Zarantonello Fixed-Point Iteration

• Zarantonello fixed-point iteration is a relaxed fixed-point scheme in Hilbert spaces that uses adaptive step sizes to drive convergence towards fixed points.
• It generalizes classical projection methods to include firmly nonexpansive mappings, cutter operators, and cyclic compositions for solving common fixed-point problems.
• The method extends to adaptive iterative linearized Galerkin schemes and is benchmarked against modern acceleration techniques like optimized Halpern anchoring for improved convergence rates.

Zarantonello fixed-point iteration is the Hilbert-space scheme

$$x_{k+1}=x_k+\theta_k(Tx_k-x_k), \qquad k=0,1,2,\dots,$$

for an operator $T:H\to H$ with nonempty fixed point set $$\operatorname{Fix}T=\{x\in H\mid Tx=x\}\neq\emptyset$$. In modern fixed-point terminology it is often called a relaxed fixed-point iteration. Equivalently, with

$T_{\theta_k}:=(1-\theta_k)I+\theta_k T,$

one has $x_{k+1}=T_{\theta_k}x_k$. The method is classical in the setting of metric projections onto closed convex sets, but later work places it in a much broader framework that includes firmly nonexpansive mappings, resolvents of maximal monotone operators, subgradient projectors, cutter operators, cyclic compositions, and adaptive iterative linearized Galerkin schemes for strongly monotone nonlinear equations (Cegielski et al., 2012, Heid et al., 2019).

1. Canonical formulation and classical projection setting

Let $H$ be a real Hilbert space with inner product $\langle\cdot,\cdot\rangle$ and norm $\|\cdot\|$. The iteration

$x_{k+1}=x_k+\theta_k(Tx_k-x_k)$

moves from $x_k$ in the direction of the fixed-point residual $T:H\to H$0, scaled by a relaxation parameter $T:H\to H$1. In the classical projection setting, $T:H\to H$2 is the metric projection $T:H\to H$3 onto a nonempty, closed and convex set $T:H\to H$4, so that

$T:H\to H$5

Zarantonello’s original work focuses on metric projections. The crucial structural fact is that $T:H\to H$6 is firmly nonexpansive: $T:H\to H$7 In particular, $T:H\to H$8 is nonexpansive and quasi-nonexpansive, and its fixed point set is exactly $T:H\to H$9. This places the iteration in a setting where the relaxed operator inherits strong metric regularity properties.

The interval $$\operatorname{Fix}T=\{x\in H\mid Tx=x\}\neq\emptyset$$0 is characteristic for the relaxation parameter. For many classes of operators considered in the modern literature, including cutters and firmly nonexpansive maps, constant relaxation $$\operatorname{Fix}T=\{x\in H\mid Tx=x\}\neq\emptyset$$1 leads to convergence under suitable assumptions, and variable relaxation $$\operatorname{Fix}T=\{x\in H\mid Tx=x\}\neq\emptyset$$2 for some $$\operatorname{Fix}T=\{x\in H\mid Tx=x\}\neq\emptyset$$3 also fits the standard convergence framework (Cegielski et al., 2012). A common misconception is that the method is specific to orthogonal projection. In fact, the same template appears for nonexpansive mappings with nonempty fixed point set, firmly nonexpansive mappings, cutter operators, and projection-type operators such as subgradient projectors.

2. Geometric operator classes and the cutter framework

A central generalization replaces the single projection operator by a cutter. For $$\operatorname{Fix}T=\{x\in H\mid Tx=x\}\neq\emptyset$$4, define the half-space

$$\operatorname{Fix}T=\{x\in H\mid Tx=x\}\neq\emptyset$$5

An operator $$\operatorname{Fix}T=\{x\in H\mid Tx=x\}\neq\emptyset$$6 is called a cutter if

$$\operatorname{Fix}T=\{x\in H\mid Tx=x\}\neq\emptyset$$7

equivalently, if

$$\operatorname{Fix}T=\{x\in H\mid Tx=x\}\neq\emptyset$$8

or, in rewritten form,

$$\operatorname{Fix}T=\{x\in H\mid Tx=x\}\neq\emptyset$$9

Geometrically, the hyperplane through $T_{\theta_k}:=(1-\theta_k)I+\theta_k T,$0 with normal $T_{\theta_k}:=(1-\theta_k)I+\theta_k T,$1 cuts the space into two half-spaces, containing $T_{\theta_k}:=(1-\theta_k)I+\theta_k T,$2 and $T_{\theta_k}:=(1-\theta_k)I+\theta_k T,$3 on opposite sides.

