Local Error-Bound Condition in Optimization

• Local error-bound condition is a quantitative measure estimating the distance from a point to a target set based on a computed residual.
• It is characterized by Lipschitzian or Hölderian bounds and is pivotal in establishing convergence rates in numerical methods.
• These conditions underpin algorithmic stability, sensitivity analysis, and exact penalty formulations across convex and nonconvex optimization frameworks.

Searching arXiv for recent and foundational papers on local error-bound conditions to ground the article. A local error-bound condition is a quantitative estimate that controls distance to a feasible set, solution set, or level set by a residual measuring constraint violation, stationarity, or function-value excess. In its most basic form, it asserts that near a reference point, the distance to the target set is bounded above by a prescribed function of the residual; the most common special cases are Lipschitzian or linear bounds, $\mathrm{dist}\le C\,\mathrm{residual}$, and Hölderian bounds, $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$. Across variational analysis, convex optimization, nonsmooth analysis, rank-constrained problems, fixed-point theory, and algorithmic convergence analysis, local error bounds serve as a bridge between local geometry and quantitative stability. They are closely tied to slope conditions, subdifferential criteria, calmness of set-valued mappings, directional derivatives, and Kurdyka–Łojasiewicz inequalities, and they frequently underwrite linear, sublinear, or even quadratic convergence results for numerical methods (Nguyen, 2017, Cuong et al., 2020, Bi et al., 2016).

1. Core definition and canonical forms

A standard local formulation considers a function $f:X\to\mathbb R\cup\{+\infty\}$ on a metric or Banach space, a level set $[f\le 0]$, and a reference point $\bar x\in[f\le 0]$. In this setting, $f$ admits a local error bound at $\bar x$ if there exist constants such as $\tau>0$ and $\delta>0$ for which

$\tau\,d\bigl(x,[f\le0]\bigr)\le f_+(x)$

whenever $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$0 is sufficiently close to $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$1, where $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$2 (Kruger et al., 2015, Cuong et al., 2020). Equivalent formulations replace $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$3 by a feasible set $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$4, a solution set, or a lower-level set, depending on the problem class (Wei et al., 2024, Nguyen, 2017).

A broader formulation uses a gauge function $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$5 and writes

$\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$6

for $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$7 near a reference point and with $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$8 restricted to a suitable level window (Nguyen, 2017). When $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$9 with $f:X\to\mathbb R\cup\{+\infty\}$0, one obtains a Hölder-type local error bound; the case $f:X\to\mathbb R\cup\{+\infty\}$1 is the Lipschitz or linear case (Nguyen, 2017).

The associated modulus is often defined by a liminf quotient. One representative form is

$f:X\to\mathbb R\cup\{+\infty\}$2

and $f:X\to\mathbb R\cup\{+\infty\}$3 has a local error bound at $f:X\to\mathbb R\cup\{+\infty\}$4 if and only if this quantity is positive (Wei et al., 2024). Closely related moduli appear in Banach-space and convex settings (Kruger et al., 2015, Wei et al., 2021).

Several widely used problem-specific instantiations retain the same structure while changing the residual. In KKT systems for variational problems, the residual is a primal-dual KKT residual $f:X\to\mathbb R\cup\{+\infty\}$5, and the bound takes the form

$f:X\to\mathbb R\cup\{+\infty\}$6

near a KKT point (Kanzow et al., 2018). In rank-constrained optimization, the residual can be the tail singular-value sum $f:X\to\mathbb R\cup\{+\infty\}$7, leading to a bound of the form

$f:X\to\mathbb R\cup\{+\infty\}$8

near the feasible set (Bi et al., 2016). In fixed-point theory, the residual is $f:X\to\mathbb R\cup\{+\infty\}$9, yielding

$[f\le 0]$0

near the fixed-point set (Treek et al., 31 Oct 2025).

This suggests that the local error-bound condition is best understood not as a single inequality, but as a local metric principle: distance to a target set is quantitatively recoverable from a nearby residual.

