---
title: Levitin-Polyak Well-Posedness in Optimization
url: https://www.emergentmind.com/topics/levitin-polyak-well-posedness
type: topic
---

# Levitin-Polyak Well-Posedness in Optimization

Searching arXiv for recent papers on Levitin-Polyak well-posedness and closely related optimization/VI formulations.
Levitin-Polyak well-posedness is a notion of stability for optimization, variational inequality, and equilibrium-type problems that requires asymptotically optimal approximate solutions to converge, or at least to possess convergent subsequences whose limits are exact solutions. In the materials considered here, the notion appears in several technically distinct but structurally related forms: as a subsequential compactness property for convex bilevel optimization [2509.18304], as convergence of generalized approximating sequences for elliptic variational inequalities [2309.04805], as a perturbation-stable metric property for split equilibrium problems [2208.07126] and split multivalued variational inequalities [2311.18060], and as an extended or infeasibility-measure-based concept for constrained optimization in infinite-dimensional spaces [2508.16462]. Across these settings, the common theme is that near-feasibility together with near-optimality, or the analogue appropriate to the problem class, is sufficient to recover exact solvability in a strong asymptotic sense.

## 1. Conceptual core and basic definitions

In convex bilevel optimization, the problem has the form
\[
\begin{aligned}
\text{(Inner problem):} \quad & g^* := \min_{x \in X} g(x) \\
\text{(Outer problem):} \quad & f^* := \min_{x \in X_g} f(x), \quad \text{where } X_g := \arg\min_{z \in X} g(z),
\end{aligned}
\]
where \(X\) is a closed convex set, and \(f,g:X\to\mathbb{R}\) are convex [2509.18304]. If \(X^* := \arg\min_{x\in X_g} f(x)\), then three notions are distinguished.

**LP Well-Posedness:** For any sequence \(\{x_t\}\subset X\) with \(\operatorname{Dist}(x_t,X_g)\to 0\) and \(f(x_t)\to f^*\), there exists a subsequence \(\{x_{t_k}\}\) and \(\bar x\in X^*\) such that \(x_{t_k}\to \bar x\) [2509.18304].

**Generalized LP Well-Posedness:** For any sequence \(\{x_t\}\subset X\) with \(f(x_t)\to f^*\) and \(g(x_t)\to g^*\), there exists a subsequence converging to some \(\bar x\in X^*\) [2509.18304].

**Strong Generalized LP Well-Posedness:** For any sequence \(\{x_t\}\subset X\) with \(\limsup_{t\to\infty} f(x_t)\le f^*\) and \(g(x_t)\to g^*\), there exists a subsequence converging to some \(\bar x\in X^*\) [2509.18304].

In constrained optimization in infinite-dimensional spaces, the extended formulation is stated for
\[
\min f(x)\qquad \text{subject to}\quad x\in M,\quad x\in Q,
\]
with feasible region \(\Omega=M\cap Q\) [2508.16462]. The problem is Levitin-Polyak well-posed in the extended sense if \(\argmin(\mathcal P)\) is nonempty and every sequence \(\{x_n\}\subset X\) satisfying \(f(x_n)\to f_*\) and \(\dist(x_n,\Omega)\to 0\) has a subsequence converging to some \(x_*\in\argmin(\mathcal P)\) [2508.16462]. The weak version requires only \(\dist(x_n,\argmin(\mathcal P))\to 0\) [2508.16462].

A further generalization replaces the distance-to-feasibility by an infeasibility measure \(\varphi:X\to[0,\infty]\), with \(\varphi(x)=0\Leftrightarrow x\) feasible. Then Levitin-Polyak well-posedness with respect to \(\varphi\) requires \(\argmin(\mathcal P)\neq\emptyset\) and that any sequence with \(f(x_n)\to f_*\) and \(\varphi(x_n)\to 0\) has a subsequence converging to some \(x_*\in\argmin(\mathcal P)\) [2508.16462]. This generalizes both the extended Levitin-Polyak and Tykhonov versions when \(\varphi(x)=\dist(x,\Omega)\) for closed \(\Omega\) [2508.16462].

These definitions show that Levitin-Polyak well-posedness is not a single formal condition but a family of asymptotic compactness requirements adapted to the native approximation model of each problem class. This suggests that its unifying content is best understood through the types of approximating sequences it controls.

