I want to analyze or approximate a random variable that is a sum of two scaled independent non central $\chi$-squared random variables with the same degrees of freedom.

For example, $$X = X_1 + a X_2$$ where $X_1 \sim \chi^2(k,\lambda_1)$ and $X_2 \sim \chi^2(k,\lambda_2)$, $k$ is the degree of freedom, $\lambda_1, \lambda_2$ are noncentral parameters.

Simulations show that pdf of $X$ has the form of non central $\chi$-squared, or Gamma, ... I did try using characteristic function $M_X(t) = M_{X_1}(t)M_{X_2}(at)$ but it does not seem that the inverse transform of $M_X(t)$ has analytic form.

I wonder if there exists well-known results for this kind of problems.

For example,

$k=7, l_1=2,l_2=5, a=3$ enter image description here

$k=7, l_1=2,l_2=5, a=1$ enter image description here

$k=7, l_1=2,l_2=5, a=-1$ enter image description here

  • $\begingroup$ You have $X= X_1 + aX_2.$ There is a standard and well known answer in case $a=1.$ If $a\ne1,$ I suspect there is no closed form. $\endgroup$ – Michael Hardy Sep 3 '20 at 23:06

There is one case which is well-known.

Lemma. Let $Q_i \sim \chi^2_{k_i}(\lambda_i)$ for $i=1,\dots,n$, be independent. Then, $Q = \sum_{i=1}^n Q_i$ is a noncentral-$\chi^2_{k}(\lambda)$, where $k=\sum_{i=1}^n k_i$ and $\lambda = \sum_{i=1}^n \lambda_i$.

Proof. The Moment Generating Function (MGF) of $Q_i$ is given by $$ \mathcal{M}_{Q_i}(t) = (1 -2t)^{-k_i/2} \exp \Big\{ \frac{\lambda_i t}{1 - 2t} \Big\} $$ The MGF of $Q$, using independence, is given by $$ \mathcal{M}_Q(t) = \prod_{i=1}^n \mathcal{M}_{Q_i}(t) = \prod_{i=1}^n (1 -2t)^{-k_i/2} \exp \Big\{ \frac{\lambda_i t}{1 - 2t} \Big\} = (1 - 2t)^{- \frac{1}{2} \sum_{i=1}^n k_i} \exp \Big\{ \frac{t}{1 - 2t} \sum_{i=1}^n \lambda_i \Big\} $$ Thus, $$ \mathcal{M}_Q(t) = (1 -2t)^{-k/2} \exp \Big\{ \frac{\lambda t}{1 - 2t} \Big\} $$ and the thesis immediately follows. $\blacksquare$

Note that, if $\lambda \neq 0$, the PDF (probability density function) of $Q$ does not admit a closed-form, as it involves Bessel functions.

The general case of a linear combination of independent $\chi^2_{k_i}(\lambda_i)$ $$ Q = \sum_{i=1}^n a_i Q_i $$ results in a so-called generalized chi-squared distribution. The probability density function (PDF) of the generalized $\chi^2$ is also complicated.

  • $\begingroup$ Thanks for the info. However, this does not answer my question. $\endgroup$ – Sabrina Cantu Apr 19 '18 at 19:30
  • $\begingroup$ Can you please better explain your question? $\endgroup$ – PseudoRandom Apr 27 '18 at 9:23
  • $\begingroup$ If I follow your reasoning, the distribution I am looking at must be generalized chi squared distribution and its PDF is complicated. It is the definition. However, simulations show that the PDF looks like the one of scaled/non-central chi squared, etc. My question is whether we can approximate my interested PDF to a chi-squared-like PDF. $\endgroup$ – Sabrina Cantu Apr 29 '18 at 8:44

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.