---
title: Trust-Region Funnel Method
url: https://www.emergentmind.com/topics/trust-region-funnel-method
type: topic
---

# Trust-Region Funnel Method

The trust-region funnel method is a globalization and acceptance mechanism for solving nonlinearly constrained optimization problems, especially prominent in sequential quadratic programming (SQP) frameworks and fully derivative-free, grey- or black-box settings. The defining attribute is a monotonically nonincreasing scalar “funnel” bound on infeasibility, which replaces classical filter methods’ multi-entry constraints with a uni-dimensional acceptance threshold. This structure enforces that all iterates remain within a shrinking feasible region, channeling the optimization process toward constraint satisfaction while enabling efficient objective reduction. The approach achieves global convergence guarantees to either Karush–Kuhn–Tucker (KKT) points or infeasible stationary points under standard constraint qualification assumptions, with demonstrated numerical and theoretical efficiency on both academic and grey/black-box test sets [2409.09208, 2511.18998, 1707.00337, 1912.12637].

## 1. Problem Formulation and Fundamentals

The trust-region funnel method addresses nonlinear programs of the form
\[
\begin{array}{ll}
\text{minimize} & f(x) \\
\text{subject to} & c(x) = 0, \quad x \geq 0
\end{array}
\]
where $f: \mathbb{R}^n \to \mathbb{R}$ is the objective and $c: \mathbb{R}^n \to \mathbb{R}^m$ enforces equality constraints; general inequalities $l \leq c(x) \leq u$ are handled via slack variables. The funnel method maintains a scalar bounding sequence $\eta_k > 0$ such that the infeasibility measure $v(x) := \|c(x)\|_1$ satisfies $v(x^k) \leq \eta_k$ for all iterates. The bound $\eta_k$ is initialized so that the starting point lies within the funnel, typically $\eta_0 = \max\{\bar{\tau}, \bar{\kappa} v(x^0)\}$ for $\bar{\tau}>0, \bar{\kappa}>1$, and is updated in a monotonic (nonincreasing) fashion [2409.09208].

In the context of equality-constrained and grey-box problems, the infeasibility measure may take different forms—for example, $v(x) = \frac12\|c(x)\|^2$ [1707.00337], or the approximation error $\theta(x) = \|r^k(w) - t(w)\|$ for local reduced models in grey-box settings [2511.18998].

## 2. Funnel Acceptance Mechanism and Switching Rules

Distinct from combined merit functions or multi-entry filters, the funnel method enforces a scalar one-sided infeasibility constraint, codified as $v(x^{(k)} + d) \leq \eta_k$ (the “funnel acceptability” condition). After each trial step, a switching rule determines whether to treat the iteration primarily as an optimality (f-type) or feasibility (h-/v-/θ-type) advancement.

Acceptance logic in the trust-region funnel SQP [2409.09208, 2511.18998]:
- **f-type step**: If the predicted reduction in the quadratic model of $f$ ($\Delta m_f^{(k)}(d)$) satisfies a threshold, i.e., $\Delta m_f^{(k)}(d) \geq \delta [v(x^{(k)})]^2$ for $\delta \in (0,1)$, and the Armijo condition $f(x^{(k)}) - f(x^{(k)}+d) \geq \sigma \Delta m_f^{(k)}(d)$ for $\sigma \in (0,1)$, the step is accepted and the funnel bound remains unchanged: $\eta_{k+1} = \eta_k$.
- **h-type (or θ-type) step**: If the feasibility is sufficiently improved $v(x^{(k)}+d) \leq \beta \eta_k$ for $\beta \in (0,1)$, the step is accepted and the funnel shrinks: $\eta_{k+1} = (1-\kappa) v(x^{(k)}+d) + \kappa \eta_k$, $\kappa \in (0,1)$.
- **Rejection**: If neither condition is met, the trial step is rejected and the trust region is contracted.

This mechanism maintains the invariant $v(x^{(k)}) \leq \eta_k$ across all iterates, ensuring convergence toward feasibility. The “switching condition” prevents endless feasibility-only progress, forcing steps to be accepted only when the objective or the constraints improve sufficiently [2409.09208, 1707.00337].

## 3. Trust-Region Subproblems and Restoration

Each iteration solves a trust-region subproblem, generally of the form:
\[
\begin{array}{ll}
\text{minimize}_d & \frac12 d^T B_k d + \nabla f(x^{(k)})^T d \\
\text{subject to} & c(x^{(k)}) + \nabla c(x^{(k)})^T d = 0 \\
                  & x^{(k)} + d \geq 0 \\
                  & \|d\|_\infty \leq \Delta^{(k,\ell)}
\end{array}
\]
where $B_k$ is an approximation to $\nabla^2_{xx} L(x^{(k)}, \lambda^{(k)})$, and $\Delta^{(k,\ell)}$ is the current trust-region radius. Subproblems may be indefinite; the trust region bounds negative curvature.

