2

I have been trying to show that given $$P_{n}(\alpha) = \alpha^{n} - a_{1}\alpha_{n-1} - a_{2}\alpha^{n-2}... - a_{n} = 0,$$ the $\alpha$'s that solve this equation (real-valued or complex) lie in the unit circle only if $\sum\limits_{i=1}^{n} a_{i} < 1$. This result is referenced at the bottom of the reply in the following link Are all $AR(p)$ processes for which $|a_1|,....,|a_p| < 1$ stationary?.

By first assuming only real roots, I have argued as follows. Suppose all roots lie inside the unit circle, i.e. $|\alpha| < 1$ for every root $\alpha$. Then for all $x \geq 1$ we must have that $P_{n}(x) > 0$ or $P_{n}(x) < 0$. If not, then by the intermediate value theorem there would exist a root of magnitude larger than $1$. Since for large enough $x$ we must have $P_{n}(x) > 0$ (since $\lim\limits_{x\rightarrow \infty} P_{n}(x) = \infty$), we conclude that $P_{n}(x) > 0$ for all $x\geq 1$. It follows that $P_{n}(1) = 1 - \sum\limits_{i = 1}^{n} a_{i} > 0$, which is equivalent to the statement $\sum\limits_{i = 1}^{n} a_{i} < 1$.

I do not see how I can extend this argument to account for complex roots. I would appreciate any help, both by reference and by explanation/hint. Thanks in advance to all that took the time to read this post.

  • Hi: There seems to be a proof here if you can decipher it. https://www.uni-ulm.de/fileadmin/website_uni_ulm/mawi.inst.050/people/lindner/brockwell-lindner2.pdf – mlofton May 03 '22 at 00:36
  • @mlofton Thank you very much. I tried looking at it, but I could not find the result in the article. Perhaps I did not understand it properly, but could you perhaps be so kind as to refer me to where in the article the result is shown? – Thomas Fjærvik May 03 '22 at 10:36
  • What you're trying to prove is false. You have reversed the condition: that inequality was asserted to be a necessary condition, not a sufficient one. Your reference is incorrect, anyway: there is no general relationship between this inequality and the locations of the roots of the polynomial. As a counterexample, consider the polynomial $P(x)=x^1-(-2)$ for which $\sum a_i=-2\lt 1$ but whose only root, $-2,$ is outside the unit circle. – whuber May 03 '22 at 13:48
  • 1
    Thomas. That paper I sent is very difficult to understand because they are actually proving something stronger by relaxing conditions in the usual ARMA. I know that there is a better proof around but I haven't been able to find it. I'll keep looking. Also, I would correct your original question because I didn't read it carefully but what Whuber said is probably correct. I just know what you're looking for so I didn't need to read it carefully. I'll look around because I know that there's a pretty recent simpler proof. – mlofton May 04 '22 at 14:27
  • This is not what i've been looking for but it's a lot clearer than the previous one that I sent. Does this one make it clearer to you ? https://eml.berkeley.edu/~powell/e241b_f06/TS-StatInv.pdf – mlofton May 04 '22 at 14:52
  • I need to read it carefully but aslanyan's answer here looks pretty instructive also.https://stats.stackexchange.com/questions/402696/stationarity-of-arp-process – mlofton May 04 '22 at 14:56
  • @mlofton Thanks for the references! I will look at the carefully tomorrow. – Thomas Fjærvik May 05 '22 at 20:18
  • @whuber Maybe I am misunderstandig something, but I cannot see that I claimed sufficiency? I am claiming (or at least trying to claim) that if all the roots of $P_{n}(\alpha)$ are inside the unit circle, then $\sum\limits_{i=1}^{n} a_{i} < 1$. Have I perhaps used the phrasing "only if" incorrectly? If so, I apologize for being unclear and I will of course edit my question. – Thomas Fjærvik May 05 '22 at 20:22
  • @mlofton As far as I can see from your references, they talk about something different than what I am asking about and not mention the condition in my post. If I am wrong, then please let me know. And thanks again for the references! – Thomas Fjærvik May 05 '22 at 21:02
  • Hi Thomas: invertibility of the MA(1) is the same as stationarity of the AR(\infty) so the second paper I linked to definitely addresses your question. It just looks different initially because it discusses the MA(1).If you read it carefull, you'll see that your question is addressed. – mlofton May 06 '22 at 00:05
  • @mlofton Here is my understanding of the second paper you referenced: It demonstrates how invertibility of a MA(q) process is linked to the stationarity of an AR(p) process. It then shows a condition for when the lag polynomial associated with an MA(q) process is invertible. This is seemingly linked to \textit{sufficient} conditions for stationarity of a process. My question (attempts to) adress a \textit{necessary} condition for stationarity. I cannot find this adressed anywhere in the paper you referenced. I would really appreciate if you could point out where it is adressed. Thanks again! – Thomas Fjærvik May 06 '22 at 10:04
  • I thought that the last section "invertibility of lag polynomials" ( on the last page ) provided a proof of what you wanted. if you say it doesn't, I believe you. Let me look on another computer because I know that I had a paper of a proof of it somewhere and I'm hoping it's on that computer. I apologize for leading you astray and I'll get back to you with success or failure today. Again, my sincere apologies. – mlofton May 06 '22 at 16:40
  • Hi Thomas: I went on to another computer and got lucky and found what I was looking for. This one is DEFINITELY a proof ( but almost impossible to find using google ). I'm not saying it's an easy proof but it's a proof. Good luck and again my apologies for leading you astray the first two times. Oh and thanks for your question because I've always wanted to try to understand the gory details so maybe this will motivate me to do that. papers.econ.ucy.ac.cy/RePEc/papers/06-08.pdf – mlofton May 06 '22 at 17:24

1 Answers1

2

Let

$$P(z) = z^n - \sum_{i=0}^{n-1} a_i z^i.$$

Notice that

$$P(1) = 1^n - \sum_{i=0}^{n-1} a_i 1^i = 1 - \sum_{i=0}^{n-1} a_i.$$

  1. If all the roots are inside the unit circle, there are no roots on the real interval $[1, \infty).$ Because the values of $P(x)$ increase to $+\infty$ as the real number $X$ grows large, $P$ must be positive everywhere on $[1,\infty).$ In particular, $P(1) \gt 0,$ equivalent to $\sum_{i=0}^{n-1} a_i \lt 1.$

  2. When $\sum_{i=0}^{n-1} a_i \lt 1,$ $P(1) \gt 0.$ However, this does not imply all the roots of $P$ lie inside the unit circle. The foregoing analysis suggests a way to construct counterexamples: give $P$ a zero in the interval $(1,\infty).$ Since $P(1)\gt 0$ and eventually $P(x)\gt 0$ for large $x,$ $P$ must have at least two zeros in this interval. The simplest possible counterexamples will be quadratic. Consider, for instance, $P(x) = (x-2)^2 = x^2 - 4x + 4,$ for which $a_1=4$ and $a_0=-4.$ $P$ has no zeros inside the unit circle but the sum of the $a_i$ is less than $1.$

Consequently, $\sum a_i\lt 1$ is a necessary but not sufficient condition for all the roots of $P$ to lie within the unit circle.

whuber
  • 322,774
  • 1
    Thanks for your reply! Everything seems to make sense now. Though the argument in point (1) seems to be the same as mine, the way you phrased made everything click. I approached it by first considering real roots, and then I planned to consider complex roots. I somehow failed to notice that it is not necessary to consider any complex roots when we have shown that the sum is less than $1$ as long as there are no real roots in $[1, \infty)$. – Thomas Fjærvik May 07 '22 at 22:05