Sidi's generalized secant method

Sidi's method

Sidi's method is a root-finding algorithm, i.e. a numerical method for solving equations of the form $f(x)=0$ . The method was published by prof. Avraham Sidi.

Sidi's method is a generalization of the secant method.

Algorithm

We call $\alpha$ the root of $f$, i.e. $f(\alpha )=0$. Sidi's method is an iterative method which generates a sequence $\{x_{i}\}$ of estimates of $\alpha$. Starting with k + 1 initial estimates $x_{1},\dots ,x_{k+1}$, the estimate $x_{k+2}$ is calculated in the first iteration, the estimate $x_{k+3}$ is calculated in the second iteration, etc. Each iteration takes as input the last k + 1 estimates and the value of $f$ for those estimates. Hence the n-th iteration takes as input the estimates $x_{n},\dots ,x_{n+k}$ and the values $f(x_{n}),\dots ,f(x_{n+k})$.

The number k must be 1 or larger: k = 1, 2, 3, .... It remains fixed during the execution of the algorithm. In order to obtain the starting estimates $x_{1},\dots ,x_{k+1}$ one could carry out a few initializing iterations with a lower value of k.

The estimate $x_{n+k+1}$ is calculated as follows in the n-th iteration. A polynomial $p_{n,k}(x)$ of degree k is fitted to the k + 1 points $(x_{n},f(x_{n})),\dots (x_{n+k},f(x_{n+k}))$. With this polynomial, the next approximation $x_{n+k+1}$ of $\alpha$ is calculated as

$x_{n+k+1}=x_{n+k}-{\frac {f(x_{n+k})}{p_{n,k}'(x_{n+k})}}$ 1

with $p_{n,k}'(x_{n+k})$ the derivative of $p_{n,k}(x_{n+k})$ at $x_{n+k}$. Having calculated $x_{n+k+1}$ one calculates $f(x_{n+k+1})$ and the algorithm can continue with the (n + 1)-th iteration.

The iterative cycle is stopped if an appropriate stop-criterion is met. Typically the criterion is that the last calculated estimate is close enough to the sought-after root $\alpha$.

To execute the algorithm effectively, Sidi's method calculates the interpolating polynomial $p_{n,k}(x)$ in its Newton form.

Convergence

Sidi showed that if the function $f$ is (k+1)-times continuously differentiable in an open interval $I$ containing $\alpha$ (i.e. $f\in C^{k}(I)$), and the initial estimates $x_{1},\dots ,x_{k+1}$ are chosen close enough to $\alpha$, then the sequence $\{x_{i}\}$ convergence to $\alpha$ (i.e. the following limit holds: $\lim \limits _{n\to \infty }x_{n}=\alpha$).

Sidi furthermore showed that the sequence converges to $\alpha$ of order $\psi _{k}$, i.e.

$\lim \limits _{n\to \infty }{\frac {|x_{n+1}|}{|x_{n}|^{\psi _{k}}}}=\mu$

with a rate of convergence $\mu >0$.

Related algorithms

Sidi's method reduces to the secant method if we take k = 1. In this case the polynomial $p_{n,1}(x)$ is the linear approximation of $f$ around $\alpha$ which is used in the n-th iteration of the secant method.

We can expect that the larger we choose k, the better $p_{n,k}(x)$ is an approximation of $f(x)$ around $x=\alpha$. Also, the better $p_{n,k}'(x)$ is an approximation of $f'(x)$ around $x=\alpha$. If we replace $p_{n,k}'$ with $f'$ in (1) we obtain that the next estimate in each iteration is calculated as

$x_{n+k+1}=x_{n+k}-{\frac {f(x_{n+k})}{f'(x_{n+k})}}$ 2

This is the Newton-Raphson method. It starts off with a single estimate $x_{1}$ so we can take k = 0 in 2. It does not require an interpolating polynomial but instead one has to evaluate the derivative $f'$ in each iteration. Depending on the nature of $f$ this may not be possible or practical.

Once the interpolating polynomial $p_{n,k}(x)$ has been calculated, one can also calculate the next estimate $x_{n+k+1}$ as a solution of $p_{n,k}(x)=0$ instead of using 1. For k = 1 these two methods are identical: it is the secant method. For k = 2 this method is known as Muller's method. For k = 3 this approach involves finding the roots of cubic function, which is unattractively complicated. This problem aggravates for even larger values of k. An additional complication is that the equation $p_{n,k}(x)=0$ will in general have multiple solutions and a prescription has to be given which of these solutions is the next estimate $x_{n+k+1}$. Muller does this for the case k = 2 but no such prescriptions appear to exist for k > 2.