Consider the sequence of stochastic processes $(X_n, n \geq 1)$, where $X_n = (X_{t;n})_{t\in \mathbb Z}$ and: \begin{equation}\label{I}\tag{SP} X_{t;n} = \sum_{j=0}^\infty \theta_{jn} \varepsilon_{t-j;n} \end{equation} with $\sum_{j=0}^\infty \theta_{jn}^2 < \infty$ and $(\varepsilon_{t;n})_{t\in \mathbb Z} \overset{\text{iid}}{\sim} \mu_n(dx)$ with zero mean and variance $\sigma_n =1$, for all $n$. Suppose: \begin{equation}\label{uan}\tag{Uan} \quad\quad \max_{0\leq j } |\theta_{jn}| \longrightarrow 0\quad (n \to \infty). \end{equation}

Note that each $X_n$ is strictly stationary. Suppose $(X_{t})_{t\in \mathbb Z}$ is another strictly stationary satisfying the following two conditions:

  1. $E[X_{t;n}^2]\longrightarrow E[X_{t}^2]< \infty$, as $n \to \infty$, for all $t$;
  2. $(X_{t_1n}, X_{t_2n},\dotsc, X_{t_pn}) \Longrightarrow (X_{t_1}, X_{t_2},\dotsc, X_{t_p})$ as $n \to \infty$ for all $t_1 <t_2<\cdots < t_p$ (weak convergence of finite dimensional vectors).

Now, fix $(t_1 <t_2<\cdots < t_p)$. We can show, starting from condition 2 and using (\ref{I}) and (\ref{uan}), that: $$\sum_{j=0}^n \left(\theta_{jn} , \theta_{(j+ t_2 - t_1)n }, \dotsc, \theta_{(j+ t_p - t_1)n } \right) \varepsilon_{t_1-j;n} \Longrightarrow (X_{t_1}, X_{t_2},\dotsc, X_{t_p})\quad (n \to \infty).$$ For simplicity, denote $X_{jn}:=\left(\theta_{jn} , \theta_{(j+ t_2 - t_1)n }, \dotsc, \theta_{(j+ t_p - t_1)n } \right) \varepsilon_{t_1-j;n}$ and $X:=(X_{t_1}, X_{t_2},\dotsc, X_{t_p})$ (note that $X_{jn}$ and $X$ depends on $t_1, t_2,\dotsc, t_p$). So, the last equation means that $$S_n := \sum_{j=0}^n X_{jn} \Longrightarrow X\quad (n \to \infty).$$ Finally, we can show that there exist a non-negative definite matrix $\Sigma$ and a Levy measure $\nu$ such that the characteristic function of $X$ is: $$\varphi_X(u)= \exp\left\{ \frac{-u' \Sigma u}{2} +\int_{\mathbb R^p} \left[e^{iu'x} - 1- i u'x \right] d\nu(x) \right\}.$$ We adopt the following notation: \begin{equation} X_{jn}\sim \nu_{jn}(dx), \,\, \nu_n(dx):= \sum_{j=0}^n\nu_{jn}(dx). \end{equation} Notice that $\nu_{jn}$ is a probability measure in $\mathbb R^p$, since it depends on $\mu_{n}(dx)$ — the probability measure of the iid $(\varepsilon_{t;n})_{t \in \mathbb Z}$ defined on borelians of $\mathbb R$ — but also depends on the vector $\left(\theta_{jn} , \theta_{(j+ t_2 - t_1)n }, \dotsc, \theta_{(j+ t_p - t_1)n } \right)\in \mathbb R^p$.

Moreover, the measure $\nu$ can be characterized as follows: let $\mathcal C_\#$ be the class of continuous and bounded functions vanishing on a neighborhood of $0$. Then: \begin{equation}\label{M}\tag{M} \int f \, \nu_n(dx) \to \int f \, \nu(dx),\quad \forall f \in \mathcal C_\# \quad (n \to \infty). \end{equation} or equivalently (See Barczy and Pap - Portmanteau theorem for unbounded measures): $$\nu_n(E) \longrightarrow \nu(E), \quad (E\,\,\ \nu\hbox{-contunity set}, 0 \notin \overline{E},\,\, n \to \infty )\label{MI}\tag{M'}$$


Notice that $E[S_n]= \int_{\mathbb R^p} x \nu_n(dx) =0$, for all $n$. I want to show that: $$\int_{\mathbb R^p} x \nu(dx) =0\label{q1}\tag{I}$$ Can we show (\ref{q1}) or can we give a counterexample?


