Averaged Operators: Theory and Applications

• Averaged operators are mappings defined as T=(1-α)I+αN with α in (0,1) that lie between the identity and nonexpansive maps, ensuring Fejér monotonicity and convergence.
• They are central to fixed-point algorithms and splitting methods by providing exact moduli, sharp residual estimates, and versatile application in proximal and projection methods.
• Their framework extends to variable-metric methods, algebraic averaging in groups and Hopf algebras, and harmonic analysis, unifying diverse areas of research.

In Hilbert-space fixed-point theory, an averaged operator is a mapping of the form

$T=(1-\alpha)I+\alpha N,\qquad \alpha\in(0,1),$

where $N$ is nonexpansive. This representation places averaged operators strictly between the identity and general nonexpansive mappings, with firmly nonexpansive operators appearing as the special case $\alpha=1/2$. The class organizes metric projections, resolvents, proximal mappings, gradient steps, and many splitting operators, and it has become a common language for convergence theory, exact composition constants, and structural generalizations in optimization, metric geometry, and algebra (Simonetto, 2017, Huang et al., 2019, Berdellima, 2020, Zhang et al., 2024).

1. Core definition and basic inequalities

A mapping $T:\mathbb{R}^n\to\mathbb{R}^n$ or, more generally, $T:H\to H$ on a real Hilbert space is nonexpansive if

$\|T(x)-T(y)\|\le \|x-y\|,\qquad \forall x,y.$

It is $\alpha$-averaged if there exists a nonexpansive operator $G$ such that

$T=(1-\alpha)I+\alpha G,\qquad \alpha\in(0,1).$

This implies that $T$ is nonexpansive and that $N$0 and $N$1 share the same fixed points (Simonetto, 2017).

A standard characterization is the inequality

$N$2

valid for all $N$3. If $N$4, this specializes to

$N$5

These inequalities yield Fejér monotonicity and the standard residual estimate

$N$6

for the fixed-point iteration $N$7 when $N$8 (Huang et al., 2019).

Firmly nonexpansive operators coincide with the case $N$9. Resolvents of maximal monotone operators and proximal mappings are firmly nonexpansive; this is the main reason averaged-operator theory interacts so closely with monotone operator splitting and proximal algorithms (Simonetto, 2017).

2. Composition laws, modulus of averagedness, and exact constants

A central problem is to quantify how averaged a given operator is. The Bauschke–Bendit–Moursi modulus of averagedness is the smallest $\alpha=1/2$0 for which a nonexpansive operator is $\alpha=1/2$1-averaged. Recent work shows that this modulus creates a sharp threshold at $\alpha=1/2$2: if an operator is averaged with a constant less than $\alpha=1/2$3, then it is a bi-Lipschitz homeomorphism, while for proximal mappings the condition $\alpha=1/2$4 holds if and only if $\alpha=1/2$5 is Lipschitz smooth (Song et al., 22 Jul 2025).

For compositions, the Ogura–Yamada coefficient plays a distinguished role. If $\alpha=1/2$6 and $\alpha=1/2$7 are $\alpha=1/2$8- and $\alpha=1/2$9-averaged, then $T:\mathbb{R}^n\to\mathbb{R}^n$0 is $T:\mathbb{R}^n\to\mathbb{R}^n$1-averaged with

$T:\mathbb{R}^n\to\mathbb{R}^n$2

Scaled relative graph methods show that this coefficient is tight for the composition class, and the same methodology yields a tight coefficient for Davis–Yin three-operator splitting (Huang et al., 2019).

Exact moduli are known for several basic operators. Song proved that for every nonempty closed convex set $T:\mathbb{R}^n\to\mathbb{R}^n$3, the metric projection $T:\mathbb{R}^n\to\mathbb{R}^n$4 satisfies $T:\mathbb{R}^n\to\mathbb{R}^n$5, sharpening the classical statement that projections are merely firmly nonexpansive. The same work shows that $T:\mathbb{R}^n\to\mathbb{R}^n$6 exactly for translations $T:\mathbb{R}^n\to\mathbb{R}^n$7, and that noninjectivity forces $T:\mathbb{R}^n\to\mathbb{R}^n$8 (Song, 2023). For orthogonal projections onto closed subspaces $T:\mathbb{R}^n\to\mathbb{R}^n$9, the composition $T:H\to H$0 has exact modulus

$T:H\to H$1

where $T:H\to H$2 is the cosine of the Friedrichs angle; this improves the general $T:H\to H$3 bound in finite-dimensional configurations and shows that the Ogura–Yamada bound is sharp in infinite-dimensional worst cases (Bauschke et al., 2023).

