Given the following set of random variables and constants,

$\newcommand{\inreala}[2]{\in \mathbb{R}^{#1 \times #2}} \newcommand{\var}{\mathrm{Var}} \newcommand{\cov}{\mathrm{Cov}} \newcommand{\corr}{\mathrm{Corr}} \begin{align*} \text{Constants} :\forall i : \alpha_i \in \mathbb{R}^1 \\ \forall i \neq j: \cov(X_i, X_j) & :=\gamma_{i,j} \neq 0 \\ \forall i : \var(X_i) &= 1 \\ D &\in \mathbb{N}\\ \text{Random variables} : \left\{ X_i \right\}_{i=1}^D, Z & \in \mathbb{R}^1\\ Z &:= \sum_{i=1}^D \alpha_i X_i\\ \end{align*}$

how can we choose $\alpha_i$ such that the correlation between $X_i$ and the $\alpha$-weighted sum $Z$ is a desired value?

More precisely, for fixed $\left\{\alpha_j\right\}_{j \neq i}$, desired correlation $c_i$: find $\alpha_i$ s.t. $\corr(X_i, Z) = c_i$.

If I understand correctly so far, this will be a function of $\left\{\gamma_{i,j} \right\}_{i \neq j}$ and $\left\{\alpha_j\right\}_{j \neq i}$.

  • 1
    $\begingroup$ What is meant by $\alpha_i \in \mathbb R^{1\times 1}$? Is $\alpha_i$ an ordered pair of real numbers $\alpha_i = (\alpha_{i,1}, \alpha_{i,2})$? Or did your excessive formalism lead you astray and the $\alpha_i$'s are ordinary real numbers because that $^{1\times 1}$ is meant literally to be $^1$ on the grounds that $1\times 1 = 1$? $\endgroup$ Jan 26, 2018 at 2:55
  • 1
    $\begingroup$ Have you computed $\operatorname{Cov}(X_i, Z)$ in terms of the $\alpha_i$? If not, then start there. The rest is (simple) linear algebra. $\endgroup$
    – whuber
    Jan 26, 2018 at 14:23
  • $\begingroup$ @DilipSarwate $\alpha_i$ are just ordinary real numbers. I simplified the notation above. $\endgroup$
    – ijoseph
    Jan 26, 2018 at 17:45
  • $\begingroup$ @whuber thanks for the tip. I've computed it for the $D=2$ case: $\operatorname{Corr}(X_1, Z) = \frac{\alpha_1 + \alpha_2 \gamma_{1,2}}{\sqrt{\alpha_1^2 + \alpha_2^2 + 2 \alpha_1 \alpha_2 \gamma_{1,2}}}$. Having issues generalizing beyond that. My linear algebra is rusty -- any further pointers appreciated :) $\endgroup$
    – ijoseph
    Jan 26, 2018 at 17:54
  • 1
    $\begingroup$ This is only a two-variable problem! When you fix all the other $\alpha_j$ you are specifying a variable $Y=\sum_{j\ne i} \alpha_j X_j.$ You can find its variance and covariance with $X_i$ easily. Your problem therefore consists of finding a single $\alpha_i$ for which the correlation of $X_i$ with $\alpha_i X_i + Y$ equals $c_i$. $\endgroup$
    – whuber
    Jan 26, 2018 at 18:52

1 Answer 1


Thanks to whuber's key insight of combining all the other summands into one other random variable, I think I figured it out:

Let $Y := \sum_{j \neq i} \alpha_j X_j$; then:

$ \newcommand{\var}{\operatorname{Var}} \newcommand{\cov}{\operatorname{Cov}} \newcommand{\corr}{\operatorname{Corr}} \newcommand\covxiy{\cov(X_i, Y)} % gamma \newcommand\corrxiz{\corr(X_i, Z)} %c \begin{align*} c_i \equiv \corr(X_i, Z) &\equiv \corr(X_i, Y+ \alpha_i X_i)\\ &= \frac{\cov( X_i, Y + \alpha_i X_i)}{\sqrt{\var(X_i) \var(Y + \alpha_i X_i)}}\\ &= \frac{\covxiy + \alpha_i }{\sqrt{1 + \alpha_i^2 + 2 \alpha_i \covxiy}}\\ \end{align*}\\$

Now, solving for $\alpha_i$:

$\begin{align*} \Rightarrow \alpha_i = \frac{ \covxiy(1 - c_i^2) - c_i\sqrt{ \covxiy^2 c_i^2 - c_i^2 - \covxiy^2 + 1 }}{c_i^2 -1 }\\ \end{align*}$

Of course, what I really wanted was a full vector of $\vec{\alpha}$ corresponding to a desired vector of $\vec{c}$, but my question didn't ask that. I suppose there's some system of above such equations that would lead to such a solution.


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.