Global Convergence Criteria

• Global Convergence Criteria are conditions that guarantee optimization algorithms reach global optima even in nonconvex, noisy, or constrained settings.
• They integrate probabilistic, analytical, and geometric methods—such as Markov chain analysis, contraction mappings, and KL inequalities—to ensure algorithmic reliability.
• These criteria inform practical designs in evolutionary computation, deep learning, and nonlinear programming, with extensive experimental validations supporting their effectiveness.

Global convergence criteria delineate the conditions under which stochastic, deterministic, and metaheuristic algorithms are provably guaranteed to locate global optima—even in the presence of nonconvexity, noise, or complex nonlinear constraints. These criteria are foundational in the mathematical analysis and practical design of optimization algorithms ranging from evolutionary computation and neural network training to experimental optimization, derivative-free methods, and constrained nonlinear programming. Rigorous convergence analysis establishes explicit conditions—often expressible in terms of problem structure, update rules, or representation geometry—such that from any initialization, the iterates of an algorithm will, with probability one or along every possible trajectory, reach a globally optimal or stationary point.

1. Probabilistic and Markov Chain Criteria

One central strand is the use of Markov chain theory to analyze population-based algorithms and stochastic global optimization. For evolutionary computation and metaheuristics, sufficiency for global convergence is typically expressed as irreducibility and aperiodicity of the induced stochastic process, together with a positive probability of entering any subset with nonzero measure (Chen et al., 2019, Glasmachers, 2017, Luo et al., 7 May 2025). For example, the Bat Algorithm (BA) is analyzed via a finite homogeneous Markov chain model of the population state, showing that the only closed communication class is the set of states containing a global optimum, and through classical arguments (Meyn–Tweedie) it is proved that the probability of convergence to the global minimizer is asymptotically one.

In evolutionary computation, the Solis–Wets framework and Rudolph's specialization require, for global convergence, that the algorithm revisit every subset of positive measure infinitely often. The Scope and Domain Measure Comparison (SDMC) criterion refines this to a necessary and sufficient measure-theoretic condition: within every finite window of $N$ generations, the union of all individuals’ reachable sets must cover the full search domain. This condition circumvents the limitations of homogeneous Markov modeling and applies even to non-stationary, adaptive algorithms (Luo et al., 7 May 2025).

2. Analytical Contraction and Fixed-Point Conditions

Global convergence is often established via contraction mapping principles in complete metric spaces (Banach’s theorem) or Lyapunov descent arguments. For ensemble learning, as in Negative Correlation Extreme Learning Machine (NCELM), the iterative update is cast as a contraction mapping on a product space of base learners' weight vectors (Perales-González, 2020). The sufficient criterion for global convergence is strictly that the Negative Correlation Learning (NCL) penalty $\lambda$ is small enough to render the aggregate mapping contractive, yielding linear convergence to a unique fixed-point.

Fixed-point iteration and Lyapunov methods extend to projection algorithms for union convex sets. Here, the existence of a descent Lyapunov function for a union upper semicontinuous operator $T$, together with local calmness (Lipschitz-like conditions), ensures global convergence of alternating and averaged projection methods, even under complex feasibility constraints (e.g., sparsity constraints, linear complementarity with $P$-matrices) (Alcantara et al., 2022).

3. Local-to-Global Slope and Metric Conditions

A significant line of research demonstrates that local geometric conditions near the initial point can imply global convergence without global convexity or knowledge of optima. Dello Schiavo et al. generalize the Kurdyka–Łojasiewicz (KL) inequality to the metric space setting, stating that if the product of a parameter function and the descending slope is bounded below (locally) in a ball around the start, gradient flow or proximal point sequences are confined within that ball and converge to a global minimizer (Schiavo et al., 2023). This leverages the chain rule for slopes and energy-dissipation inequalities, providing explicit rates (exponential, polynomial) and guaranteeing global optimality in highly nonconvex or non-smooth landscapes, as long as the slope is sufficiently large on the relevant level set.

4. Polyak–Łojasiewicz-Type Criteria and Block Coordinate Frameworks

Generalized Polyak–Łojasiewicz (PL) and weak-PL conditions state that, for strongly nonconvex objectives, a uniform lower bound on the proximal forcing function (ratio of descent to optimality gap) suffices for global linear or sublinear convergence of arbitrary-block descent methods (Csiba et al., 2017). The grand unification here is the "proportion function," which measures what fraction of possible descent is realized by each block. Strong PL yields linear convergence; weak PL (a local version) and even milder smoothness yield global sublinear rates and near-stationarity in finite time for coordinate/gradient descent, randomized, cyclic, greedy, or adaptive block-selection schemes.