This framework unifies several operator classes. Orthogonal projections $T_{\theta_k}:=(1-\theta_k)I+\theta_k T,$4, resolvents of maximal monotone operators, subgradient projectors, and firmly nonexpansive mappings are all cutters. If $T_{\theta_k}:=(1-\theta_k)I+\theta_k T,$5, then $T_{\theta_k}:=(1-\theta_k)I+\theta_k T,$6 is closed and convex. Every cutter is quasi-nonexpansive: $T_{\theta_k}:=(1-\theta_k)I+\theta_k T,$7

Within this language, the Zarantonello iteration is simply the usual relaxation of a cutter. If one defines

$T_{\theta_k}:=(1-\theta_k)I+\theta_k T,$8

then setting $T_{\theta_k}:=(1-\theta_k)I+\theta_k T,$9 yields

$x_{k+1}=T_{\theta_k}x_k$0

which is exactly the relaxed operator underlying Zarantonello iteration. Thus the classical scheme is a special case of a generalized relaxation in which the step can depend on the current state (Cegielski et al., 2012).

3. Cyclic compositions, extrapolation, and local acceleration

The cutter framework is particularly useful for common fixed point problems. Given a finite family of cutter operators $x_{k+1}=T_{\theta_k}x_k$1 with nonempty common fixed point set

$x_{k+1}=T_{\theta_k}x_k$2

one may form the cyclic composition

$x_{k+1}=T_{\theta_k}x_k$3

The basic cyclic scheme is then the Picard iteration

$x_{k+1}=T_{\theta_k}x_k$4

which reduces to the classical cyclic projection method when $x_{k+1}=T_{\theta_k}x_k$5.

The generalized Zarantonello-type scheme for the composite operator takes the form

$x_{k+1}=T_{\theta_k}x_k$6

with $x_{k+1}=T_{\theta_k}x_k$7 and a step size function $x_{k+1}=T_{\theta_k}x_k$8. When $x_{k+1}=T_{\theta_k}x_k$9 outside $H$0, the paper terms this an extrapolation of $H$1: the step along $H$2 is larger than in the basic Zarantonello iteration.

A distinguished choice is the maximal step size

$H$3

where $H$4 and $H$5. This choice is constructed so that $H$6 is again a cutter. In this way, the classical single-operator iteration extends systematically to composite operators, state-dependent step sizes, and local acceleration strategies of Dos Santos type (Cegielski et al., 2012).

The Dos Santos principle is line-search-like but remains within the fixed-point geometry. For projection methods onto hyperplanes or related feasibility operators, the accelerated point on the line $H$7 is selected by a problem-specific formula for $H$8. The paper emphasizes that this acceleration is local: each step is optimally adjusted along the current line, but global speed-up is not guaranteed a priori.

4. Convergence mechanisms

For a cutter $H$9, the relaxed operator

$\langle\cdot,\cdot\rangle$0

is strongly quasi-nonexpansive for $\langle\cdot,\cdot\rangle$1. This property is the analytical basis for convergence of Zarantonello-type iterations in the cutter setting (Cegielski et al., 2012).

For cyclic composites, the central convergence statement is expressed for

$\langle\cdot,\cdot\rangle$2

Under the common fixed point assumption and the cutter hypotheses on the factors $\langle\cdot,\cdot\rangle$3, one has the estimate

$\langle\cdot,\cdot\rangle$4

This yields Fejér monotonicity with respect to $\langle\cdot,\cdot\rangle$5, and in particular $\langle\cdot,\cdot\rangle$6.

The demiclosedness assumptions identify the weak limit. If $\langle\cdot,\cdot\rangle$7 is demiclosed at $\langle\cdot,\cdot\rangle$8, then every weak cluster point belongs to $\langle\cdot,\cdot\rangle$9, and Fejér arguments give weak convergence of the whole sequence to a point in $\|\cdot\|$0. If each $\|\cdot\|$1 is demiclosed at $\|\cdot\|$2, then the limit lies in the common fixed point set

$\|\cdot\|$3

These results generalize the single-operator picture in two directions. First, they allow families of operators treated through their composition. Second, they allow extrapolation parameters $\|\cdot\|$4 that may exceed $\|\cdot\|$5, provided the induced operator remains strongly quasi-nonexpansive. The standard restriction $\|\cdot\|$6 therefore survives in a broadened form rather than disappearing.

