The secant method is an iterative root-finding algorithm for solving an equation . Starting from two approximations, it constructs the straight line through their points on the graph of and uses that line’s intersection with the horizontal axis as the next approximation. It is closely related to Newton’s method, but replaces the derivative with a slope computed from two function values. For sufficiently good initial approximations near a simple root, it generally converges faster than linearly without requiring explicit derivative evaluations.
Iteration and geometric interpretation
Given initial values and , the secant method generates
The update requires . If either current approximation already satisfies , the calculation can terminate without another update.
The line through and has equation
Setting gives the secant formula. Thus each iteration replaces the function locally by a linear interpolant and finds the root of that interpolant.
The method always uses the two most recent iterates: after computing , the point is discarded. It does not require the two function values to have opposite signs, and it does not preserve an interval containing a root. The new approximation can lie outside the interval between the previous two.
Relationship to Newton’s method
Newton’s method uses
Replacing its derivative by the divided difference
produces the secant method.
Unlike a finite-difference Newton implementation that evaluates the function at a separately chosen nearby point, the secant method reuses a previous iterate and its stored function value. After the two initial evaluations, each successful iteration requires only one new function evaluation.
This makes the method useful when derivatives are unavailable or expensive. Newton’s method has a higher generic local convergence order at a simple root, but the relative computational cost depends on how easily the derivative can be obtained, including whether automatic differentiation is available.
Local convergence
A root is simple if
If is sufficiently smooth near a simple root and the two initial approximations are sufficiently close to it, the secant iteration converges locally, subject to the update remaining defined.
Let . A Taylor expansion gives the leading error relation
in the usual nondegenerate case where .
To identify the convergence order, suppose asymptotically that
Combining this form with the two-step error relation yields
The positive solution is
the golden ratio. The method therefore has generic superlinear convergence at a simple root: it is asymptotically faster than a linearly convergent method but has lower order than Newton’s quadratic convergence.
This order is a local, asymptotic result rather than a guarantee for arbitrary starting points. Special cases can behave differently. For an affine function with nonzero slope, one secant update finds the root in exact arithmetic. If , the leading error term vanishes and convergence can be faster. At a multiple root, where , the simple-root analysis does not apply, and convergence is generally only linear.
Example
Consider
with initial approximations and . The first update is
Subsequent approximations are:
| 0 | 1.0000000000 |
| 1 | 2.0000000000 |
| 2 | 1.3333333333 |
| 3 | 1.4000000000 |
| 4 | 1.4146341463 |
| 5 | 1.4142114385 |
| 6 | 1.4142135621 |
These approach the positive root . The iterates need not approach the root monotonically or alternate around it.
Implementation and numerical limitations
A basic implementation stores the two most recent approximations and their function values. Each iteration computes the secant step, evaluates the function at the new point, and shifts the stored pair.
Several issues affect reliability:
- Zero denominator. Distinct points can have equal function values, making the secant update undefined.
- Nearly equal function values. Their subtraction can suffer loss of relative accuracy in floating-point arithmetic. A small estimated slope may also produce a very large step.
- Stagnation. Rounding may make the new approximation equal to an existing one even when the root has not been accurately located.
- Domain violations. An extrapolated point can leave the domain of the function, where evaluation may be undefined or fail.
- Lack of global convergence. Poor starting points can lead to divergence, cycling, or convergence to a different root from the one intended.
These limitations concern both the iteration’s mathematical behavior and its numerical stability.
Common stopping tests examine the residual and the step size. A combined absolute and relative step test has the form
A small step alone does not establish that a root has been found, because it may result from stagnation. A small residual alone does not necessarily imply a small error in . Near a simple root,
so the relation between residual and root error depends on the local slope. This is an aspect of the conditioning of the root-finding problem.
Comparison with related methods
Bisection
The bisection method begins with a sign-changing interval for a continuous function and repeatedly halves it. The intermediate value theorem ensures that the interval contains a root. Bisection provides a predictable interval-width reduction and a direct error bound.
The secant method is usually faster near a simple root, but it does not provide the same bracketing guarantee.
False position
The false-position method, also called regula falsi, uses the same straight-line intersection formula but a different rule for retaining points. It preserves endpoints with opposite function signs rather than always retaining the two most recent approximations.
The distinction is algorithmic, not merely terminological. Standard false position preserves a bracket but may converge slowly when one endpoint remains fixed for many iterations; the ordinary secant method sacrifices that bracket to achieve its generic superlinear local convergence.
Safeguarded and hybrid methods
Hybrid algorithms combine interpolation steps with a bracketing fallback. Brent’s method, for example, combines secant steps, inverse quadratic interpolation, and bisection. Under its usual continuity and bracketing assumptions, safeguarding retains convergence reliability while allowing faster interpolation steps when appropriate.
Extensions to several variables
For nonlinear systems, quasi-Newton methods generalize the idea of updating derivative information from changes in function values. Instead of estimating a scalar slope, they update an approximation to a Jacobian matrix or its inverse. A multidimensional secant condition constrains that approximation, but unlike the scalar case, a single condition does not uniquely determine the matrix.