Must the model be quadratic?

No. Quadratic models are common, but other local models can be used within suitable trust-region frameworks.

Does a good ratio always enlarge the radius?

Exact policies vary, and some also check whether the step reached the region boundary.

Can the region handle an indefinite Hessian?

Yes. Bounding the step makes the quadratic subproblem bounded on a compact region, with appropriate algorithms handling negative curvature.

AI Glossary / Mathematics and optimization / Intermediate

Trust region

Definition

A trust region is a neighborhood around the current point within which an optimization algorithm uses a local model to propose a step. The algorithm compares the model's predicted improvement with the improvement actually observed, then adjusts the region and decides whether to accept the step. The region expresses a computational limit on model-based movement, not statistical confidence.

What it means

An objective function may be complicated, while a local linear or quadratic approximation is easier to optimize. That approximation can be informative nearby and misleading farther away. Restricting the proposed step helps control this mismatch.

The radius is adaptive. Good agreement can justify a larger region; poor agreement suggests a smaller one. The algorithm also uses its acceptance rule to decide whether the current point changes. Shrinking the region and accepting a point are distinct decisions, so a rejected trial can still provide useful information about model reliability.

How it works

Build a model m(p), often f(x)+gᵀp+(1/2)pᵀBp, around the current point x. Choose a step that sufficiently improves this model subject to a restriction such as ‖p‖≤Δ. Scaling can change the shape of the region to account for different variable units.

Evaluate the true objective at x+p. When predicted reduction is positive, form the ratio ρ=[f(x)−f(x+p)]/[m(0)−m(p)]. A ratio near one indicates good agreement in reduction; a negative ratio means the trial worsened the true objective despite a predicted improvement.

Update the radius and acceptance decision according to the chosen algorithm. The subproblem may be solved approximately, and the model can use exact, approximate, or sampled curvature information. Convergence guarantees depend on model quality, sufficient-decrease conditions, and other assumptions; the word “trust” is not itself a guarantee.

Three practical examples

1. Nonlinear parameter estimation: restrict steps when a local model of fitting error is unreliable far from current parameter values.

2. Large-scale smooth optimization: combine a trust-region framework with iterative quadratic-model solves that avoid costly exact factorizations.

3. Derivative-free optimization: optimize a local surrogate within a region where available function samples support its accuracy, updating that region after new evaluations.

Think of it this way

Think of navigating with a detailed local sketch rather than a complete map. You choose a nearby destination, check whether the actual journey matches the sketch, and decide how far to rely on the next sketch.

The analogy does not mean the region is a physical safe zone or that all points within it are feasible. It limits the use of a mathematical approximation; operational constraints and safety requirements must be represented and checked separately.

Worked example

Consider an original objective f(x)=x⁴−2x² at x=0. Here f(0)=0, f′(0)=0, and f″(0)=−4. The quadratic model is m(p)=−2p², which decreases without bound if no step restriction is imposed.

With trust-region radius Δ=2, a model minimizer is p=2. The predicted reduction is m(0)−m(2)=8. But the true objective is f(2)=16−8=8, so actual reduction is 0−8=−8. The ratio is ρ=−1. A normal positive-threshold acceptance rule rejects this worsening trial and reduces the radius.

Suppose the new radius is 1. A model minimizer p=1 predicts reduction 2, while f(1)=1−2=−1 gives actual reduction 1. Thus ρ=0.5. Under an illustrative acceptance threshold of 0.1, this trial is accepted.

The symmetric negative steps behave identically. Radius reduction has transformed an unhelpful large proposal into a useful smaller one. The exact radius-update factor is a stated teaching choice, not a universal rule. These hand calculations also show how negative curvature can matter even when the gradient is zero.

Illustrative case study

Fictional case: Alder Calibration fits a nonlinear model whose early quadratic approximations are poor. A large proposed parameter change predicts a strong reduction but actually increases the fitting error.

The optimizer rejects that trial, shrinks its region, and evaluates a more local step. Analysts retain both predicted and actual reductions in their diagnostics, rather than labeling the rejected evaluation as meaningless. They compare final fit and evaluation cost with another method. This scenario illustrates an algorithmic response without claiming measured performance or guaranteed global optimality.

Common misconceptions and mistakes

A trust region is not a confidence interval and does not quantify the probability that a parameter lies inside it. Nor is it automatically the feasible set of the original problem.

It differs from an ordinary line search, which typically chooses a direction and then varies distance along it. A trust-region subproblem can change the direction as its radius changes.

Do not divide by zero or a nonpositive predicted reduction without appropriate handling. Also, small accepted steps do not alone prove success: the model may be poor, the radius may have collapsed, or stopping conditions may need investigation.

Business use

Trust-region methods help manage the tension between an inexpensive local approximation and a more costly true objective. Their usefulness depends on obtaining sufficient progress without wasting evaluations on unsuitable trials.

Monitor actual and predicted reduction, radius changes, gradient or stationarity measures, and constraint violations. Scale variables meaningfully so one coordinate does not dominate the step norm merely because of units. Report the algorithm's convergence status and local nature where applicable; an adaptive region does not turn a nonconvex problem into a globally solved one.

Check your understanding

Why is the radius-2 trial rejected in the worked example even though it minimizes the local model?

Answer: The model predicts improvement 8, but the true objective worsens by 8. The negative agreement ratio exposes the model's poor prediction at that distance.

Related terms

Local model
Line search
Derivative-free optimization
Newton's method
quadratic subproblem
actual reduction
predicted reduction
negative curvature
variable scaling

Learn next

Recalculate the worked ratio for radius 0.5 and compare it with radii 1 and 2. Then examine a two-variable quadratic model with different coordinate scales. Study sufficient model decrease and approximate subproblem solvers before assuming that every implementation must solve the trust-region subproblem exactly.

Sources

Primary source: NEOS Guide, “Trust-Region Methods”, https://neos-guide.org/guide/algorithms/trust-region/ . The subproblem and acceptance-ratio formulas were read. The supplied Probabilistic Machine Learning book download could not be retrieved, so its exact passage was not checked. The quartic example is independently derived and does not reproduce a specific solver trace.

Previous: Trust Next: Trust region policy optimization