Compact R-Continuity Overview

• Compact R-continuity is a local quantitative continuity property for set-valued mappings in Banach spaces, controlling truncated residuals using a modulus.
• It relaxes full R-continuity by applying estimates within any fixed compact set, making it valuable for convergence analysis and stability in nonsmooth algorithms.
• Its finite-dimensional characterization via graph closedness and its role in residual-based error bounds provide actionable insights for algorithm design.

Searching arXiv for recent and directly relevant papers on compact R-continuity and closely related continuity/compactness notions. arXiv search query: "compact R-continuity set-valued mappings Banach spaces"

Compact R-continuity is a local quantitative continuity property for set-valued mappings $\mathcal A:\mathbb X\rightrightarrows\mathbb Y$ between Banach spaces. At a reference point $\bar x$, it requires that, inside any fixed compact region $K\subset\mathbb Y$, the truncated values $\mathcal A(x)\cap K$ remain within a modulus-controlled enlargement of $\mathcal A(\bar x)$ when $x$ is close to $\bar x$. In current arXiv usage, the notion is studied as a relaxation of R-continuity, aimed at robustness of solution residuals in inclusions and at convergence analysis for nonsmooth algorithms, while remaining closely tied in finite dimensions to closedness of the graph (Le et al., 2 Sep 2025).

1. Definition and quantitative form

Let $\mathcal A:\mathbb X\rightrightarrows\mathbb Y$ be a set-valued map between Banach spaces, and let $\bar x\in\mathbb X$ with $\mathcal A(\bar x)\neq\emptyset$. R-continuity at $\bar x$0 means that there exist a radius $\bar x$1 and a nondecreasing continuity modulus $\bar x$2 such that $\bar x$3, $\bar x$4 as $\bar x$5, and

$\bar x$6

Equivalently,

$\bar x$7

where $\bar x$8 is the excess (Le et al., 2 Sep 2025).

Compact R-continuity at $\bar x$9 weakens this by imposing the estimate only after truncation by an arbitrary compact subset of the range: for every compact $K\subset\mathbb Y$0, there exist $K\subset\mathbb Y$1 and a nondecreasing $K\subset\mathbb Y$2 with $K\subset\mathbb Y$3 and $K\subset\mathbb Y$4 as $K\subset\mathbb Y$5 such that

$K\subset\mathbb Y$6

Equivalently,

$K\subset\mathbb Y$7

Notion	Estimate	Scope
R-continuity	$K\subset\mathbb Y$8	All values
Compact R-continuity	$K\subset\mathbb Y$9	Values inside a fixed compact $\mathcal A(x)\cap K$0
R-Lipschitz continuity	$\mathcal A(x)\cap K$1	Linear modulus
R-Hölder continuity	$\mathcal A(x)\cap K$2	Power modulus

This formalism is explicitly one-sided. It controls how far values at nearby points can move away from $\mathcal A(x)\cap K$3, but it does not impose a symmetric Hausdorff-type estimate. The paper therefore places it between stronger inverse-mapping notions such as metric regularity or the Aubin property and weaker residual notions such as calmness (Le et al., 2 Sep 2025).

2. Structural properties and finite-dimensional characterization

A central result is that compact R-continuity is strictly weaker than full R-continuity. The standard example is the map $\mathcal A(x)\cap K$4 defined by

$\mathcal A(x)\cap K$5

For any fixed compact $\mathcal A(x)\cap K$6, the branch $\mathcal A(x)\cap K$7 eventually leaves $\mathcal A(x)\cap K$8 as $\mathcal A(x)\cap K$9, so $\mathcal A(\bar x)$0 is compactly R-Lipschitz at $\mathcal A(\bar x)$1. It is not R-continuous at $\mathcal A(\bar x)$2, because no modulus can control the unbounded branch $\mathcal A(\bar x)$3 relative to $\mathcal A(\bar x)$4 (Le et al., 2 Sep 2025).

In finite dimensions, compact R-continuity is characterized by closedness of the graph at the reference point. If $\mathcal A(\bar x)$5 and $\mathcal A(\bar x)$6 is closed, then

$\mathcal A(\bar x)$7

where graph closedness at $\mathcal A(\bar x)$8 means that $\mathcal A(\bar x)$9 and $x$0 with $x$1 imply $x$2 (Le et al., 2 Sep 2025).