First, I can show that condition 1 and (\ref{uan}) imply: \begin{equation}\label{ui}\tag{UI} \int_{\mathbb R^p} |x|^2 \nu_{n}(dx) = \sum_{j=0}^n \int_{\mathbb R^p} |x|^2 \nu_{jn}(dx) \longrightarrow \int_{\mathbb R^p} |x|^2 \nu(dx)< \infty,\quad ( n\to \infty) \end{equation}

It is worth noting that we can define $$m_n(B):= \int_{B} |x|^2 \nu_n(dx)< \infty\quad\hbox{ and }\quad m(B):= \int_{B} |x|^2 \nu(dx)< \infty$$ for all borelian $B$. Since $E[S_n]=0$, we have $$\int_{\mathbb R^p} \frac{x}{|x|^2} m_n(dx)=0$$ and (\ref{q1}) is equivalent to: $$\int_{\mathbb R^p} \frac{x}{|x|^2} m(dx)$$ I would venture to say that this is true due to (\ref{ui}), using an argument similar to uniform integrability.

Sorry if I contextualized the issue too much, but I needed to avoid counterexamples like this answer. In this same question, we can also find more technical details about the context given here.

  • $\begingroup$ Please interpret $(\epsilon_{t:n})_{t\in\mathbb{Z}}\overset{\text{iid}}{\sim}\nu_n(dx)$(a compound Poisson Distribution?) $\endgroup$
    – JGWang
    Oct 10, 2023 at 9:58
  • $\begingroup$ No. It shouldn't be a CP distribution. It is just that, for each $n$, a sequence iid according to a distribution $\nu_n$. For example $(\epsilon _{t,n})_{t \in \mathbb Z}\sim N(0, \sigma_n)$ or $(\epsilon _{t,n})_{t \in \mathbb Z }\sim Bernoulli (1/n)$. $\endgroup$
    – PSE
    Oct 10, 2023 at 17:57
  • $\begingroup$ What the expression (Uan) means? $\max\limits_{1\le j\le n}|\theta_{jn}|\to0$ or $\max\limits_{j\in\mathbb{Z}_+}|\theta_{jn}|\to0$. $\endgroup$
    – JGWang
    Oct 11, 2023 at 2:48
  • $\begingroup$ Okay, you can assume the second option. I'll fix this now. Some minutes later: fixed. $\endgroup$
    – PSE
    Oct 11, 2023 at 3:32
  • $\begingroup$ @JGWang now we have $(\varepsilon_{t;n})_{t \in \mathbb Z} \overset{iid}{\sim} \mu_n(dx)$. $\endgroup$
    – PSE
    Nov 11, 2023 at 9:56

1 Answer 1


At first, we consider an example.

Let \begin{gather*} f(x)=\frac{I_{\{x>0\}}(x)}{2x^2(1\vee x^2)} =\frac{I_{\{(0,1)\}}(x)}{2x^2} + \frac{I_{\{[1,\infty)\}}(x)}{2x^4},\\ \nu(\mathrm{d}x)=f(x)\,\mathrm{d}x, \end{gather*} where $\nu$ is a $\sigma-$finite measure on $(\mathbb{R},\mathscr{B}(\mathbb{R}))$ and \begin{equation*} \int_0^\infty (1\wedge x^2)\nu(\mathrm{d}x)<\infty \tag{1} \end{equation*} Also, let \begin{align*} \psi_f(u)&=\int_{0}^{\infty}(e^{iux}-1-iux)\,\nu(\mathrm{d}x)\\ &=\int_{0}^{\infty}(e^{iux}-1-iux)f(x)\,\mathrm{d}x,\\ \bar{\phi}_f(u)&=\exp(\psi_f(u))=\exp\Big[\int_{0}^{\infty} (e^{iux}-1-iux)f(x)\,\mathrm{d}x\Big]. \tag{2} \end{align*} Then, from (1), $\bar{\phi}_f(u)$ is a (infinitely divisible) characteristic function and \begin{equation*} \bar{\phi}_f^\prime(0)=\psi_f^\prime(0)=0, \quad -\psi_f^{\prime\prime}(0)=\int_{0}^{\infty}x^2f(x)\,\mathrm{d}x=1. \tag{3} \end{equation*}