5. Representation and Ordering-Based Criteria in Learning

Global convergence in function-approximation and learning algorithms often depends not merely on approximability, but on the preserved order structure or feature geometry. In policy gradient methods for finite-arm bandits, the necessary and sufficient criterion for Natural Policy Gradient (NPG) is that projection onto the representable subspace preserves the top action's rank ("optimal-action-preservation") (Mei et al., 2 Apr 2025). For vanilla (“softmax”) Policy Gradient, a sufficient condition is "non-domination" of features and full reward-ordering preservation: the representation must realize the full ordering of rewards and have strictly diagonally dominant features. Importantly, approximation error (distance of true reward from its projection) is not an adequate criterion; only ordering matters for convergence.

6. Global Convergence Under Weak Constraint Qualifications

In constrained nonlinear programming, classical constraint qualifications (CQs) such as nondegeneracy (LICQ) and Robinson's CQ (MFCQ) are typically required for global convergence of sequential quadratic programming (SQP) and Lagrangian methods. However, for nonlinear second-order cone programming (NSOCP), the paper (Andreani et al., 2021) introduces weaker CQs, namely constant rank constraint qualification (CRCQ) and constant positive linear dependence (CPLD), formulated in terms of eigenvector-induced linear independence involving only two vectors per cone constraint. These suffice for global convergence: that is, for any sequence of approximate Karush–Kuhn–Tucker (KKT) points (arising from SQP, ALM, etc.), if the limit point satisfies seq-CPLD, it is a KKT point, and multiplier boundedness, metric subregularity, and error bounds follow automatically—without requiring the multipliers to remain bounded a priori.

7. Algorithmic Design, Practicality, and Experimental Validation

Global convergence criteria are not only theoretical constructs but inform and guide practical algorithmic design:

• In experimental optimization, sufficient conditions for feasible-side global convergence (SCFO) rigorously enforce both strict feasibility of every iterate (via Lipschitz bounds and projections) and monotonic decrease of experimental cost (Bunin et al., 2014).
• Power proximal point and augmented Lagrangian methods (ALM) achieve globally optimal rates under inexact updates and non-Euclidean prox terms, with an implicit adaptive penalty schedule dictated by the exponent in the penalty norm (Oikonomidis et al., 2023).
• For deep neural networks in nonconvex regimes, a two-phase training algorithm coupled with an "expressivity condition" at initialization provably yields global convergence as long as the last hidden layer is full-rank after a random perturbation; GC is guaranteed for standard architectures/data and is verifiable numerically (Kawaguchi et al., 2021).
• Recursive Adam-type algorithms for on-line system identification and dynamic RNN training possess global convergence under weak mixing and regularity assumptions, with their average-update trajectories equivalent (in expectation) to scaled stochastic gradient or sign–sign algorithms, as formally shown by ODE averaging (Ljung’s method) (Wigren et al., 6 Jan 2025).

Direct experimental validations, benchmark tests, and practical parameter selection underscore the predictive and practical value of these criteria, confirming that only algorithms and settings conforming to the precise structural or representation condition enjoy rapid global convergence in practice (Chen et al., 2019, Luo et al., 7 May 2025, Kawaguchi et al., 2021, Wigren et al., 6 Jan 2025).

Summary Table: Representative Global Convergence Criteria

Algorithm/Domain	Sufficient Criterion	Foundational Property
Evolutionary/Metaheuristics	SDMC covering measure (H5), irreducibility	Markov chain theory, measure
Extreme Learning Machine (NCELM)	Contractive mapping on ensemble weights	Banach fixed-point theorem
Proximal/Projection Methods	Lyapunov function, upper-semicontinuous maps	Fixed-point/descending sequence
SGD, Adam	Stepsize & noise control, Lyapunov stability	ODE averaging, martingale descent
Policy Gradients (PG/NPG)	Top-rank preservation, ordering, non-domination	Representation geometry
NSOCP (SQP/ALM)	seq-CPLD/seq-CRCQ weaker CQ	Eigenvector linear independence
Deep Learning (2-phase)	Expressivity condition, rank of hidden layer	Real-analytic rank, convex reduction
Gradient Flow/Proximal Metric	Local slope × function monotonicity near start	Metric KL/Kurdyka–Łojasiewicz

These global convergence criteria reflect an overview of probabilistic, analytic, geometric, and structural properties tailored to the problem and algorithm, making explicit the boundary between regimes where thermodynamically or algorithmically rapid convergence is mathematically predictable and those where only local or heuristic progress can be guaranteed.