2. Metric, slope, and subdifferential characterizations

One major line of theory characterizes local error bounds by slope conditions. For a lower-semicontinuous function on a complete metric space, sufficient conditions can be stated through the nonlocal slope $[f\le 0]$1, while in normed or Banach spaces one can use Fréchet or Clarke subdifferential slopes (Cuong et al., 2020). In convex problems these slope notions collapse to the distance from the origin to the convex subdifferential, so local error bounds become equivalent to lower bounds on $[f\le 0]$2 near the reference point (Nguyen, 2017, Cuong et al., 2020).

In the convex l.s.c. setting, a classical criterion uses the strict outer subdifferential slope

$[f\le 0]$3

If this quantity is positive, then a local error bound holds, and in fact

$[f\le 0]$4

The same condition is also necessary (Kruger et al., 2015). This identifies local error bounds with a first-order nondegeneracy property of the subdifferential.

For convex functions on Banach spaces, a more recent primal characterization uses directional derivatives rather than dual-space information. If $[f\le 0]$5 denotes the right-hand directional derivative, then

$[f\le 0]$6

and positivity of the minimum directional derivative at the reference point is exactly the criterion for robustness of the local error bound under small perturbations (Wei et al., 2024). For a single convex inequality in finite dimensions, the condition

$[f\le 0]$7

is equivalent to stability of the local error bound under arbitrarily small perturbations, and the best modulus is its reciprocal (Wei et al., 2021).

For locally Lipschitz and regular functions, the modulus can be compared with quantities derived from the outer limiting subdifferential. In particular, the distance from $[f\le 0]$8 to the outer limiting subdifferential of the support function of $[f\le 0]$9 gives an upper estimate of the local error-bound modulus, and for lower-$\bar x\in[f\le 0]$0 functions equality holds with $\bar x\in[f\le 0]$1 (Li et al., 2016).

This body of results shows that local error bounds are not merely geometric inequalities; they are equivalent, in many settings, to positivity of slope, subdifferential, or directional-derivative quantities near the reference point.

3. Calmness, multifunctions, and variational geometry

A second major perspective formulates local error bounds through calmness of associated set-valued mappings. In rank-constrained optimization, let

$\bar x\in[f\le 0]$2

Then $\bar x\in[f\le 0]$3 is the feasible set, and $\bar x\in[f\le 0]$4 is calm at $\bar x\in[f\le 0]$5 for each $\bar x\in[f\le 0]$6 if and only if there exist constants $\bar x\in[f\le 0]$7 and $\bar x\in[f\le 0]$8 such that

$\bar x\in[f\le 0]$9

In this framework, calmness and the local Lipschitzian error bound are equivalent (Bi et al., 2016).

The same logic appears in perturbed KKT systems. For Banach-space variational problems, calmness of the perturbed KKT system at a reference primal-dual point is equivalent to a residual-type local error bound for distance to the primal variable and multiplier set (Kanzow et al., 2018). Specifically, for small perturbations $f$0, calmness yields

$f$1

and this is equivalent to the local residual estimate

$f$2

for nearby $f$3 (Kanzow et al., 2018).

In mathematical programs with vanishing constraints, the local error-bound property is expressed either through perturbations of an extended feasible system in variables $f$4 or, equivalently, directly in terms of residuals

$f$5

Under MPVC-generalized quasinormality, there exist $f$6 and $f$7 such that

$f$8

for all $f$9 sufficiently near the feasible point (Khare et al., 2018).

These formulations emphasize that the local error-bound condition is often equivalent to metric regularity properties of the constraint system. This suggests a general variational principle: when perturbations in data produce proportionate perturbations in the nearby solution map, a local error bound emerges as the corresponding primal estimate.

4. Lipschitzian versus Hölderian local bounds

The linear or Lipschitzian case is the most familiar, but many problem classes admit only Hölderian bounds. The general local Hölder form

$\bar x$0

appears prominently in semi-algebraic and polynomial settings (Nguyen, 2017, Chen et al., 2 Oct 2025). For the canonical rank-constrained affine feasibility set

$\bar x$1

with residual

$\bar x$2

there exist constants $\bar x$3 and $\bar x$4 such that on any compact set $\bar x$5,

$\bar x$6

where

$\bar x$7

The exponent is explicit and dimension-dependent (Chen et al., 2 Oct 2025).

The following table summarizes representative local forms appearing across the literature.

