0

Suppose $a_i,q_i,b_0$ are positive real numbers. I need to solve the following equation for $k$

$$\sum_i^d a_i \exp(-q_i k)=b_0$$

Is this a well-known problem? One my special cases has $a_i=q_i$

In my application $d> 20000$, $q_i\approx 0$. Wondering if equation structure makes it possible to solve faster than Newton's method.

Yaroslav Bulatov
  • 2,655
  • 11
  • 23
  • 1
    If the $q$ are all small, and if you expect $k$ to be of moderate size, then $e^{-q_ik}\approx 1-q_ik$ and the whole problem becomes one simple linear equation in $k$. At the very least, this will provide for a good starting guess for a Newton method for solving the "real" problem. – Wolfgang Bangerth Mar 01 '23 at 22:31
  • I think it becomes $(1-q_i)^k$. This is actually the equation I started with, before switching to $e$ approximation, assuming some transform or series expansion of $e$ might be useful here – Yaroslav Bulatov Mar 01 '23 at 23:03
  • Cross-posted on dsp.SE – Yaroslav Bulatov Mar 02 '23 at 17:52
  • 1
    The first order Taylor expansions of $e^{-qk}$ and $(1-q)^k$ are both $1-qk$ (as you can see by multiplying out what $(1-q)^k$ actually is). If $qk\ll 1$, you can choose whichever you want, and $1-qk$ is definitely the easier one. – Wolfgang Bangerth Mar 02 '23 at 22:13
  • Newton should converge in very few steps and without numerical problems, since the function and all its derivatives are monotonic; and each step costs 2x as much as evaluating the function itself. It seems hard to beat if you consider these points. – Federico Poloni Mar 04 '23 at 21:46

1 Answers1

3

If $q_i$ are small, you might get away with solving for k with only one summand where q_i is smallest, as that will be the dominant term. The other terms where q_i are larger will be small due to the exponential function. As the comments suggest, that might just serve as a good starting point for a newton iteration.

MPIchael
  • 2,935
  • 10
  • 19
  • Good point, that approach gives a lower bound on true value. And I can get upper bound by averaging $q_i$'s together and using Jensen's – Yaroslav Bulatov Mar 03 '23 at 08:18