This equivalence has two immediate consequences. First, in finite dimensions compact R-continuity is often inexpensive to verify, because graph closedness is standard for many subdifferential and monotone-operator constructions. Second, once graph closedness is combined with local boundedness of the values, compact R-continuity upgrades to full R-continuity. A plausible implication is that compact R-continuity isolates precisely the amount of continuity needed in bounded algorithmic regimes, without requiring global control of all branches of the graph.

The notion is therefore neither purely topological nor fully metric-regular. Its defining estimate retains a modulus $x$3, but the compact truncation makes that modulus compatible with set-valued maps whose distant values may be uncontrolled or even unbounded.

3. Analytic mechanisms: inverse maps, full rank, and Łojasiewicz inequalities

The paper connects compact R-continuity to classical analytic inequalities through inverse solution maps. Let $x$4 be $x$5, $x$6, and consider

$x$7

If $x$8 has full rank on the solution set $x$9, then $\bar x$0 is compactly R-Lipschitz at $\bar x$1. Under the additional hypothesis that $\bar x$2 is R-continuous at $\bar x$3, the map is actually R-Lipschitz there (Le et al., 2 Sep 2025).

For scalar analytic $\bar x$4 with $\bar x$5, the classical Łojasiewicz inequality states that for every compact $\bar x$6 there exist $\bar x$7 such that

$\bar x$8

This yields compact R-Hölder continuity of the inverse map $\bar x$9 at $\mathcal A:\mathbb X\rightrightarrows\mathbb Y$0, because any $\mathcal A:\mathbb X\rightrightarrows\mathbb Y$1 satisfies

$\mathcal A:\mathbb X\rightrightarrows\mathbb Y$2

hence

$\mathcal A:\mathbb X\rightrightarrows\mathbb Y$3

Thus the inverse solution mapping inherits a compact modulus directly from the analytic geometry of the zero set (Le et al., 2 Sep 2025).

The reverse implication does not hold in general. The non-analytic $\mathcal A:\mathbb X\rightrightarrows\mathbb Y$4 function

$\mathcal A:\mathbb X\rightrightarrows\mathbb Y$5

does not satisfy a Łojasiewicz inequality, yet its inverse has full R-continuity at $\mathcal A:\mathbb X\rightrightarrows\mathbb Y$6 because its graph is closed and locally bounded (Le et al., 2 Sep 2025). This distinguishes compact R-continuity from the usual analytic or subanalytic Łojasiewicz framework: the former is a direct property of the inverse set-valued map, not an inequality imposed on an objective function.

4. Inclusions, residual stability, and Hoffman-type interpretation

The primary application concerns inclusions

$\mathcal A:\mathbb X\rightrightarrows\mathbb Y$7

with solution set

$\mathcal A:\mathbb X\rightrightarrows\mathbb Y$8

Here compact R-continuity is used on the inverse mapping $\mathcal A:\mathbb X\rightrightarrows\mathbb Y$9, not on $\bar x\in\mathbb X$0 itself. If $\bar x\in\mathbb X$1 are residuals with $\bar x\in\mathbb X$2, then R-continuity of $\bar x\in\mathbb X$3 at $\bar x\in\mathbb X$4 implies

$\bar x\in\mathbb X$5

hence

$\bar x\in\mathbb X$6

If only compact R-continuity is available, the same conclusion holds provided $\bar x\in\mathbb X$7 is bounded, so that all iterates lie in a compact set $\bar x\in\mathbb X$8 and one can use

$\bar x\in\mathbb X$9

instead (Le et al., 2 Sep 2025).

This gives a Hoffman-type stability principle: approximate solutions with small residuals are close to the exact solution set. The framework is weaker than metric regularity or the Aubin property, because it does not require inverse estimates uniform in both domain and range neighborhoods. It is stronger than calmness, because it controls all points in $\mathcal A(\bar x)\neq\emptyset$0, at least inside any fixed compact region (Le et al., 2 Sep 2025).

The resulting viewpoint is residual-centric. Instead of deriving convergence from a full error-bound inequality or a KL/PL inequality, one asks only whether vanishing residuals force vanishing distance to the solution set. Compact R-continuity answers that question positively on bounded sets.

