Does it require second derivatives of each residual?

The basic method does not; it uses residuals and their Jacobian.

Is it identical to gradient descent?

No. Its least-squares step incorporates the local sensitivity structure through the Jacobian.

Can it fit inconsistent observations?

Yes. Least squares seeks a compromise; the minimum residual need not be zero.

AI Glossary / Mathematics and optimization / Advanced

Gauss - Newton algorithm

Definition

The Gauss - Newton algorithm is an iterative method for nonlinear least squares. It improves a parameter estimate by linearizing the residual functions and solving a linear least-squares problem for the next step, avoiding the residual second derivatives required by a full Newton Hessian.

What it means

A model may depend nonlinearly on its parameters even though its fitting objective is a sum of squared discrepancies. Gauss - Newton uses the current parameter values to construct a simpler local approximation of those discrepancies, solves that approximation, and then repeats at the updated values.

The approximation is local. A successful step does not establish that the model is correct, that all observations can be matched exactly or that the final estimate is a global optimum. Starting values, parameter scaling and the shape of the residual functions affect performance.

How it works

Write the objective as phi(x) = one-half the sum of r_i(x) squared. Let r be the residual vector and J its Jacobian, whose entries describe how each residual changes with each parameter.

Approximate r(x + p) by r(x) + Jp. The Gauss - Newton step minimizes the squared length of that linearized residual. With full column rank, it satisfies J transpose J p = −J transpose r. Numerically, solving the least-squares problem with a suitable factorization is preferable to explicitly inverting this matrix.

The exact objective Hessian also contains the sum of r_i times the Hessian of r_i. Gauss - Newton omits that term. The approximation can work especially well near a well-conditioned solution with small residuals. Line searches, damping or trust-region methods can control steps when the local approximation is unreliable.

Three practical examples

1. Sensor calibration: fit nonlinear calibration parameters to measured reference values. Residuals express discrepancies between predicted and observed readings, with weights if measurement precision differs.

2. Curve fitting: estimate parameters of a nonlinear response curve. A linear least-squares solve supplies each update, even though the complete fitting problem remains nonlinear.

3. Geometric estimation: adjust coordinates or pose parameters to reduce discrepancies between predicted and observed features. Derivative accuracy and weakly determined parameter directions can strongly affect the solution.

Think of it this way

Imagine adjusting several controls to bring a collection of gauges toward their targets. At the current setting, estimate how each control changes each gauge. Use those local sensitivities to choose an adjustment, then measure the sensitivities again because the relationship changes as the controls move.

Worked example

Consider an original one-parameter problem with residuals r_1(x) = x squared − 4 and r_2(x) = x − 2. The objective is phi(x) = [(x squared − 4) squared + (x − 2) squared] / 2. Both residuals vanish at x = 2, which therefore gives the global minimum value zero.

Start at x = 1. The residual vector is (−3, −1), and the Jacobian is the column (2, 1). Thus J transpose J = 5 and J transpose r = −7. The step satisfies 5p = 7, so p = 1.4 and the new parameter is 2.4.

At the starting point, the objective is (9 + 1)/2 = 5. At 2.4, the residuals are 1.76 and 0.4, giving objective (3.0976 + 0.16)/2 = 1.6288. The step improves the fit but overshoots the zero-residual solution.

The linearized residual predicted by the first step is (−3 + 2 × 1.4, −1 + 1.4) = (−0.2, 0.4). Its predicted objective is 0.1. The actual objective, 1.6288, is larger because the first residual is curved. This difference illustrates why evaluating the real objective after a step matters.

A second Gauss - Newton step uses Jacobian (4.8, 1). It gives p = −1106/3005, approximately −0.368053, and x becomes 6106/3005, approximately 2.031947. The objective falls to approximately 0.008806.

For comparison, the exact objective Hessian at x = 1 is 5 + 2(−3) = −1. The Gauss - Newton approximation is 5. These are different curvature models, despite using the same objective gradient. Neither should be confused with the residual Jacobian itself.

Illustrative case study

In a fictional calibration project, an engineering team fits a nonlinear response model to reference measurements. A basic Gauss - Newton implementation reduces the objective for one starting estimate but behaves poorly for another.

The team checks the analytical Jacobian against numerical directional differences, rescales the parameters and compares full steps with controlled steps. It records both the predicted improvement from the linearized residual and the actual improvement in the original objective.

The revised workflow makes fitting more dependable, but the team still checks residual patterns on separate measurements. A numerically converged parameter estimate can reveal an inadequate model rather than a finished calibration.

Common misconceptions and mistakes

Gauss - Newton does not make the original model linear. It solves a sequence of local linear least-squares subproblems.

It does not generally use the complete Hessian. The omitted residual-curvature term explains why its behavior can differ from Newton's method.

A full-rank Jacobian can provide a descent direction away from stationarity, but a full step need not decrease the actual objective. Step control and convergence are separate issues.

Small parameter changes alone do not prove a satisfactory fit. Check stationarity, residual size, conditioning and the validity of the model. Rank deficiency requires appropriate handling rather than blind matrix inversion.

Business use

The algorithm is useful when an operational model is naturally fitted by squared residuals and its derivatives can be evaluated reliably. It can reduce the derivative burden compared with full Newton optimization.

Teams should assess repeatability across reasonable starting values, sensitivity to measurement noise and the meaning of the fitted parameters. Report fitting error separately from predictive performance. If residual scales differ substantially, the chosen weighting determines which discrepancies dominate the fit.

Check your understanding

Answer: What is the first step in the example? p = 1.4, taking x from 1 to 2.4.

Why does the predicted objective differ from the actual one?

Answer: The linearized residual omits nonlinear curvature away from the current point.

What extra term does the full Hessian contain?

Answer: The sum of each residual multiplied by its own Hessian.

Related terms

Jacobian
Levenberg - Marquardt algorithm
Line search
Trust region
Nonlinear least squares
Newton's method
parameter estimation

Learn next

Reproduce the two updates and compare a full step with a shortened step. Then examine a problem with nonzero residual at the solution and investigate how Jacobian conditioning affects the fitted parameters.

Jacobian

Sources

David Bindel, Cornell CS 4220, Gauss-Newton and Nonlinear Least Squares, April 10, 2026, Newton and Gauss-Newton discussion: https://www.cs.cornell.edu/courses/cs4220/2026sp/lec/2026-04-10.html

The polynomial calculation and calibration case are original illustrations.

Previous: Gating mechanism Next: Gaussian blur