Operator or class	Exact modulus / coefficient	Source
Projection $T:H\to H$4, $T:H\to H$5	$T:H\to H$6	(Song, 2023)
Reflector $T:H\to H$7	$T:H\to H$8	(Song, 2023)
Composition $T:H\to H$9	$\|T(x)-T(y)\|\le \|x-y\|,\qquad \forall x,y.$0	(Huang et al., 2019)
$\|T(x)-T(y)\|\le \|x-y\|,\qquad \forall x,y.$1 for subspaces	$\|T(x)-T(y)\|\le \|x-y\|,\qquad \forall x,y.$2	(Bauschke et al., 2023)

These exact constants matter because the residual constant in fixed-point iteration is $\|T(x)-T(y)\|\le \|x-y\|,\qquad \forall x,y.$3. Smaller averagedness coefficients therefore produce strictly sharper $\|T(x)-T(y)\|\le \|x-y\|,\qquad \forall x,y.$4 residual bounds (Huang et al., 2019).

3. Fixed-point algorithms and convex optimization

Averaged operators sit at the center of modern first-order convex optimization because many algorithms can be written as fixed-point iterations of averaged maps. The Krasnosel’skiĭ–Mann scheme

$\|T(x)-T(y)\|\le \|x-y\|,\qquad \forall x,y.$5

reduces to the plain iteration $\|T(x)-T(y)\|\le \|x-y\|,\qquad \forall x,y.$6 when $\|T(x)-T(y)\|\le \|x-y\|,\qquad \forall x,y.$7 itself is already $\|T(x)-T(y)\|\le \|x-y\|,\qquad \forall x,y.$8-averaged and $\|T(x)-T(y)\|\le \|x-y\|,\qquad \forall x,y.$9. Forward–backward splitting, Douglas–Rachford splitting, and ADMM all fit this template (Giselsson et al., 2016).

For smooth $\alpha$0 with $\alpha$1-Lipschitz gradient and convex $\alpha$2, forward–backward splitting

$\alpha$3

is $\alpha$4-averaged with

$\alpha$5

Douglas–Rachford splitting has the form

$\alpha$6

with $\alpha$7, and ADMM appears as Douglas–Rachford splitting on the dual (Giselsson et al., 2016).

This perspective extends directly to time-varying convex optimization. If $\alpha$8 is a sequence of $\alpha$9-averaged operators and $G$0 moves with bounded drift

$G$1

then the running iteration

$G$2

tracks the moving solution set. Under bounded operator images, the mean fixed-point residual satisfies

$G$3

while contractive cases yield direct tracking-error bounds. Running projected gradient, proximal point, forward–backward splitting, Douglas–Rachford splitting, and running ADMM all fall under this framework (Simonetto, 2017).

Residual-based line search can be added without breaking the averaged-operator analysis. A general scheme chooses $G$4 along the residual direction $G$5 and accepts a trial step when

$G$6

For nonexpansive $G$7, residual norms converge; if $G$8, then $G$9 and $T=(1-\alpha)I+\alpha G,\qquad \alpha\in(0,1).$0; if $T=(1-\alpha)I+\alpha G,\qquad \alpha\in(0,1).$1 is contractive, linear convergence follows (Giselsson et al., 2016).

4. Variable-metric, set-valued, and nonlinear-geometric generalizations

One major extension replaces scalar relaxation by a changing metric. In variable-metric algorithms driven by averaged operators, one works with a sequence $T=(1-\alpha)I+\alpha G,\qquad \alpha\in(0,1).$2 of self-adjoint strongly positive operators and studies compositions that are averaged with respect to the $T=(1-\alpha)I+\alpha G,\qquad \alpha\in(0,1).$3-norm. This yields a general variable-metric iteration with errors and relaxation, together with weak convergence and strong-convergence criteria, and it specializes to variable-metric forward–backward and primal–dual splitting (Glaudin, 2018).

A more explicit operator-valued replacement of scalar averaging is the notion of an $T=(1-\alpha)I+\alpha G,\qquad \alpha\in(0,1).$4-averaged operator:

$T=(1-\alpha)I+\alpha G,\qquad \alpha\in(0,1).$5

where $T=(1-\alpha)I+\alpha G,\qquad \alpha\in(0,1).$6 is nonexpansive and $T=(1-\alpha)I+\alpha G,\qquad \alpha\in(0,1).$7 is a self-adjoint linear operator satisfying $T=(1-\alpha)I+\alpha G,\qquad \alpha\in(0,1).$8 with $T=(1-\alpha)I+\alpha G,\qquad \alpha\in(0,1).$9. Such maps are $T$0-averaged in the metric induced by $T$1, and the extended iteration

$T$2

converges to a fixed point under a variable-metric Fejér condition. This formalism is used to analyze semismooth Newton and active-set methods for sparse convex optimization, as well as operator-averaged forward–backward and primal–dual schemes (Simões, 2021).