## 2. Approximating sequences and convergence mechanisms

For elliptic variational inequalities in a real Hilbert space \(X\), the problem is to find \(u\in K\) such that
\[
(Au,v-u)_x + j(v) - j(u) \ge (f,v-u)_x \qquad \forall v\in K,
\]
where \(K\subset X\) is nonempty, closed, and convex, \(A:X\to X\) is strongly monotone and Lipschitz, and \(j:X\to\mathbb R\) is convex and locally Lipschitz [2309.04805]. A sequence \(\{u_n\}\subset X\) is a Levitin-Polyak generalized approximating sequence if there exist \(\{w_n\}\subset X\) with \(w_n\to 0\), and \(0<\epsilon_n\to 0\), such that
\[
u_n+w_n\in K,
\]
and
\[
(Au_n,v-u_n)_x + j(v)-j(u_n) + \epsilon_n |v-u_n|_x \ge (f,v-u_n)_x \qquad \forall v\in K
\]
[2309.04805]. The variational inequality is well-posed in the sense of Levitin-Polyak if every such approximating sequence converges strongly in \(X\) to the unique solution \(u\) [2309.04805].

The same paper formulates a more general necessary and sufficient convergence criterion. If \(u\) is the unique solution, then for any sequence \(\{u_n\}\subset X\),
\[
u_n\to u \text{ in } X
\]
if and only if there exist \(0<\epsilon_n\to 0\) such that both
\[
d(u_n,K)\to 0
\]
and
\[
\forall v\in K:\quad (Au_n,v-u_n)_x + j(v)-j(u_n)+\epsilon_n(1+|v-u_n|_x)\ge (f,v-u_n)_x
\]
hold [2309.04805]. The classical LP-approximating sequence is a special case of this criterion, because \(u_n+w_n\in K\) with \(w_n\to 0\) implies \(d(u_n,K)\le |w_n|_x\to 0\) [2309.04805]. Corollary 17 therefore yields Levitin-Polyak well-posedness under the standard assumptions [2309.04805].

In split multivalued variational inequalities, generalized approximating sequences are similarly built from controlled infeasibility and vanishing residual inequalities. If \(S\) is the solution set, the problem is LP well-posed when \(S\) is a singleton and every generalized approximating sequence converges to that unique solution; it is LP well-posed in the generalized sense when \(S\) is nonempty and every generalized approximating sequence has a subsequence converging to some element of \(S\) [2311.18060]. In split equilibrium problems under perturbations, generalized approximating sequences likewise allow small infeasibility and perturbed equilibrium inequalities, and LP well-posedness by perturbations requires convergence of every such sequence to the unique solution when perturbations vanish [2208.07126].

A recurring structural feature is that Levitin-Polyak well-posedness is formulated for inexact trajectories rather than exact iterates. In that respect it is a natural notion for iterative computation, perturbation analysis, and asymptotic stability.

## 3. Convex bilevel optimization: duality, compactness, and LP well-posedness

The convex bilevel setting in "Convergence, Duality and Well-Posedness in Convex Bilevel Optimization" makes the role of Levitin-Polyak well-posedness particularly explicit [2509.18304]. The paper identifies two challenges: strong duality is not guaranteed due to the lack of Slater constraint qualification, and convergence of algorithms is not guaranteed even when usual suboptimality gap bounds are present, because super-optimal solutions can occur [2509.18304].

The central duality statement is Theorem 3.1:
\[
\boxed{
\text{Strong duality holds for (1.3)} \iff
\begin{array}{l}
\text{Whenever } \limsup_{t\to\infty} f(x_t)\le f^*,\ g(x_t)\to g^*, \\
\text{then } \lim_{t\to\infty} f(x_t)=f^*.
\end{array}
}
\]
Thus strong duality for the Lagrangian dual of the value function formulation is equivalent to the claim that asymptotically vanishing primal suboptimality in the sense \(\limsup f(x_t)\le f^*\) and \(g(x_t)\to g^*\) forces exact convergence of outer function values, excluding persistent super-optimality [2509.18304].

Two sufficient conditions are then separated. Condition 4.1 is the proximity or metric subregularity property
\[
\text{If } \{x_t\}\subset X \text{ with } g(x_t)\to g^*, \text{ then } \operatorname{Dist}(x_t,X_g)\to 0.
\]
This is an error bound on \(g\), connected to metric subregularity or error bounds, and it guarantees strong duality and hence \(f(x_t)\to f^*\) [2509.18304]. However, it does not guarantee convergence in distance to the optimal solution set \(X^*\); the paper gives an example where sequences converge in function value but do not approach \(X^*\) in distance [2509.18304].

Condition 4.2 is compactness:
\[
X^*=\arg\min_{x\in X_g} f(x)\ \text{ is compact.}
\]
This includes cases where \(X\) is bounded, or when a coercivity condition holds on \(f\) or on \(\bar h:=\max\{f-f^*,g-g^*\}\) [2509.18304]. Under this condition, strong duality holds and convergence in both function value and distance to \(X^*\) is ensured [2509.18304].

