2
$\begingroup$

Let $X$ be a R.V whose pdf is given by $$f(x)=p\frac{1}{\sqrt{2\pi\sigma_1^2}}\exp\left(-\frac{(x-\mu_1)^2}{2\sigma_1^2}\right)+ (1-p)\frac{1}{\sqrt{2\pi\sigma_2^2}}\exp\left(-\frac{(x-\mu_2)^2}{2\sigma_2^2}\right)$$

clearly $X\sim pN(\mu_1,\sigma_1^2)+(1-p)N(\mu_2,\sigma_2^2)=N(p\mu_1+(1-p)\mu_2,p^2\sigma_1^2+(1-p)^2\sigma_2^2)$

therefore $\kappa=E(e^X)-1=e^{p\mu_1+(1-p)\mu_2+\frac{1}{2}(p^2\sigma_1^2+(1-p)^2\sigma_2^2)}-1$.

However, if I let $Y=e^X$, then Let $G(y)$ be the cdf of $Y$, then $G(y)=P(Y<y)=P(e^X<y)=p(X<\ln(y))=F_X(\ln(y))$ there fore $g(y)=f(\ln(y))\frac{1}{y}$ so the pdf $g(y)$ of $Y$ is given by

$$g(y)=p\frac{1}{y\sqrt{2\pi\sigma_1^2}}\exp\left(-\frac{(\ln(y)-\mu_1)^2}{2\sigma_1^2} \right) + (1-p) \frac{1}{y\sqrt{2\pi\sigma_2^2}} \exp\left(-\frac{(\ln(y)-\mu_2)^2}{2\sigma_2^2}\right)$$

i.e $Y=p\times \mathrm{logNormal}(\mu_1,\sigma_1^2)+(1-p)\times \mathrm{logNormal} (\mu_2,\sigma_1^2)$

where $\mathrm{logNormal}(\mu_1,\sigma_1^2)$ means a log-normal distribution.

hence $\kappa=E(Y)-1=pE(\mathrm{logNormal}(\mu_1,\sigma_1^2))+(1-p) E(\mathrm{logNormal} (\mu_2,\sigma_2^2))= pe^{\mu_1+\frac{\sigma_1^2}{2}}+(1-p)e^{\mu_2+\frac{\sigma_2^2}{2}}-1$.

My question is : why I have different values for $\kappa$? Could some one explain for me where I am wrong ? thank you for your time.

$\endgroup$
  • $\begingroup$ The first part of your post is incorrect. The distribution of a mixture is not the distribution of a weighted sum. $\endgroup$ – user76844 Aug 28 '14 at 16:15
  • $\begingroup$ Your second line is incorrect. $X$ has a mixture normal distribution and not a normal distribution as the last equality on the second line asserts. Your calculation of the pdf of $Y$ as a mixture of log-normal distributions is correct; the statement made on line 3 is not. $\endgroup$ – Dilip Sarwate Aug 28 '14 at 16:15
  • $\begingroup$ Eupraxix1981: why is that ? so what is the distribution of $X$ ? I found it here en.wikipedia.org/wiki/… $\endgroup$ – Dave Nguyen Aug 28 '14 at 16:20
  • $\begingroup$ Dilip : isn't if $Y\sim logNormal(\mu,\sigma^2)$ then $E(Y)=e^{\mu+0.5\sigma^2}$? $\endgroup$ – Dave Nguyen Aug 28 '14 at 16:22
2
$\begingroup$

With $p$ denoting a constant in the interval $(0,1)$, suppose $X$ has density $$f(x)=p\frac{1}{\sqrt{2\pi\sigma_1^2}}\exp\left(-\frac{(x-\mu_1)^2}{2\sigma_1^2}\right)+ (1-p)\frac{1}{\sqrt{2\pi\sigma_2^2}}\exp\left(-\frac{(x-\mu_2)^2}{2\sigma_2^2}\right)$$ which is a mixture density. Then, $$\begin{align} F_X(\alpha) &= P\{X \leq \alpha\}\\ &=\int_{-\infty}^{\alpha} f(x)\,\mathrm dx\\ &= \int_{-\infty}^{\alpha}p\frac{1}{\sqrt{2\pi\sigma_1^2}}\exp\left(-\frac{(x-\mu_1)^2}{2\sigma_1^2}\right)+ (1-p)\frac{1}{\sqrt{2\pi\sigma_2^2}}\exp\left(-\frac{(x-\mu_2)^2}{2\sigma_2^2}\right)\,\mathrm dx\\ &= p\int_{-\infty}^{\alpha}\frac{1}{\sqrt{2\pi\sigma_1^2}}\exp\left(-\frac{(x-\mu_1)^2}{2\sigma_1^2}\right)\,\mathrm dx+(1-p)\int_{-\infty}^{\alpha} \frac{1}{\sqrt{2\pi\sigma_2^2}}\exp\left(-\frac{(x-\mu_2)^2}{2\sigma_2^2}\right)\,\mathrm dx\\ &= p\Phi\left(\frac{\alpha-\mu_1}{\sigma_1}\right) +(1-p)\Phi\left(\frac{\alpha-\mu_1}{\sigma_1}\right)\\ &\neq \Phi\left(\frac{\alpha-(p\mu_1+(1-p)\mu_2)}{\sqrt{p^2\sigma_1^2+(1-p)^2\sigma_2^2}}\right) \end{align}$$

