Replacing a derivative with a slope
Newton's method uses a tangent. When a derivative is unavailable or expensive, a secant through two previous points supplies an approximation.
Definition 6.2.1
(Secant iteration)
Let with . The secant iteration sets
This is the zero of the line interpolating at those points. It requires two initial values and a nonzero denominator. It is not a bracketing method: even if the initial points lie on opposite sides of a root, a later point need not remain between them.
Worked example 6.2.2
(One secant step on a quadratic)
Take with , . Then and , so . Check with the interpolating line: its slope is , and it vanishes at , in agreement. Since , the new point lies to the left of the root , at distance from it, much closer than (distance ) or (distance ).
The local error product
Let be a root and write . The key estimate involves two consecutive errors.
Lemma 6.2.3
(Error product for one secant step)
Let on an interval containing , with , and on . If both lie in , then the denominator of the secant step is nonzero and
Proof
(Interpolation error and the slope of the secant)
Let be the linear interpolant of through and . Interpolation with remainder at (which also holds when equals a node) gives
for some in the interval containing the three points. The slope of is the divided difference , whose magnitude is at least by the mean value theorem: is continuous and cannot change sign on . In particular it is nonzero, so the step is defined. Since , equals the slope magnitude times . Combining these facts proves the bound. If either iterate is already , stop; no further iteration is needed.
Definition 6.2.4
(Superlinear convergence)
A sequence with converges superlinearly if , where .
Theorem 6.2.5
(Local superlinear convergence of the secant method)
Suppose near a simple root , meaning and . There is a neighborhood such that any two distinct starting points in it produce a well-defined secant sequence converging to , unless a step reaches the root and the algorithm stops. If it never reaches the root, then
Thus the convergence is superlinear in this ratio sense.
Proof
(A fixed contraction factor, then a vanishing ratio)
Shrink a closed interval around so that the derivative has a positive lower bound and the second derivative has a finite upper bound . Choose small enough that the closed -neighborhood lies inside this interval and , where . If both current errors are at most , the product estimate gives . Induction keeps all points in the neighborhood and makes the errors tend to zero geometrically. As long as the current point is not a root, the next point differs from it; strict monotonicity of on the interval then makes the next denominator nonzero. Finally .
Proof X-ray
Superlinear is stronger than a fixed geometric factor
The product estimate first gives a fixed contraction factor , which keeps the iteration in the controlled neighborhood. Only after both errors tend to zero can we improve the ratio bound to . This proof does not require a claimed golden-ratio convergence order or an estimate of at each step.
Examples and limitations
Worked example 6.2.6
(Secant iterates for the square root of two)
For with , ,
The errors are , and , and the ratios and are small. The next iterate is with , and . These iterates approach ; the local theorem guarantees convergence once two consecutive iterates lie in a sufficiently small neighborhood, but it does not by itself prove that every arbitrary starting pair reaches that neighborhood.
Common mistake
(The secant method does not keep the root between the newest points)
In the example above, and lie on opposite sides of , and so do and . But and are both below , so the root is not between them, and the next step jumps to the other side, above (indeed ). Unlike bisection, the secant method does not maintain a sign-changing bracket, so the pair of newest points carries no error certificate.
Counterexample mode
A zero denominator can occur before reaching a root
For , the distinct starting points and give . The secant denominator is zero even though neither point is a root. A robust implementation must detect this situation; a local theorem near one simple root does not justify ignoring it globally.
The assumptions also matter at multiple roots. For with positive distinct starting values, the recurrence simplifies to . The derivative vanishes at the root, so there is no positive lower bound near zero and the theorem above does not apply.
Worked example 6.2.7
(A double root removes superlinear convergence)
For and positive , the recurrence above is equivalent to . Writing gives the Fibonacci recurrence , so grows like with and . The error ratio is , which tends to , not to . For example , gives , that is, : the successive ratios are , approaching . The sequence converges, but only linearly.
Practice and solutions
Quick check
(Compute one secant step for from , .)
Evaluate at both points first.
Solution · Answer
and , so .
Quick check
(Why can the local theorem ensure the divided difference is nonzero?)
Relate the divided difference to .
Solution · Answer
For two distinct points in the chosen interval, is continuous and bounded away from zero, so the mean value theorem bounds the divided-difference magnitude below by .
Quick check
(Does the local theorem guarantee convergence from , for ?)
Compare the function values at the two points.
Solution · Answer
No: these points are not in one controlled simple-root neighborhood, and their function values are equal, so the first denominator is zero.
Quick check
(If the product bound holds and both errors tend to zero without reaching the root, explain why the error ratios tend to zero.)
Divide the product bound by .
Solution · Answer
Dividing by gives a ratio at most .