3
$\begingroup$

I have a continues function $f:[a,b]\to \mathbb{R}$. And want to approximate this function by $p_n(x)$ , ($p_n(x)$ is a polynomial of maximum degree $n.$)

To do this, the value of below integral must be minimized.

Must the value of $\int\limits_0^1 |f(x)-p_n(x)|\, dx$ be minimize.

Whit the help of optimization , how this problem solved? how can determine the $p_n(x)$?

$\endgroup$
10
  • 2
    $\begingroup$ The larger $n$ the better the approximation unless $f$ is itself a polynomial of lower order. $\endgroup$
    – user121049
    Nov 19 '17 at 8:37
  • 1
    $\begingroup$ As @user121049 says, as $n$ increases the approximation gets better. You shouldn’t be solving for “the best $n$.” Rather, you should come up with a methodology that uses $n$ as a parameter. $\endgroup$ Nov 19 '17 at 14:25
  • 2
    $\begingroup$ Does it have to be the integral of the absolute value of the differences you are minimising? If you minimised the square instead then the problem is a lot easier. It becomes very similar to doing linear regression. $\endgroup$
    – user121049
    Nov 19 '17 at 18:25
  • 1
    $\begingroup$ @StellaBiderman Except on very short intervals near the point of expansion, Taylor approximation is very bad at achieving $L^p$ closeness. Even for analytic functions (for which it is guaranteed to eventually work) it is still generally slow. $\endgroup$
    – Ian
    Nov 19 '17 at 20:25
  • 1
    $\begingroup$ And for $L^2$ you can use Legendre polynomials. $\endgroup$ Nov 20 '17 at 3:23
2
$\begingroup$

If one considers the linear space $P_n$ of polynomials with degree $\leq n$, one can, via the Gram-Schmidt process, create an orthonormal basis of polynomials with respect to the inner product

$$\langle f,g \rangle = \int_0^1 f(t)g(t) dt.$$

Let these polynomials be $u_0,\cdots,u_n$. Then, one can project $f$ onto $P_n$ as

$$f_n(x) = \sum_{i=0}^n \langle f,u_i\rangle u_i(x).$$

This approximation $f_n(x)$ is the polynomial in $P_n$ that minimizes

$$\int_0^1 |f(x)-f_n(x)|^2 dx.$$

(As an aside, this is very similar to how a periodic function's Fourier series is constructed, using the already orthogonal basis $1,\cos(x),\sin(x),\cos(2x),\sin(2x),\cdots$ of the trigonometric polynomials.)

$\endgroup$
6
  • 3
    $\begingroup$ Note that this is $L^2$ minimization, which is a relatively easy problem. $L^1$ minimization as in the OP is not as easy as that. $\endgroup$
    – Ian
    Nov 19 '17 at 20:03
  • $\begingroup$ @Ian The way I read the question, it asked about a close approximate of a function by polynomials, not specifically about the $L^1$ norm (maybe just putting that down as an idea, and it is a good one). The way I've always seen this approximation done is with the $L^2$ norm, as it is much nicer. If this is more of an open-ended question, I think $L^2$ norm is fine. If not, can OP please indicate this, and I can delete my answer? $\endgroup$ Nov 20 '17 at 3:38
  • $\begingroup$ @Carl Schidkraut hello, thanks for your answer, in answer , where you use of optimization? $\endgroup$
    – linkho
    Nov 20 '17 at 9:15
  • $\begingroup$ @linkhochon I don't directly use optimization, but the approximation polynomial $f_n(x)$ is the one that minimizes $$\int_0^1 |f(x)-f_n(x)|^2 dx.$$ If you want, I can try to find you a source for this; I'm not sure how to prove it offhand. $\endgroup$ Nov 20 '17 at 14:56
  • $\begingroup$ @CarlSchildkraut The solution to this would be Legendre polynomials with some coordinate transformation, since these polynomials are orthogonal over the integral interval $[-1,1]$ instead of $[0,1]$. $\endgroup$ Nov 20 '17 at 18:56

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.