If the quadratic program (QP) is infeasible, the algorithm enters a **restoration phase**: it ignores objective descent and minimizes a measure of violation (e.g., inelastic formulation with elastic variables for feasibility), ensuring that the linearized constraints can again be satisfied. Restoration is continued until feasibility is restored or local infeasibility is detected [2409.09208, 2511.18998, 1912.12637].

Trust-region radii are dynamically updated: expanded if the step is accepted and the trial step is on the boundary, contracted otherwise. In some frameworks, radii for objective-progress and feasibility-progress are updated separately, yielding an efficient balance [1912.12637].

## 4. Algorithmic Structure and Pseudocode

The trust-region funnel algorithm generally exhibits a double-loop structure: an outer loop iterating over major iterates, and an inner loop adjusting the trust-region size until a funnel-acceptable step is found. A high-level pseudocode [2409.09208] is:

```
Input: x⁰, multipliers, parameters (σ, δ, β, κ, τ̄, κ̄, Δ₀)
Initialize: η₀ = max{τ̄, κ̄·v(x⁰)}, phase ← Optimality, k←0
While not converged do
   ℓ←0;  Δ←Δ₀
   Repeat
     If phase=Optimality:
        Solve TR-QP (possibly enter restoration if infeasible)
     Else:
        Solve FQP for feasibility
     Let x_trial = xᵏ + d
     Accept step if funnel/optimality criteria satisfied; update η, Δ
     If not accepted, contract Δ
   Until accept
   k←k+1
Return x*, multipliers
```
This core structure admits various enhancements, such as two-phase approaches (first attain $\epsilon$-feasibility, then optimize over the feasible manifold) [1707.00337], surrogate modeling for black-box constraints [1912.12637], and flexible restoration mechanisms.

## 5. Convergence Theory and Complexity

Under standard regularity conditions—compactness, $C^2$-smoothness, bounded Hessians, Mangasarian–Fromovitz constraint qualification (MFCQ)—the trust-region funnel method is globally convergent. Specifically:
- The funnel rules force the infeasibility $v(x^k) \to 0$ as $k \to \infty$.
- Any accumulation point $x^*$ is either feasible and KKT, or a Fritz John point if MFCQ fails [2409.09208, 2511.18998].
- Iteration complexity is $O(\epsilon^{-3/2})$ for $\epsilon$-feasibility or $\epsilon$-stationarity [1707.00337]. Surrogate-based methods in black/grey-box settings typically have first-order rates $O(\epsilon^{-2})$ under fully linear models [1912.12637].
- The approach can be generalized: for higher-order models, improved complexity bounds (e.g., $O(\epsilon^{-4/3})$) are attainable [1707.00337].

## 6. Practical Performance and Implementation

The trust-region funnel approach has been implemented in multiple open-source frameworks:
- **Uno** (C++) [2409.09208] unifies several nonlinearly constrained optimization solvers and supports both line-search and trust-region funnel globalization.
- **DEFT-FUNNEL** [1912.12637] targets constrained black- and grey-box problems with polynomial surrogate models, adapting the funnel acceptance to model-based objective and constraint evaluations.
- **TRF-Solver** (Pyomo) [2511.18998] supports analytic derivatives for glass-box parts, with multiple local reduced-model (RM) options for black-box evaluation. Benchmarking against classical filter methods on grey-box engineering benchmarks, the trust-region funnel algorithm matched or outperformed filter-based variants in terms of black-box evaluation cost and robustness.

Representative results highlight:
- For academic CUTEst problems, trust-region funnel SQP requires fewer constraint/Hessian evaluations than filter-based counterparts and avoids the need for convexification of indefinite QPs [2409.09208].
- In black- and grey-box settings, the funnel method accelerates “funneling” toward feasible regions, often saving expensive function evaluations, especially with low-order surrogates [2511.18998, 1912.12637].
- High-fidelity (GP, Taylor) surrogates yield similar performance across funnel/filter methods; efficiency gains for the funnel are most pronounced for lower-fidelity or derivative-free contexts.

## 7. Relationships to Other Globalization Methods and Extensions

The trust-region funnel method generalizes and, in certain contexts, supersedes classical filter methods by providing a simpler acceptance mechanism (single infeasibility threshold) rather than maintaining a history of dominated pairs. While error-free behaviors of filter and funnel methods are similar for most test cases, the funnel occasionally admits greater nonmonotonicity in infeasibility, which can improve robustness but may exacerbate second-order convergence issues due to Maratos phenomena (remediable via second-order corrections) [2409.09208].

Two-phase extensions (e.g., first funneling to the feasible set, then optimizing on the feasible manifold) can reduce overall complexity and improve efficiency, especially if the first phase incorporates both feasibility and objective progress [1707.00337]. Surrogate-based and grey-box extensions further broaden applicability to domains where analytic derivatives are unavailable or only partially available, relying on polynomial or Gaussian process local modeling coupled with the funnel mechanism for global convergence [2511.18998, 1912.12637].

Source: https://www.emergentmind.com/topics/trust-region-funnel-method