2
$\begingroup$

Given a function $1/\sqrt{x^2 -k^2}$ where k is a constant with a small imaginary part, how do you go about constructing a rational approximation? I am interested in the L_p (p=2 or $\infty$) norm of the difference being small on the real line. Both the theoretical and the practical implementation is of interest.

$\endgroup$
2
  • $\begingroup$ What do you mean by rational? Isn't your function rational? $\endgroup$ Oct 5, 2013 at 17:36
  • $\begingroup$ Sorry, forgot a square root. Updated. $\endgroup$
    – Steven
    Oct 5, 2013 at 17:48

2 Answers 2

1
$\begingroup$

By translation and scaling, we may assume WLOG $k = i$, i.e. your function is $1/\sqrt{x^2 + 1}$. Moreover, by symmetry we may assume the approximating function is even. So taking $x^2 = t$, we want to approximate $1/\sqrt{t+1}$ by rational functions of $t$ on $[0,\infty)$.

Now Maple's {\tt minimax} can approximate by rational functions, but it requires a bounded interval. So we map $[0,\infty)$ to $[0,1]$ by $1/(t+1) = s$. Thus we take $f(s) = 1/\sqrt{1/s} = \sqrt{s}$ on $[0,1]$. Now e.g. for a best uniform approximation of that by polynomials of degrees $5$ on $[0,1]$, we take

g:= numapprox:-minimax(sqrt(s),s=0..1,5,1,'maxerror');

Due to a possible bug, this produces an error in Maple 17. In Maple 15 I got

$$ g := 0.0278445029+(4.753636971+(-20.64608236+(47.77480263+(-49.59144735+18.70909011 s)s)s)s)s$$

with maximum error of $ 0.02784459798$. Substutute $s = 1/(1+x^2)$ to get a rational function of $x$ approximating $1/\sqrt{1+x^2}$ on $(-\infty,\infty)$ with this same error.

$\endgroup$
2
  • $\begingroup$ I am interested when $x$ is real so I can only rescale by reals and translate by reals. Therefore I do not see how I can take without loss of generality $k=i$. Could you explain this? $\endgroup$
    – Steven
    Oct 6, 2013 at 18:58
  • $\begingroup$ Oh, sorry, wrong function. I seem to have assumed it was $1/\sqrt{(x-a)^2 + b^2}$. $\endgroup$ Oct 6, 2013 at 19:34
1
$\begingroup$

Eleven years too late.

Using @Robert Israel's approach, the $[n,n]$ Padé approximants $P_n$ $$\sqrt s =\frac{1}{\sqrt{2}}\,\frac {1+\sum_{i=1}^n a_i\,\left(s-\frac{1}{2}\right)^i} {1+\sum_{i=1}^n b_i\,\left(s-\frac{1}{2}\right)^i}+O\left(\left(s-\frac{1}{2}\right)^{2n+1}\right) $$ can easily be built starting from the infinite series $$\sqrt s=\sum_{i=1}^\infty 2^{i-\frac{1}{2}} \binom{\frac{1}{2}}{i}\left(s-\frac{1}{2}\right)^i$$

The table below reports, for a given $n$ the coefficients as well as the norm $$\Phi_n=\int_0^1 \big[\sqrt s-P_n\big]^2\,ds$$ $$\left( \begin{array}{cccc} n & a_i & b_i & \Phi_n \\ 1 & \left\{\frac{3}{2}\right\} & \left\{\frac{1}{2}\right\} & 9.6\times 10^{-4} \\ 2 & \left\{\frac{5}{2},\frac{5}{4}\right\} & \left\{\frac{3}{2},\frac{1}{4}\right\} & 1.35\times 10^{-4} \\ 3 & \left\{\frac{7}{2},\frac{7}{2},\frac{7}{8}\right\} & \left\{\frac{5}{2},\frac{3}{2},\frac{1}{8}\right\} & 3.62\times 10^{-5} \\ 4 & \left\{\frac{9}{2},\frac{27}{4},\frac{15}{4},\frac{9}{16}\right\} & \left\{\frac{7}{2},\frac{15}{4},\frac{5}{4},\frac{1}{16}\right\} & 1.34\times 10^{-5} \\ 5 & \left\{\frac{11}{2},11,\frac{77}{8},\frac{55}{16},\frac{11}{32 }\right\} & \left\{\frac{9}{2},7,\frac{35}{8},\frac{15}{16},\frac{1}{32}\right\} & 6.05\times 10^{-6} \\ \end{array} \right)$$

$\endgroup$

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.