5. Algorithmic role: R-class schemes, proximal methods, and descent dynamics

The paper formalizes this residual viewpoint through the class of R-class algorithms. An iterative scheme $\mathcal A(\bar x)\neq\emptyset$1 belongs to the R-class if there exist a function $\mathcal A(\bar x)\neq\emptyset$2 with $\mathcal A(\bar x)\neq\emptyset$3 and constants $\mathcal A(\bar x)\neq\emptyset$4 such that for every $\mathcal A(\bar x)\neq\emptyset$5 there exists

$\mathcal A(\bar x)\neq\emptyset$6

The basic convergence theorem states that if $\mathcal A(\bar x)\neq\emptyset$7, then R-continuity of $\mathcal A(\bar x)\neq\emptyset$8 at $\mathcal A(\bar x)\neq\emptyset$9 implies $\bar x$00, and boundedness of $\bar x$01 plus compact R-continuity of $\bar x$02 at $\bar x$03 implies the same conclusion (Le et al., 2 Sep 2025).

Several standard algorithms fit this template. For the generalized proximal scheme

$\bar x$04

the stationarity condition yields

$\bar x$05

with

$\bar x$06

Thus $\bar x$07 and $\bar x$08 give an R-class estimate (Le et al., 2 Sep 2025).

For the Proximal Point Algorithm

$\bar x$09

one has

$\bar x$10

so

$\bar x$11

For maximally monotone $\bar x$12, the graph is closed, hence $\bar x$13 is compactly R-continuous at zero; boundedness of the iterates and vanishing increments then give $\bar x$14, and in fact convergence of the whole sequence to a point in $\bar x$15 (Le et al., 2 Sep 2025).

The same pattern is applied to nonsmooth optimization under the Attouch–Bolte–Svaiter-type conditions

$\bar x$16

and

$\bar x$17

or the analogous variant with $\bar x$18. Gradient descent, proximal algorithms for convex $\bar x$19, Boosted DCA, and a DCA variant with smooth $\bar x$20 are treated in this way. The output is not merely that cluster points are stationary, but that the full iterate sequence satisfies

$\bar x$21

under boundedness and graph closedness of the subgradient mapping (Le et al., 2 Sep 2025).

6. Related notions, distinctions, and scope

The phrase “compact R-continuity” should be distinguished from several older compactness-based continuity formalisms. In the theory of compact-preserving functions, a map $\bar x$22 is called compact-preserving if $\bar x$23 is compact for every compact $\bar x$24; on strong Fréchet spaces this admits a local characterization, and on locally connected strong Fréchet spaces compact-preserving plus the Darboux property characterizes ordinary continuity (Banakh et al., 2012). For polynomials on normed spaces, continuity is equivalent to mapping compact subsets into compact subsets (Gámez-Merino et al., 2011). These are image-preservation principles for single-valued maps, not modulus-controlled inclusion estimates for set-valued inverse problems.

A second nearby line of work studies generalized continuity and compactness through sequence methods. Upward and downward half quasi-Cauchy continuity preserve one-sided quasi-Cauchy behavior of sequences, and the associated compactness notions require every sequence in a set to admit a corresponding half quasi-Cauchy subsequence (Cakalli, 2012). This suggests a structural analogy with compact R-continuity, but the underlying primitives are sequence classes rather than excess estimates for set-valued maps.

A third comparison comes from approach theory. On the space of cumulative distribution functions, the continuity approach structure has a relative sequential compactness index equal to an escape index measuring tail mass, yielding a quantitative Prokhorov theorem (Berckmoes, 2015). Here again compactness and continuity are quantified, but the ambient object is an approach space rather than a set-valued map between Banach spaces.

A further terminological source of ambiguity is Hausdorff continuity on hyperspaces of compact sets. Continuous convexity measures on compact subsets of $\bar x$25 are defined on $\bar x$26 or on the non-singleton subspace $\bar x$27, with continuity understood as $\bar x$28 in Hausdorff distance implying $\bar x$29 (Douzal et al., 2023). This is a continuity theory on spaces of compact sets, not the compact-truncated residual continuity studied for set-valued inclusions.

Within its own framework, compact R-continuity has clear current limitations. The theory in (Le et al., 2 Sep 2025) is mostly finite-dimensional; extending it to infinite-dimensional Banach settings is nontrivial because compact subsets are much more restrictive there, and local boundedness no longer implies local compactness. The results are mainly qualitative, establishing $\bar x$30 rather than rates. The paper also identifies as an open direction a systematic understanding of when PLK or KŁ inequalities imply R- or compact R-continuity of inverse subgradient maps (Le et al., 2 Sep 2025).

Compact R-continuity therefore occupies a specific position in modern variational analysis: it is a residual-based, compact-range quantitative continuity notion for set-valued maps, weaker than full R-continuity, often equivalent in finite dimensions to graph closedness, and strong enough to convert vanishing residuals into convergence to solution sets for broad classes of inclusions and algorithms (Le et al., 2 Sep 2025).