The $\chi^2$ distribution PDF is $$f_{\chi^2}(x;k) = \frac{1}{2^{k/2}\Gamma(k/2)} x^{k/2 - 1} \mathrm{e}^{-x/2} \mathbf{1}_{x \geq 0}$$

I am trying to find a polynomial approximation to this density function. Power series expansion for the exponential $\mathrm{e}^x$ works fine, but it's an infinite series. I an looking at possibilities to attain a more compact approximation. Approximation using orthogonal polynomials is one option I'm looking at.

Are there better alternatives to power series expansion of $\mathrm{e}^x$ for approximating this pdf?

  • 1
    $\begingroup$ Do you require a polynomial or would other approximations, such as Gaussians, work? (For instance, the cube root of $x$ is approximately Gaussian for modest to large $k$.) $\endgroup$ – whuber Nov 10 '11 at 21:01
  • $\begingroup$ See if this is helpful. $\endgroup$ – Sasha Nov 10 '11 at 21:09
  • 2
    $\begingroup$ Since $f(x)\to0$ when $x\to\infty$, $|f(x)-P(x)|\to\infty$ when $x\to\infty$ hence $(f-P)$ is unbounded for every polynomial $P$ except $P=0$. You could explain the sense in which $P$ should approximate $f$. $\endgroup$ – Did Nov 10 '11 at 22:31
  • 1
    $\begingroup$ If you're approximating on a bounded interval, you might use Chebyshev series or the Remez algorithm. As Didier noted, polynomials won't give you a good approximation on all of $[0,\infty)$. You could, however, approximate by rational functions. $\endgroup$ – Robert Israel Nov 11 '11 at 0:02
  • $\begingroup$ @RobertIsrael I want to approximate on all of $[0,\inf]$. I can see why approximation with polynomials could be problem as suggested by Didier. What approach is suggested for rational function approx? $\endgroup$ – sauravrt Nov 11 '11 at 17:35

Cecil Hastings, in his book Approximations for Digital Computers, gives a rational function approximation to $\exp(-x)$ for $0 \leq x < \infty$ that gives results that are good to $\approx 6$ digits:

$$\exp(-x)\approx R(x)=\frac1{(1+c_1 x+c_2 x^2+c_3 x^3+c_4 x^4+c_5 x^5+c_6 x^6)^4}$$


$$\begin{align*} c_1&=0.2499986842\\ c_2&=0.0312575832\\ c_3&=0.0025913712\\ c_4&=0.000171562\\ c_5&=5.4302\times 10^{-6}\\ c_6&=6.906\times 10^{-7} \end{align*}$$

Here's a plot of $R(x)-\exp(-x)$:

plot of R(x)-\exp(-x)

Hastings's book gives other, simpler approximations you can use if you have less stringent accuracy requirements. You can now plug this into your $\chi^2$ PDF expression to get the approximation you need.


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.