The formula
For a single equation f(x) = 0:
x_(n+1) = x_n − f(x_n) / f'(x_n)
Reading: the correction is how far off you are, divided by how fast the function is changing where you stand. Geometrically, you drop a tangent line at the current guess and jump to where that line hits the axis.
For a system of m nonlinear equations F(x) = 0 in m unknowns:
J(x_n) · Δx = −F(x_n), then x_(n+1) = x_n + Δx
Reading: same move, but the derivative becomes the Jacobian matrix J (every partial of every equation), and each iteration means solving one linear system. This is the version inside load-flow solvers and nonlinear FEA codes — Newton's method turns one hard nonlinear problem into a short stack of linear ones.
Near a simple root r, the error obeys:
e_(n+1) ≈ ( f''(r) / (2·f'(r)) ) · e_n²
Reading: the new error is proportional to the square of the old one. Squaring a small number is a violent operation — this is why the digit count roughly doubles per step.
Watch it happen on f(x) = x² − 2 (computing √2), starting from x₀ = 1: the iterates run 1.5, 1.41667, 1.4142157, 1.41421356237469. Correct digits: 1, then 3, then 6, then 12. Four iterations and you're at double-precision noise. Bisection would need about forty steps to get there.
Where you meet it
- Power system load flow. The steady-state grid problem — bus voltages and angles that satisfy the power balance at every node — is a big set of nonlinear equations, and the industry-standard solver is Newton-Raphson on the power mismatch vector. When a planning engineer says "the case didn't converge," this iteration is the thing that didn't converge, and the usual suspects are a bad flat start or a system genuinely near voltage collapse.
- SPICE and every circuit simulator descended from it. A diode's
I = I_s·(exp(V/(n·V_T)) − 1)makes the DC operating point a nonlinear problem, so the simulator linearizes each device around the present guess and Newton-iterates. The dreaded "no convergence in DC analysis" message, and the GMIN-stepping and source-stepping tricks that fix it, exist entirely to babysit this loop. - Nonlinear FEA. Contact, plasticity, and large deflection make stiffness depend on displacement, so the solver applies load in increments and runs Newton iterations against the residual force vector, rebuilding the tangent stiffness matrix as it goes. When the run log shows "cutback" after "cutback," that's the code shrinking the load step because Newton stopped converging from where it stood.
- Bench and hand-calc territory. The Colebrook friction factor, Mach number from an area ratio, normal depth in an open channel, liquid level in a horizontal tank from a volume reading, Kepler's equation in orbit propagation — all implicit equations with no closed-form inverse. A few Newton steps in a spreadsheet cell or a script beats guess-and-check every time.
How it works
The method makes exactly one bet: that your function looks like its tangent line near the current guess. Where that bet is good, convergence is spectacular — quadratic, digits doubling per step. Where it's bad, the method does not degrade gracefully; it lies to you with confidence, firing your next guess somewhere the tangent line pointed but the function never went.
Everything practical follows from that bet. You need f' to be healthy: a derivative near zero means a nearly flat tangent, and a nearly flat tangent crosses the axis in the next county — one step near a local extremum can throw the iterate out of the basin entirely. You need a decent starting point: Newton is a local method with no global guarantees, and which root you land on (or whether you land at all) depends on where you start. The basins of attraction are famously fractal — the Newton-fractal pictures are pretty, but the engineering reading is sober: near a basin boundary, tiny changes in the initial guess change the answer you converge to. Converging is not the same as converging to the root you meant.
The failure modes are worth knowing by name. Cycling: on f(x) = x³ − 2x + 2 starting from x₀ = 0, the iterates bounce 0 → 1 → 0 → 1 forever — no divergence warning, no convergence either. Multiple roots: at a repeated root, f' vanishes along with f, and convergence drops from quadratic to linear — the method still works, it just stops being fast, and that slowdown is itself a diagnostic that your root is multiple. Runaway: on functions like x^(1/3), every Newton step overshoots the root by more than the last, and the iteration diverges from any nonzero start.
