(r=k=2\); \(r=2\), \(k=3\); \(r=k=3\); \(r=2\), \(k\ge 4\); \(r=3\), \(k\ge 4\); \(4\le r\le k\). In the following we shall present the details only for the case where \(4\le r\le k\); the other cases can be treated using similar (and simpler) arguments.
Assume now that \(4\le r\le k\). From (33), we have
We shall show that the asymptotic covariance in (36) can be determined by using only the first two terms in (38) and (39). Indeed, it is easily seen that (37) holds true if it can be shown that
- (i)
\(n {\mathbb {E}}[(m_{r,n}-\mu _r)(m_{k,n}-\mu _k)]=\mu _{r+k}-\mu _r\mu _k\),
- (ii)
\(n{\mathbb {E}}[m_{1,n}m_{k-1,n}(m_{r,n}-\mu _r)] \rightarrow \mu _{r+1}\mu _{k-1}\),
- (iii)
\(n{\mathbb {E}}[m_{1,n}m_{r-1,n}(m_{k,n}-\mu _k)] \rightarrow \mu _{r-1}\mu _{k+1}\),
- (iv)
\(n{\mathbb {E}}(m_{1,n}^2m_{r-1,n}m_{k-1,n}) \rightarrow \sigma ^2 \mu _{r-1}\mu _{k-1}\),
- (v)
\(n {\mathbb {E}}[m_{1,n}^{k-j_2} m_{j_2,n}(m_{r,n}-\mu _r)] \rightarrow 0\), \(j_2=2,\ldots ,k-2\),
- (vi)
\(n{\mathbb {E}}[m_{1,n}^k (m_{r,n}-\mu _r)] \rightarrow 0\),
- (vii)
\(n {\mathbb {E}}(m_{1,n}^{k+1-j_2} m_{j_2,n}m_{r-1,n}) \rightarrow 0\), \(j_2=2,\ldots ,k-2\),
- (viii)
\(n{\mathbb {E}}(m_{1,n}^{k+1} m_{r-1,n}) \rightarrow 0\),
- (ix)
\(n {\mathbb {E}}[m_{1,n}^{r-j_1} m_{j_1,n}(m_{k,n}-\mu _k)] \rightarrow 0\), \(j_1=2,\ldots ,r-2\),
- (x)
\(n{\mathbb {E}}(m_{1,n}^{r+1-j_1} m_{j_1,n} m_{k-1,n}) \rightarrow 0\), \(j_1=2,\ldots ,r-2\),
- (xi)
\(n {\mathbb {E}}(m_{1,n}^{r+k-j_1-j_2} m_{j_1,n}m_{j_2,n}) \rightarrow 0\), \(j_1=2,\ldots ,r-2\), \(j_2=2,\ldots ,k-2\),
- (xii)
\(n{\mathbb {E}}(m_{1,n}^{r+k-j_1} m_{j_1,n}) \rightarrow 0\), \(j_1=2,\ldots ,r-2\),
- (xiii)
\(n {\mathbb {E}}[m_{1,n}^{r} (m_{k,n}-\mu _k)] \rightarrow 0\),
- (xiv)
\(n{\mathbb {E}}(m_{1,n}^{r+1} m_{k-1,n}) \rightarrow 0\),
- (xv)
\(n {\mathbb {E}}(m_{1,n}^{r+k-j_2} m_{j_2,n}) \rightarrow 0\), \(j_2=2,\ldots ,k-2\), and
- (xvi)
\(n{\mathbb {E}}(m_{1,n}^{r+k}) \rightarrow 0\).
