7
$\begingroup$

For the standard Cauchy distribution, do fractional moments exist? If $Y \sim C(1,0)$, is it possible to evaluate $E(Y^{1/3})$?

$\endgroup$
2
  • 2
    $\begingroup$ Note that the Cauchy takes negative values, so in effect almost all fractional moments are ruled out immediately - as they are for the normal, say (assuming you expect real-valued answers). $\endgroup$
    – Glen_b
    Commented Jun 3, 2014 at 4:02
  • 1
    $\begingroup$ maybe complex moments $\endgroup$
    – Aksakal
    Commented Sep 5, 2017 at 21:00

4 Answers 4

6
$\begingroup$

Let's apply the law of the unconscious statistician, so we want to be able to evaluate

$\int_{-a}^b \frac{x^k}{1+x^2} dx$

in the limit as $a,b$ each $\to\infty$.

As mentioned in comments, for most fractional $k$ there's problems for negative $x$.

Let's step aside from that issue, and just talk about convergence of the right side.

Consider then, the convergence of

$I(b)=\int_0^b \frac{x^k}{1+x^2} dx\,$, for $0<k<1$

First consider splitting into $\int_0^1$ and $\int_1^b$ for $b>1$

for $\int_0^1 \frac{x^k}{1+x^2} dx$ note that when $x$ and $k$ are between 0 and 1 and that $x^k\leq 1$, so that integral is bounded.

Now note that for $x>0$, we have $\frac{1}{(1+x^2)}<x^{-2}$, so

$\int_1^b \frac{x^k}{1+x^2} dx<\int_1^b x^{k-2} dx = \left. \frac{x^{k-1}}{k-1}\right|_1^b = \frac{b^{k-1}-1}{k-1}=\frac{1-b^{k-1}}{1-k}$

which doesn't get any bigger than $\frac{1}{1-k}$ (when $b>1$ and $0<k<1$, as here).

So in the limit as $b\to\infty$, the integral converges.

As a result, $\int_{-\infty}^\infty \frac{|x|^k}{1+x^2} dx$ converges, and so when $x^k$ is real, $\int_{-\infty}^\infty \frac{x^k}{1+x^2} dx$ will converge to a real answer.

Specifically, the $\frac{1}{3}$ moment of a standard Cauchy is therefore 0 (as are all reciprocal-of-odd-positive-integer moments), by symmetry.


Note that according to Wolfram alpha (I didn't try to integrate it myself) $\int_{0}^\infty \frac{x^{1/3}}{1+x^2} dx=\frac{\pi}{\sqrt{3}}$.

$\endgroup$
2
  • 1
    $\begingroup$ The integral is easy to evaluate using the Calculus of Residues. For the moment to be well-defined and real you need $k$ to be a rational number with odd denominator or even numerator. Only in the latter case will the moment be nonzero. $\endgroup$
    – whuber
    Commented Sep 6, 2017 at 14:37
  • 1
    $\begingroup$ @whuber indeed, that's exactly how I'd have tackled it if I was going to try to evaluate it myself. $\endgroup$
    – Glen_b
    Commented Sep 7, 2017 at 2:56
3
$\begingroup$

To complement @Glen_b's answer, I asked maple for help, and it gives the following answer, which is a complex number since we ask for fractional powers of negative numbers! If such a complex moment is of any use I do not know, probably one would have to take the moments of the positive and negative part separately, as is done in definition of The Mellin transform https://en.wikipedia.org/wiki/Mellin_transform As that page shows, these fractional moments is essentially the Mellin transform. The Mellin transform is used in studying products of random variables.

The result is (I did not check on maple ...) $$ \int_{-\infty}^\infty x^\alpha \frac1{\pi (1+x^2)}\; dx = \frac{(-1)^\alpha + 1}{2 \cos (\pi\alpha/2)}, \quad \alpha < 1 $$ That's mostly curiosa, more interesting is the fractional moment of the absolute value $$ 2\int_0^\infty |x|^\alpha \frac1{\pi (1+x^2)}\; dx = \sec(\pi\alpha/2), \quad \alpha<1 $$ Some papers which seems to have more details: On Distribution of Product of Stable Laws and On the Use of Fractional Calculus for the Probabilistic Characterization of Random Variables and MELLIN-BARNES INTEGRALS FOR STABLE DISTRIBUTIONS AND THEIR CONVOLUTIONS (I did'nt look much at them yet, but they seem to be relevant)

$\endgroup$
2
$\begingroup$

The fractional moments $E[x^k]$ do exist for $k<1$. You have to be careful with fractional moments of distributions with negative values, of course. The value $x^k$ may end up being a complex number.

The absolute moments $E[|x|^k]$ were studied in Goria, M. N. "Fractional absolute moments of the Cauchy distribution." Quaderni di Statistica e Matematica applicata alle scienze Economico-Sociali 1 (1992): 3-9.

$\endgroup$
1
  • $\begingroup$ As I said elsewhere, this is really the Mellin transform. $\endgroup$ Commented Sep 6, 2017 at 14:38
0
$\begingroup$

Let $\mu \in \mathbb{R}$ and $\sigma > 0$ be the location and scale parameters of a Cauchy distribution respectively. Let $r > 0$ and $\theta \in (0, \pi)$ such that $r\exp(i \theta) = \mu + \sigma i$, where $i$ denotes the imaginary unit. Let $-1 < \text{Re}(a) < 1$. Assume that $x^a = |x|^a \text{sign}(x), \ x \ne 0$ and $0^a = 0$. Then, by (3.0.3) in V. M. Zolotarev, One-dimensional stable distributions, AMS, 1986 https://bookstore.ams.org/mmono-65,

$$ E[|X|^a] = \frac{\cos\left(a (\pi /2 - \theta)\right)}{\cos(a\pi /2)} $$ and $$ E[X^a] = \frac{\sin\left(a (\pi/2 - \theta) \right)}{\sin(a \pi/2)}, \ a \ne 0, $$ and $$ E[X^a] = 1 - \frac{2\theta}{\pi}, \ a = 0. $$

$\endgroup$

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.