For systems, two more realities. Each iteration costs a Jacobian build plus a linear solve, which is why production codes hold the Jacobian for several iterations (modified Newton), approximate it from iteration history (quasi-Newton — Broyden's update in circuit and structural codes, BFGS in optimization), or fall back on it only when convergence stalls. And a raw full Newton step is often too aggressive from a poor start, so real solvers damp it — take a fraction of Δx, check that the residual actually dropped, and only then accept. Load-flow programs, FEA codes, and SPICE all wrap Newton in this scaffolding; the textbook iteration is the engine, not the whole vehicle.
The mistake people actually make: declaring victory on the wrong test. A small step |x_(n+1) − x_n| does not mean a small residual |f(x)|, and vice versa — a flat function passes the residual test miles from the root, a steep one passes the step test while f is still large. Check both, in sane units. Poor scaling (one unknown in pascals, its neighbor in meters) quietly wrecks both the Jacobian conditioning and the convergence test, and it's the first thing to fix when a multi-physics Newton loop misbehaves.
One more place you use it daily without noticing: hardware and library routines for square roots, reciprocals, and divisions are Newton iterations on f(x) = x² − a or f(x) = 1/x − a, chosen precisely because a couple of digit-doubling steps from a table lookup gets full machine precision.
History
The method carries two names and earned at least three. Around 1669, Isaac Newton — building on a slower perturbation technique of François Viète from around 1600, with related root-extraction methods appearing in al-Kāshī's 1427 The Key to Arithmetic — worked out a faster scheme and demonstrated it on the cubic x³ − 2x − 5 = 0 [1][5]. His procedure looks strange to modern eyes: he noted the root was near 2, substituted x = 2 + p to get a new polynomial in the error p, kept only the linear terms to estimate p ≈ 0.1, then rewrote the polynomial again in the next error and repeated [1][5]. No derivative appears anywhere; the account sat in manuscripts, with the fuller treatment written around 1671 and not printed as the Method of Fluxions until 1736 [1][2]. In the Principia (1687) he turned the same machinery on his first non-polynomial target, Kepler's equation x − e·sin x = M [1][5]. The root of his demonstration cubic, for the record, is 2.0945514815… — his first two corrections give 2.1, then 2.0946.
Joseph Raphson beat Newton's method into print and into usable shape. His 1690 Analysis aequationum universalis computed each successive correction directly from the original polynomial instead of Newton's tedious re-expansion, making the scheme genuinely iterative — same formula every pass [1][2][5]. Raphson considered his derivation a different method from Newton's, and the Royal Society's records credited him with a rule that "doubles the known figures of the root … by each Operation" — quadratic convergence, observed on the ledger two centuries before it was proved [2].
The version engineers actually use belongs to Thomas Simpson. In 1740 — in Essays on Several Curious and Useful Subjects — he restated the iteration in terms of fluxions, Newton's calculus, which freed it from polynomials and made it work on any differentiable equation; he also extended it to systems of two equations and pointed it at optimization by driving derivatives to zero [1][3][4][5]. Both Newton and Raphson had worked without explicit derivatives at all. Historians who have traced the record — Ypma's 1995 SIAM Review account is the standard one — conclude that "Newton-Simpson" would be the fairer name for the general method [4][5]. Simpson, characteristically, is instead remembered for Simpson's rule, which he got from Newton [3]. The rigorous convergence theory for systems arrived much later, with Leonid Kantorovich's work in the late 1940s extending Newton's method to equations in function spaces [1][5].
Related tools
- /tools/pipe-pressure-drop — the Colebrook friction factor is implicit; solvers Newton-iterate it
- /tools/oblique-shock — the θ–β–M relation has no closed-form inverse for wave angle
- /tools/isentropic-flow — Mach number from an area ratio is a root-finding problem
- /tools/manning-open-channel — normal depth hides inside Manning's equation implicitly
- /tools/tank-volume-horizontal — level from volume inverts a transcendental segment-area formula