Standard deviation of random variable $A =\sigma_{1}=5$

Standard deviation of random variable $B=\sigma_{2}=4$

Then the variance of A+B is:

$Var(w_{1}A+w_{2}B)= w_{1}^{2}\sigma_{1}^{2}+w_{2}^{2}\sigma_{2}^{2} +2w_{1}w_{2}p_{1,2}\sigma_{1}\sigma_{2}$


$p_{1,2}$ is the correlation between the two random variables.

$w_{1}$ is the weight of random variable A

$w_{2}$ is the weight of random variable B


The figure below plots the variance of A and B as the weight of A changes from 0 to 1, for the correlations -1 (yellow),0 (blue) and 1 (red).

alt text

How did the formula result in a straight line (red) when the correlation is 1? As far as I can tell, when $p_{1,2}=1$, the formula simplifies to:

$Var(w_{1}A+w_{2}B)= w_{1}^{2}\sigma_{1}^{2}+w_{2}^{2}\sigma_{2}^{2} +2w_{1}w_{2}\sigma_{1}\sigma_{2}$

How can I express that in the form of $y=mx+c$?

Thank you.

  • $\begingroup$ Don't you mean $Var(w_1 A + w_2 B)$, since you weigh them? $\endgroup$ Dec 12 '10 at 9:20
  • $\begingroup$ @Raskolnikov: Thank you for pointing that out. I have edited it. $\endgroup$
    – Sara
    Dec 12 '10 at 10:41

Using $w_1 + w_2 = 1$, compute

$$\eqalign{ \text{Var}(w_1 A + w_2 B) &= \left( w_1 \sigma_1 + w_2 \sigma_2 \right)^2 \cr &= \left( w_1(\sigma_1 - \sigma_2) + \sigma_2 \right)^2 \text{.} } $$

This shows that when $\sigma_1 \ne \sigma_2$, the graph of the variance versus $w_1$ (shown sideways in the illustration) is a parabola centered at $\sigma_2 / (\sigma_2 - \sigma_1)$. No portion of any parabola is linear. With $\sigma_1 = 5$ and $\sigma_2 = 4$, the center is at $-5$: way below the graph at the scale in which it is drawn. Thus, you are looking at a small piece of a parabola, which will appear linear.

When $\sigma_1 = \sigma_2$, the variance is a linear function of $w_1$. In this case the plot would be a perfectly vertical line segment.

BTW, you knew this answer already, without calculation, because basic principles imply the plot of variance cannot be a line unless it is vertical. After all, there is no mathematical or statistical prohibition to restrict $w_1$ to lie between $0$ and $1$: any value of $w_1$ determines a new random variable (a linear combination of the random variables A and B) and therefore must have a non-negative value for its variance. Therefore all these curves (even when extended to the full vertical range of $w_1$ ) must lie to the right of the vertical axis. That precludes all lines except vertical ones.

Plot of the variance for $\rho = 1 - 2^{-k}, k = -1, 0, 1, \ldots, 10$:

alt text


It isn't linear. The formula says it isn't linear. Trust your mathematical instinct!

It only appears linear in the graph because of the scale, with $\sigma_{1}=5$ and $\sigma_{2}=4$. Try it yourself: calculate the slopes at a few places and you will see that they differ. You can exaggerate the difference by picking $\sigma_{1}=37$, say.

Here is some R code:

a <- 5; b <- 4; p <- 1
f <- function(w) w^2*a^2 + (1-w)^2*b^2 + 2*w*(1-w)*p*a*b
curve(f, from = 0, to = 1)

If you would like to check some slopes:

(f(0.5) - f(0.4)) / 0.1
(f(0.8) - f(0.7)) / 0.1

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.