Set-valued variants arise through union averaged nonexpansive operators. Such an operator has the form

$T$3

where each branch $T$4 is single-valued and averaged, and the selector $T$5 is outer semicontinuous. This class is closed under unions, convex combinations, and compositions. Around strong fixed points, Krasnosel’skiĭ–Mann-type iterations converge locally, and proximal mappings of min-convex functions are union $T$6-averaged. As a result, forward–backward and Douglas–Rachford methods extend to min-convex regularization and union-of-convex feasibility models (Dao et al., 2018).

Beyond Hilbert spaces, averagedness has been reformulated in CAT(0) geometry through $T$7-firmly nonexpansive operators defined with the discrepancy functional

$T$8

The defining inequality is

$T$9

This notion reproduces nonexpansiveness, admits composition and convex-combination rules, and yields weak convergence of Picard iterates for quasi $N$00-firmly nonexpansive maps. Cyclic and averaged projection algorithms in complete CAT(0) spaces are particular cases (Berdellima, 2020).

A further enlargement is the class of $N$01-conically nonexpansive operators, $N$02 with $N$03. In this setting averaged operators correspond to $N$04, and resolvents $N$05 of $N$06-comonotone operators satisfy

$N$07

This connects averagedness, resolvents, cocoercivity, and proximal mappings for hypoconvex functions (Bauschke et al., 2019).

5. Algebraic averaging operators, groups, and Hopf algebras

A distinct algebraic usage of the term concerns operators satisfying multiplicative averaging identities rather than metric nonexpansiveness. In a commutative $N$08-algebra $N$09, an averaging operator is an $N$10-module endomorphism $N$11 such that

$N$12

The image $N$13 is a subalgebra, $N$14, and $N$15 induces a Lie bracket

$N$16

making $N$17 a solvable Lie algebra of derived length $N$18 (Cao, 2014).

Recent work extends averaging from associative and Lie algebras to groups and Hopf algebras. An averaging operator on a group $N$19 is a map $N$20 such that

$N$21

If $N$22, then $N$23 is idempotent and satisfies the pointed identity

$N$24

An averaging group induces a disemigroup via

$N$25

and, when $N$26, a rack via

$N$27

Smooth averaging Lie groups differentiate to averaging Lie algebras, and averaging operators on groups correspond exactly to averaging coalgebra maps on the group Hopf algebra $N$28 by linear extension (Zhang et al., 2024).

The same work defines averaging Hopf algebras by requiring a coalgebra map $N$29 to satisfy

$N$30

It also constructs the free averaging group on a set $N$31 using a normal-form language of averaging words and an inductively defined multiplication and averaging operator with the universal property expected of a free object (Zhang et al., 2024).

These algebraic notions are motivated in part by Koszul duality: averaging operators are presented as Koszul-dual to weight-zero Rota–Baxter operators across associative and Lie settings, and the group/Hopf constructions are designed to preserve that duality pattern (Zhang et al., 2024).

6. Harmonic-analytic and operator-theoretic uses of averaging

In noncommutative harmonic analysis, “differential transforms attached to averaging operators” compare two averaging mechanisms at each scale. On $N$32, with Hardy–Littlewood averages $N$33 and dyadic conditional expectations $N$34, the operators

$N$35

satisfy endpoint and strong-type bounds:

$N$36

and hence

$N$37

Here the “differentiation” is discrete in scale rather than generated by a continuous derivation (Xu, 2021).

A different line of work studies operators obtained by averaging a family $N$38. Because $N$39 is not a Banach space, weak-type $N$40 bounds do not automatically pass to sums or averages. Under uniform weighted endpoint estimates

$N$41

with control through $N$42, the averaged operator $N$43 or $N$44 retains essentially the same weighted weak-type behavior for all $N$45 weights, up to a logarithmic factor in the $N$46 characteristic (Baena-Miret et al., 2023).

In scattering theory on singular spectra, averaged wave operators are formed by Cesàro or Abel–Poisson means of

$N$47

For singular unitary operators $N$48 and bounded $N$49, if $N$50, then the weak limit of the difference of the averaged past and future wave operators is zero; consequently, the averaged past and future wave operators either both exist or both fail to exist, and if they exist, they coincide (Bessonov, 2011).

Across these settings, averaged operators appear in two complementary roles. In fixed-point theory they are structural contractions built from nonexpansive maps; in algebra and analysis they are frequently literal averages over symmetries, scales, or operator families. The common thread is that averaging creates additional rigidity—quantitative in optimization, categorical in algebra, and regularizing in harmonic and spectral analysis.