Setting	Residual	Local bound
Convex inequality system	$\bar x$8	$\bar x$9 (Wei et al., 2021)
Rank-constrained set	$\tau>0$0	$\tau>0$1 (Bi et al., 2016)
Rank-constrained affine set	$\tau>0$2	$\tau>0$3 (Chen et al., 2 Oct 2025)
Averaged fixed-point map	$\tau>0$4	$\tau>0$5 (Treek et al., 31 Oct 2025)
Second-order critical set	$\tau>0$6	$\tau>0$7 (Yue et al., 2018)

The Hölderian case is also central in general error-bound theory. For a lower-semicontinuous function, local error bounds of the form

$\tau>0$8

with $\tau>0$9 arise naturally from Kurdyka–Łojasiewicz inequalities and from polynomial or subanalytic geometry (Nguyen, 2017). In the Moreau-envelope setting, one encounters local Hölder error bounds expressed by the gap $\delta>0$0 or by the distance to lower level sets (Wang et al., 2023).

A plausible implication is that the distinction between Lipschitzian and Hölderian local error bounds reflects the local singularity structure of the target set: tame or nondegenerate geometries often yield exponent $\delta>0$1, while more singular semi-algebraic geometries lead to exponents strictly below $\delta>0$2.

5. Stability under perturbations and robustness criteria

A recurring theme is that the local error-bound property is meaningful only if it survives perturbations. For proper convex l.s.c. functions on Banach spaces, one can define arbitrary, convex, and linear $\delta>0$3-perturbations near $\delta>0$4, and the “radius” of perturbation preserving the local error-bound property is exactly the boundary-subdifferential slope

$\delta>0$5

with the same critical value for arbitrary, convex, and linear perturbation families (Kruger et al., 2015). In particular, if $\delta>0$6, then every perturbation in the corresponding family still admits a local error bound with strictly positive modulus (Kruger et al., 2015).

A closely related directional-derivative criterion was established for convex functions on Banach spaces. If

$\delta>0$7

then there exist $\delta>0$8 and $\delta>0$9 such that every sufficiently small perturbation $\tau\,d\bigl(x,[f\le0]\bigr)\le f_+(x)$0 satisfies $\tau\,d\bigl(x,[f\le0]\bigr)\le f_+(x)$1 (Wei et al., 2024). The same conclusion can be checked using only affine perturbations (Wei et al., 2024). For a single convex inequality, stability under small perturbations is essentially equivalent to the non-zero minimum of the directional derivative over the sphere (Wei et al., 2021).

In the generalized framework of quantitative error bounds, sufficient local conditions can be expressed by nonlocal slopes, Clarke subdifferential slopes, or Fréchet subdifferential slopes, depending on whether the ambient space is metric, Banach, or Asplund (Cuong et al., 2020). The need for completeness and lower semicontinuity in these results highlights that robustness of local error bounds is tied to the existence of nearby minimizers furnished by Ekeland-type principles.

This suggests that local error bounds are best viewed as stable first-order regularity properties rather than isolated inequalities. Their persistence under perturbations is frequently the decisive feature for applications in sensitivity analysis and algorithm design.

6. Relations to identifiability, KL theory, and manifold restrictions

Local error bounds interact strongly with the Kurdyka–Łojasiewicz framework. For a lower-semicontinuous function, a local KL inequality of the form

$\tau\,d\bigl(x,[f\le0]\bigr)\le f_+(x)$2

is equivalent to a nonlinear local error bound involving $\tau\,d\bigl(x,[f\le0]\bigr)\le f_+(x)$3 and the distance to a lower-level set (Nguyen, 2017). In the analytic or subanalytic case, the Łojasiewicz gradient inequality therefore yields a local Hölder error bound directly (Nguyen, 2017). In the unified error-bound framework for randomized algorithms, KL-type conditions are subsumed by a local unified error-bound condition through the choice

$\tau\,d\bigl(x,[f\le0]\bigr)\le f_+(x)$4

(Yang et al., 18 Mar 2026).