In other words, it is not true (as you state on Line 3 of your question, and continue to assert dogmatically) that $X$ has a normal distribution with mean $p\mu_1+(1-p)\mu_2)$ and variance $p^2\sigma_1^2+(1-p)^2\sigma_2^2$. That is, $$X\sim N(p\mu_1+(1-p)\mu_2,p^2\sigma_1^2+(1-p)^2\sigma_2^2)\quad \mathbf{is~a~false~statement}$$

As a result, your calculation $\kappa=E(e^X)-1=e^{p\mu_1+(1-p)\mu_2+\frac{1}{2}(p^2\sigma_1^2+(1-p)^2\sigma_2^2)}-1$ is not correct.

$\endgroup$
  • $\begingroup$ Thanks Dilip :)) I got it. $\endgroup$ – Dave Nguyen Aug 29 '14 at 15:10
2
$\begingroup$

Let $\mu := p\mu_1+(1-p)\mu_2$ and $\sigma^2 := p^2\sigma_1^2 + (1-p)^2\sigma_2^2$. Then, one can show that, the pdf of $Y$ is: $$g(y)=\frac{1}{y\sqrt{2\pi\sigma^2}}\exp\left(-\frac{(\ln(y)-\mu)^2}{2\sigma^2}\right),$$ not the one that you display (your $Y$ presentation as linear combination of log-normal variables is wrong: if $X=pX_1 +(1-p)X_2$, then $Y={\rm e}^X = {\rm e}^{pX_1}{\rm e}^{(1-p)X_2}$, not $p{\rm e}^{X_1}+(1-p){\rm e}^{X_2}$; also, $f$, pdf of $X=pX_1+(1-p)X_2$, is NOT $pf_1 + (1-p)f_2$, where $f_1$ and $f_2$ are pdf's of $X_1$ and $X_2$; lastly, see Dilip Sarwarte comment on what is really your starting hypothesis on $X$ being normal variable, as say a linear combination of independent normal variables, or on its pdf being a mixture of normal pdf's).

$\endgroup$
  • $\begingroup$ But what you wrote is exactly $X=N(p\mu_1+(1-p)\mu_2,p^2\sigma_1^2+(1-p)^2\sigma_2^2)$? I am confused now :(( Here is what I have Let $G(y)$ be the cdf of $Y$, then $G(y)=P(Y<y)=P(e^X<y)=p(x<ln(y))=F_X(ln(y))$ there fore $g(y)=f(ln(y))\frac{1}{y}$ $\endgroup$ – Dave Nguyen Aug 28 '14 at 16:26
  • $\begingroup$ Yes. $Y={\rm e}^X$ and $X$ is normal with params $\mu$ and $\sigma^2$. Hence $Y$ is log-normal. Hence $g$ form. $\endgroup$ – ir7 Aug 28 '14 at 16:29
  • $\begingroup$ Here is what I have Let $G(y)$ be the cdf of $Y$, then $G(y)=P(Y<y)=P(e^X<y)=p(X<ln(y))=F_X(ln(y))$ there fore $g(y)=f(ln(y))\frac{1}{y}$ $\endgroup$ – Dave Nguyen Aug 28 '14 at 16:31
  • 1
    $\begingroup$ $f$, pdf of $X=pX_1+(1-p)X_2$, is NOT $pf_1 + (1-p)f_2$, where $f_1$ and $f_2$ are pdf's of $X_1$ and $X_2$. $\endgroup$ – ir7 Aug 28 '14 at 16:48
  • 1
    $\begingroup$ @Nguyen That's ok. In my answer, I mentioned Dilip Sawarte's comment. Basically, if you stop claiming that $X$ is normal, then $f=pf_1+(1-p)f_2$, as given directly by the problem. Hence, your $g$ is correct. If you claim that $X$ is normal (with $\mu$ and $\sigma$), then my $g$ is correct. $\endgroup$ – ir7 Aug 28 '14 at 19:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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