We now proceed to verify (i)–(xvi). Since \({\mathbb {E}}(m_{r,n})=\mu _r\) and \({\mathbb {E}}(m_{k,n})=\mu _k\), we have
which shows (i). Also, (ii), (iii) and (iv) follow by straightforward computations; e.g., for (ii) we have
while (iii) is similar to (ii), and (iv) can be deduced from
The vanishing limits (vi)–(viii) and (x)–(xvi) are by-products of Lemmas 6 and 7 with \(\alpha =\delta =\nu =r+k\), since \({\mathbb {E}}|X|^{r+k}<\infty \). Indeed, we have \(|n{\mathbb {E}}(m_{1,n}^{r+k})|\le n{\mathbb {E}}|m_{1,n}|^{r+k} =n^{-(r+k-2)/2}{\mathbb {E}}|\surd {n}(\bar{X}_n-\mu )|^{r+k}\rightarrow 0\), which verifies (xvi). Also, using Hölder’s inequality with \(p=(r+k)/(r+k-j_2)>1\), we obtain (xv) as follows:
because \(n^{-(r+k-j_2-2)/2}\rightarrow 0\) and \({\mathbb {E}}|\surd {n}(\bar{X}_n-\mu )|^{r+k}\rightarrow \sigma ^{r+k}{\mathbb {E}}|Z|^{r+k} <\infty \); (xii) is similar to (xv). For the limit (xiv) we have
and similarly for (viii). In order to prove (xiii), it is sufficient to show that \(n{\mathbb {E}}[m_{1,n}^r m_{k,n}]\rightarrow 0\) and \(n{\mathbb {E}}(m_{1,n}^r)\rightarrow 0\). The second limit is obvious since, as for (xvi), one can easily verify that \(|n{\mathbb {E}}(m_{1,n}^r)|\le n^{-(r-2)/2}{\mathbb {E}}|\surd {n}(\bar{X}_n-\mu )|^r =n^{-(r-2)/2}O(1)\rightarrow 0\). For the first limit, we have
Limit (vi) is similar to (xiii), and its proof is omitted. Regarding (xi), we have
Similarly, for (x) we have
while (vii) is similar to (x).
It remains to verify (v) and (ix); but, since they are similar, it suffices to prove (v). If \(j_2\in \{2,\ldots ,k-3\}\) (and hence, \(k\ge 5\) and \(j_2<k-2\)), we have
and it suffices to prove that \(n{\mathbb {E}}(|m_{1,n}|^{k-j_2}|m_{j_2,n}m_{r,n}|)\rightarrow 0\) and \(n {\mathbb {E}}(|m_{1,n}|^{k-j_2}|m_{j_2,n}|)\rightarrow 0\). For the first quantity, we have
because \(k-j_2-2>0\). Similarly, for the second quantity we have
because \(k-j_2-2>0\). Finally, it remains to study the limit (v) when \(j_2=k-2\); in this case the above limits do not necessarily vanish. However, since \(j_2=k-2\) we have
and direct computations show that
and
Hence, when \(j_2=k-2\) we have
and the proof is complete. \(\square \)
Proof of Theorem 3
Observe that \(M_{k,n}-\mu _k=g_{k,k}({\varvec{m}}_{k,n})-g_{k,k}(\pmb {\mu }_{k})\); see in Sect. 2. Also, , where , see (1). Hence, Lemma 5 applies to \({\varvec{X}}_n={\varvec{m}}_{k,n}\), provided (21) is fulfilled for \({\varvec{m}}_{k,n}\), i.e., provided that . Because \(\nabla g_{k,k}(\pmb {\mu }_{k})=(-\,k\mu _{k-1},0,\ldots ,0,1)'\), we get \([\nabla g_{k,k}(\pmb {\mu }_{k})]'({\varvec{m}}_{k,n}-\pmb {\mu }_{k})=-\,k\mu _{k-1}m_{1,n}+(m_{k,n}-\mu _k)\). Since \({\mathbb {E}}(m_{j,n})=\mu _j\) for all n and j, we get \({\mathbb {E}}[-\,k\mu _{k-1}m_{1,n}+(m_{k,n}-\mu _k)]=0\). Also,
because \(v_k^2=0\) by the assumed singularness. Therefore, \([\nabla g_{k,k}(\pmb {\mu }_{k})]'({\varvec{m}}_{k,n}-\pmb {\mu }_{k})=0\) with probability one and, thus, in a trivial sense. Now, a simple calculation, since \(\nabla g_{k,k}(\pmb {\mu }_{k})=(-\,k\mu _{k-1},0,\ldots ,0,1)'\), shows that
i.e.,
etc. Applying (22), we see that \(n(M_{k,n}-\mu _k)\) converges weakly to the distribution of \(\frac{1}{2}{\varvec{W}}'_k\mathbf {H}_k(\pmb {\mu }_k){\varvec{W}}_k =\frac{1}{2}k(k-1)\mu _{k-2} W_1^2 - k W_1 W_{k-1}\), while, by (1), the distribution of \((W_1,W_{k-1})'\) is given by (24). \(\square \)
About this article
Cite this article
Afendras, G., Papadatos, N. & Piperigou, V.E. On the limiting distribution of sample central moments. Ann Inst Stat Math 72, 399–425 (2020). https://doi.org/10.1007/s10463-018-0695-4
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10463-018-0695-4