A more geometric viewpoint links local error bounds to identifiability. For a proper closed function with minimizer set $\tau\,d\bigl(x,[f\le0]\bigr)\le f_+(x)$5 and an identifiable manifold $\tau\,d\bigl(x,[f\le0]\bigr)\le f_+(x)$6, local ambient-space error bound

$\tau\,d\bigl(x,[f\le0]\bigr)\le f_+(x)$7

around $\tau\,d\bigl(x,[f\le0]\bigr)\le f_+(x)$8 is equivalent to local manifold-restricted error bound

$\tau\,d\bigl(x,[f\le0]\bigr)\le f_+(x)$9

around the same point, under identifiability assumptions (Wu et al., 4 May 2026). In the convex partly smooth setting, the corresponding subdifferential estimate in ambient space is equivalent to a Riemannian-gradient estimate on the manifold (Wu et al., 4 May 2026).

For $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$00-regularized models, strict complementarity yields an identifiable manifold on which the restricted function is smooth; the local manifold error bound is then equivalent to the ambient subdifferential error bound and to the classical proximal error bound used in local linear convergence analyses (Wu et al., 4 May 2026).

This line of work shows that local error bounds are compatible with active-manifold reduction. A plausible implication is that, once identification occurs, the relevant local error-bound constant may be interpreted through the geometry of the smooth restricted problem rather than the full ambient nonsmooth model.

7. Algorithmic consequences and representative applications

Local error bounds frequently convert residual decay into actual distance-to-solution decay, and this is the mechanism behind many local convergence results.

For averaged fixed-point iterations, a local Lipschitz error bound is both necessary and sufficient for local linear convergence. If $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$01 is continuous and $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$02-averaged with nonempty fixed-point set, then

$\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$03

in a neighborhood is equivalent to

$\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$04

for some $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$05 (Treek et al., 31 Oct 2025). Piecewise linear operators satisfy such local error bounds through Hoffman-type arguments, which yields linear convergence for methods such as Douglas–Rachford and ADMM in settings without strong convexity (Treek et al., 31 Oct 2025).

For the cubic regularization method applied to $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$06 nonconvex minimization, the local error-bound condition

$\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$07

near the set of second-order critical points is equivalent, under the stated hypotheses, to a quadratic growth condition, and it implies that the iterates converge at least Q-quadratically to a second-order critical point (Yue et al., 2018). The paper verifies this condition for phase retrieval and low-rank matrix recovery, with overwhelming probability, thereby obtaining Q-quadratic convergence to a global minimizer from arbitrary initialization (Yue et al., 2018).

For inertial forward-backward algorithms in Hilbert space, a local error bound of proximal-gradient type,

$\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$08

when the residual is small and the function value is bounded above, is used to derive improved convergence rates of function values and strong convergence of iterates for several choices of inertial parameters, including the original FISTA and FISTA_CD (2007.07432).

For asynchronous distributed optimization, the Luo–Tseng error-bound condition

$\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$09

under a local neighborhood condition is the key ingredient behind R-linear convergence of the asynchronous algorithm to stationary solutions (Cannelli et al., 2020).

For augmented Lagrangian methods in Banach spaces, a local primal-dual error bound implies Q-linear convergence of KKT residuals and of the primal-dual iterates, and yields boundedness of the penalty-parameter sequence (Kanzow et al., 2018). For the inexact Moreau envelope Lagrangian method, a uniform local error-bound condition on subsets of the feasible set is used to absorb a negative term in the potential-descent estimate and thereby establish an $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$10 complexity bound for finding an $\mathrm{dist}\le C\,\mathrm{residual}^{\theta}$11-KKT point (Huang et al., 27 Feb 2025).

In rank-constrained optimization, local Lipschitzian error bounds have two notable consequences. First, moving the rank constraint into the objective yields an exact penalty formulation once the penalty parameter exceeds a threshold (Bi et al., 2016). Second, error bounds for the iterates of a multi-stage convex relaxation approach can be derived, and the bounds are nonincreasing as the number of stages increases (Bi et al., 2016).

Taken together, these results show that the local error-bound condition is a central regularity hypothesis in modern convergence theory. It is weaker than strong convexity in many settings, yet strong enough to support exact penalization, stability under perturbation, and sharp local convergence estimates ranging from polynomial decay to linear rates and Q-quadratic convergence (Cannelli et al., 2020, Yue et al., 2018, Treek et al., 31 Oct 2025).