1
$\begingroup$

I have a problem of the following structure: For some scalar $g$, functions $F(z)$ and $h(z)$ defined on $[0,\bar{z}]$ , and a non-linear operator $\phi(F,z)$ (in reality, $F$ and $h$ are vector valued)

The $F(z)$ and $g$ must fulfill the ODE st. boundary values: $$ 0 = -F'(z;g) + \phi(F,z;g)\\ F(0) = 0\\ F(\bar{z}) = 1 $$ and the equilibrium condition: $$ 0 = -c + \int_{0}^{\bar{z}}h(z;g)F'(z;g) d z $$

My question is: what the appropriate algorithm to solve this numerically? I can solve the non-linear ODE with finite differences (and some shooting method in a more complicated variation), but a nested iteration with the integral equilibrium condition is difficult and expensive.

What about collocation? For example, pick the $N$th order chebyshev approximation. The chebyshev roots and polynomials (adapted to the $[0,\bar{z}]$ bounds) are $\{z_n\}_{n=0}^{N}$ such that $z_0 = 0$ and $z_N = \bar{z}$ and the polynomial of order $k$ evaluated at the $n$ root is $T_{nk} = T_n(z_n)$ and the vector of all orders at the $n$th root is $T_n$

Then if I assume a vector $a_0,\ldots a_N$ and a vector of these $F(z) \approx \sum_{a=0}^{N} a_n T_n(z)$ and therefore, $$F(z_n) = a \cdot T_n$$

Since chebyshevs are easy to differentiate, assume another set of values such that $F'(z_n) = a \cdot U_n$

Then I can apply the ODE and boundary values to get the usual system in $a$. For the integral constraint, I already have the chebyshev nodes, so if I assume weights $w_k$ for the usual chebyshev quadrature, then I get a final equation in $g$ $$ 0 = -c + \sum_{k=0}^N w_k h(z_k;g) \, a \cdot U_k $$

Does that sound like a defensible algorithm (using a nonlinear solver to find the $a,g$)? Are there other approaches which might work better?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

Your proposal sounds like a Galerkin-type approach. Specifically, it looks to be a spectral method rather than a finite element method. Collocation is typically associated with Runge-Kutta like methods for ODEs. Spectral methods are probably a fine approach here as long as you expect your solution to be smooth.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Thanks, I haven't entirely understood the difference - I thought it had to do with whether you minimize a residual directly or have an approximation hold exactly at a set of nodes. I think economists refer to this sort of thing where you have your operator hold exactly at the set of roots of the polynomial as collocation, of which implicit runge-kutta ends up a special case (rather than minimizing an actual integral). Regardless of semantics, it sounds like my approach to use the nodes for both quadrature and the BVP is reasonable? $\endgroup$
    – jlperla
    Commented Jul 21, 2014 at 18:42

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.