5. Strongly monotone equations and adaptive iterative linearized Galerkin schemes

A second modern realization of Zarantonello iteration arises in the solution of nonlinear operator equations

$\|\cdot\|$7

on a real Hilbert space $\|\cdot\|$8, where $\|\cdot\|$9 is Lipschitz continuous and strongly monotone. In the unified iterative linearization framework, one chooses $x_{k+1}=x_k+\theta_k(Tx_k-x_k)$0, or more precisely $x_{k+1}=x_k+\theta_k(Tx_k-x_k)$1 with $x_{k+1}=x_k+\theta_k(Tx_k-x_k)$2 the Riesz map. The iteration then becomes

$x_{k+1}=x_k+\theta_k(Tx_k-x_k)$3

and in weak form

$x_{k+1}=x_k+\theta_k(Tx_k-x_k)$4

This is exactly the Zarantonello fixed-point iteration in that framework (Heid et al., 2019).

The paper assumes Lipschitz continuity with constant $x_{k+1}=x_k+\theta_k(Tx_k-x_k)$5, strong monotonicity with constant $x_{k+1}=x_k+\theta_k(Tx_k-x_k)$6, and the existence of a Gâteaux differentiable potential $x_{k+1}=x_k+\theta_k(Tx_k-x_k)$7 with $x_{k+1}=x_k+\theta_k(Tx_k-x_k)$8. For Zarantonello iteration, the energy decrease condition holds for any

$x_{k+1}=x_k+\theta_k(Tx_k-x_k)$9

Under these assumptions, the general convergence theory yields convergence of $x_k$0 to the unique solution $x_k$1.

The same framework is coupled to Galerkin discretization. On a finite-dimensional subspace $x_k$2, the discrete Zarantonello iteration reads

$x_k$3

Adaptive iterative linearized Galerkin algorithms monitor the linearization error

$x_k$4

against a posteriori estimator values

$x_k$5

and iterate on the current space while $x_k$6. When $x_k$7, the space is enriched and the process repeats. In the finite element applications studied in the paper, this yields convergence $x_k$8 and linear decay of a combined energy-estimator quantity. The numerical comparisons reported there place Zarantonello alongside Kačanov and Newton iterations: Zarantonello is globally convergent under the monotonicity assumptions and fits naturally into adaptive finite element discretizations, but it is slower in iteration count than Newton when high accuracy is sought (Heid et al., 2019).

6. Optimal acceleration and related modern interpretations

Recent complexity theory places classical relaxed fixed-point methods in a sharper hierarchy. For nonexpansive and contractive operators, optimized Halpern-type anchoring yields exact optimal worst-case rates for the squared fixed-point residual $x_k$9. In the nonexpansive case, the optimized Halpern rate is

$T:H\to H$00

while in the contractive case the optimal rate is of order $T:H\to H$01, and for maximal $T:H\to H$02-strongly monotone inclusions the corresponding optimal proximal-point residual rate is of order $T:H\to H$03 (Park et al., 2022).

This has a direct consequence for the interpretation of classical Zarantonello-type iterations. Relaxed Picard, Krasnosel’skiĭ–Mann, projection, and resolvent schemes belong to the broad class of deterministic fixed-point methods driven by evaluations of a nonexpansive or averaged operator. In the abstract residual metric used in the complexity analysis, classical relaxed projection and Mann-type schemes with $T:H\to H$04 residual behavior are therefore strictly suboptimal, whereas optimized Halpern-type anchoring is optimal (Park et al., 2022). This does not invalidate the classical method; it locates it as a baseline iteration whose convergence guarantees are robust but not complexity-optimal in the black-box model.

A related but terminologically distinct interpretation appears for linear systems. The paper “Zur iterativen Loesung von linearen Gleichungssystemen” studies a fixed-point iteration

$T:H\to H$05

and modifies it to

$T:H\to H$06

so that the new iteration matrix has spectral radius less than one, even when the original iteration diverges. That paper does not use the name “Zarantonello,” and presents the method through state feedback, eigenvalue placement, and Riccati equations; the connection is conceptual rather than terminological (Karl et al., 2013). It nevertheless illustrates a recurrent theme in the modern reading of Zarantonello’s iteration: a problematic fixed-point map is altered by a structured relaxation or feedback mechanism so that convergence becomes enforceable.

Taken together, these developments suggest a stable taxonomy. Zarantonello fixed-point iteration is the basic relaxation process

$T:H\to H$07

for projection, nonexpansive, firmly nonexpansive, cutter, and monotone-operator settings. Cyclic composition and extrapolation extend it to common fixed point problems; adaptive ILG places it in PDE discretization; and optimized Halpern-type methods define the current complexity frontier against which the classical scheme is measured.