# Fixed Point for finding a root

I need to solve this equation (find $\lambda$) using numerical methods:

$\displaystyle N_0e^{\lambda}+v\frac{e^{\lambda}-1}{\lambda}-N_1 = 0$

All other terms are constant and known.

N0 = 1000000; v = 435000; N1 = 1564000;

I need to solve it by using bisection, false position, newton Raphson and fixed point. I have already solved it by using the three first approaches, however I am stuck in the fixed point. I cannot find a function, such that finding the roots of that function represents the equation and also $|f'(\alpha)| < 1$ where alpha is near the root (Condition that must be satisfied so that the method converges. I have tried several combinations, like dividing by N0, changing the starting approximation (I know that the root is around 0.1). Could somebody help me?

• Treat the $e^\lambda$ as a separate variable and solve for $\lambda$. I don't have the time at the moment to check the work, but this is a method that does converge to machine precision pretty quickly. The function looks like $$\lambda_{i+1}=\frac{v(e^{\lambda_i}-1)}{n_1-n_0e^{\lambda_i}}$$ – Terra Hyde Sep 3 '15 at 5:06
• It actually converges to zero, which is the solution to $\lambda(N_0 e^\lambda - N_1) + \nu (e^\lambda - 1) = 0$. It actually runs away from the root since $$\left|\frac{d}{d\lambda}\left(\frac{\nu (e^\lambda - 1)}{n_1 - n_0 e^\lambda}\right)\right| \approx 1.28 > 1.$$ Sure, you can solve for $e^\lambda$ and do the inverse iterations. – uranix Sep 3 '15 at 7:35
• And that is what I get for working while tired. Thanks. :) – Terra Hyde Sep 3 '15 at 14:34

If you know that the root is around $0.1$ then $$e^\lambda - 1 \approx \lambda$$ so the solution is close to the solution of $$N_0 e^\lambda + \nu = N_1$$ Solving it for $\lambda$ yields $$\lambda = \log \frac{N_1 - \nu}{N_0}.$$ Let's use that as the initial guess. $$\lambda_0 = \log \frac{1.564 - 0.435}{1} \approx 0.121332.$$ Recall that we approximated $\frac{e^\lambda - 1}{\lambda}$ by $1$ to obtain the approximate solution. Now let's bring that term back $$\lambda = \log \left(N_1 - \nu \frac{e^{\lambda} - 1}{\lambda}\right) - \log{N_0}.$$ Instead of approximating the term with $1$, let's use the last known approximate value for $\lambda$. That would give the following iterative method: $$\lambda_{n+1} = \log \left(N_1 - \nu \frac{e^{\lambda_n} - 1}{\lambda_n}\right) - \log{N_0}.$$ Here are the first 10 iterations of the method: $$\begin{array}{ccc} n & \lambda_n & \text{error}\\ 0 & 1.21332000000000\times 10^{-1} & 2.03341\times 10^{-2} \\ 1 & 9.66817916349767\times 10^{-2} & 4.31614\times 10^{-3} \\ 2 & 1.01904141090938\times 10^{-1} & 9.06211\times 10^{-4} \\ 3 & 1.00807223823552\times 10^{-1} & 1.90706\times 10^{-4} \\ 4 & 1.01038042978185\times 10^{-1} & 4.01133\times 10^{-5} \\ 5 & 1.00989491349787\times 10^{-1} & 8.43834\times 10^{-6} \\ 6 & 1.00999704757899\times 10^{-1} & 1.77507\times 10^{-6} \\ 7 & 1.00997556283298\times 10^{-1} & 3.73402\times 10^{-7} \\ 8 & 1.00998008234252\times 10^{-1} & 7.85485\times 10^{-8} \\ 9 & 1.00997913162376\times 10^{-1} & 1.65234\times 10^{-8} \\ 10 & 1.00997933161588\times 10^{-1} & 3.47584\times 10^{-9} \end{array}.$$ The method has $|f'(\alpha)| \approx 0.210$ near the root (gives $\approx 0.7$ correct digits per iteration).
The other method, based on inverse iterations proposed by Terra Hyde may be obtained if you solve the equation for $e^\lambda$ treating $\lambda$ as independent variable $$e^\lambda = \frac{N_1 + \nu \lambda}{N_0 + \nu \lambda}\\ \lambda_{n+1} = \log \frac{N_1 + \nu \lambda_n}{N_0 + \nu \lambda_n}.$$ The method has $|f'(\alpha)| \approx 0.771$ near the root (gives $\approx 0.11$ correct digits per iteration).