Now we are ready to construct a special $(\epsilon_{t;n})$ and $\theta_{jn}$. Let $(\epsilon_{t;n})_{t\in\mathbb{Z}}\stackrel{\mathrm{iid}}{\sim}\mu_n(dx)$, where the characteristic function of $\mu_n$ is \begin{equation*} \phi_{\epsilon;n}(u)\stackrel{\text{def}}{=}\int_{\mathbb{R}}\exp(iux)\mu_n(dx) =\exp\Big[\frac1n \psi_f(\sqrt{n}u)\Big], \end{equation*} then \begin{equation*} \mathsf{E}[\epsilon_{t;n}]=0, \qquad \mathsf{var}[\epsilon_{t;n}]=1,\qquad \forall n\ge 1. \tag{4} \end{equation*} Meanwhile, let \begin{equation*} \theta_{jn}=\frac{1}{\sqrt{n}}, \qquad 0\le j\le n-1,\quad \forall n\ge 1. \tag{5} \end{equation*} Hence, \begin{align*} X(t;n)&\stackrel{\text{def}}{=}X_{t;n}=\sum_{j=0}^{n-1}\theta_{jn}\epsilon_{t-j;n}=\frac{1}{\sqrt{n}} \sum_{j=0}^{n-1}\epsilon_{t-j;n},\qquad \forall n\ge 1,\\ \phi_{X(t;n)}&\stackrel{\text{def}}{=}\mathsf{E}[\exp(iX(t;n)u)] =[\phi_{\epsilon;n}(u/\sqrt{n})]^n \\ &=\exp[\psi_f(u)]=\bar{\phi}(u), \qquad \forall n\ge 1.\tag{6} \end{align*} From (4)(5), (Uan) and conditions 1,2 hold. If $\phi_{X(t_1)}(u) $ is the characteristic function of one dimensional distribution of limit process $X=\{X_t\} $, than, from (6), \begin{equation*} \phi_{X(t_1)}(u)=\phi_{X(t;n)}=\exp[\psi_f(u)]=\bar{\phi}(u). \end{equation*}

Now from above example we could discuss your questions.

Question 1 In above example, \begin{equation*} \int_{\mathbb{R}}|x|\nu(\mathrm{d}x)=+\infty. \end{equation*}

Question 2 In fact, if \begin{equation*} \int_{\mathbb{R}^p}|x|^2 \nu(\mathrm{d}x)<\infty, \tag{7} \end{equation*} $\bar{\phi}_f(u) $, defined by (2), may be $\phi_{X(t_1)}(u) $, the characteristic function of one dimensional distribution of limit process $X=\{X_t\} $. Hence your question transfer to ``if $\nu $ satisfy (7), under what conditions we have $\nu(\mathbb{R}^p)<\infty$ or $\nu(\mathbb{R}^p)=\infty$''.

  • $\begingroup$ Thank you so much for your answer. I have already reviewed your answer and it seems mathematically valid, although I expected that the distribution of sequences $X_n$ would not be the same as the limit. But mathematically, your counterexample is valid. Thanks. I just didn't understand very well regarding the possibility of $\nu(\mathbb R)= \infty $ or $\nu(\mathbb R)<\infty $. Could you comment on this? $\endgroup$
    – PSE
    Nov 26, 2023 at 21:24
  • $\begingroup$ @PSE Thank you for your replication. Regarding the Question II(in the early version of this post), I could not find a reasonable answer in term $\nu_n$ right now. If you still interesting in it, I agree the Thomas Kojar's suggestion. $\endgroup$
    – JGWang
    Nov 27, 2023 at 3:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.