Proposition 4.4 then establishes a precise equivalence: compactness of \(X^*\), generalized LP well-posedness, and strong generalized LP well-posedness are equivalent; standard LP well-posedness also implies compactness [2509.18304]. This is one of the clearest available statements of the relation between LP well-posedness and the geometry of the optimal set. In this setting, LP well-posedness is not merely a qualitative stability label; it is equivalent to compactness of the solution set and underwrites convergence beyond function values.

Corollary 4.5 sharpens this point quantitatively. Under compactness, there exists a continuous function \(c:\mathbb R_+^2\to\mathbb R_+\) with \(c(0,0)=0\) such that
\[
|f(x)-f^*| \ge c\left(\operatorname{Dist}(x,X^*),\ g(x)-g^*\right)
\qquad \forall x\in X.
\]
Hence any sequence with \(f(x_t)\to f^*\) and \(g(x_t)\to g^*\) must also satisfy \(\operatorname{Dist}(x_t,X^*)\to 0\) [2509.18304]. This is a modulus-type error bound derived from generalized LP well-posedness via Theorem 2.2 of Huang and Yang, 2006, as cited in the paper [2509.18304].

## 4. Metric characterizations and perturbation frameworks

In split equilibrium problems in real Banach spaces, LP well-posedness is characterized through approximate solution sets [2208.07126]. If \(S\) denotes the solution set and \(S(\epsilon)\) the \(\epsilon\)-approximate solution set, then Theorem 1 states that the split equilibrium problem is LP well-posed by perturbations if and only if \(S\) is nonempty and
\[
\operatorname{diam}(S(\epsilon))\to 0 \quad \text{as } \epsilon\to 0
\]
[2208.07126]. Under upper semicontinuity assumptions on the perturbed bifunctions, this is equivalently stated as
\[
S(\epsilon)\neq\emptyset\ \forall \epsilon>0,\qquad \operatorname{diam}(S(\epsilon))\to 0
\]
[2208.07126].

For generalized LP well-posedness by perturbations, the relevant metric criterion is
\[
H(S(\epsilon),S)\to 0 \quad \text{as } \epsilon\to 0,
\]
together with nonempty compact \(S\), where \(H(\cdot,\cdot)\) denotes the Hausdorff metric [2208.07126]. In finite-dimensional settings, generalized LP well-posedness is equivalent to
\[
S(\epsilon)\neq \emptyset\ \forall \epsilon>0,\qquad \mu(S(\epsilon))\to 0,
\]
where \(\mu\) is the Kuratowski measure of noncompactness [2208.07126]. Theorem 5 further states that in finite-dimensional Banach spaces, under suitable monotonicity, continuity, and convexity assumptions, LP well-posedness by perturbations is equivalent to existence and uniqueness of the solution [2208.07126].

The split multivalued variational inequality setting has closely parallel metric characterizations. Let \(S(\epsilon)\) denote the \(\epsilon\)-approximate solution set. Then the problem is LP well-posed if and only if
\[
S(\epsilon)\neq \emptyset\ \forall \epsilon>0,\qquad \operatorname{diam}(S(\epsilon))\to 0 \quad \text{as } \epsilon\to 0
\]
[2311.18060]. Generalized LP well-posedness is equivalent to
\[
S(\epsilon)\neq \emptyset\ \forall \epsilon>0,\qquad \mu(S(\epsilon))\to 0 \quad \text{as } \epsilon\to 0
\]
[2311.18060]. In finite-dimensional Hilbert spaces, if \(S(\epsilon)\) is nonempty and bounded for some \(\epsilon>0\), then LP well-posedness is equivalent to uniqueness of the solution, while generalized LP well-posedness is equivalent to nonemptiness of the solution set [2311.18060].

These metric characterizations expose a common logic: LP well-posedness is equivalent to collapse of approximate solution sets. In the singleton case the collapse is measured by vanishing diameter; in the nonunique case it is measured by Hausdorff convergence to a compact solution set or by vanishing measure of noncompactness. This suggests a general interpretation of LP well-posedness as an asymptotic concentration principle for approximate solutions.

## 5. Relations to Tykhonov well-posedness and extended variants

The constrained optimization framework in infinite-dimensional spaces explicitly compares Tykhonov and Levitin-Polyak notions [2508.16462]. Tykhonov well-posedness is the analogous property restricted to feasible sequences \(x_n\in\Omega\), whereas Levitin-Polyak well-posedness allows approximate feasibility through \(\dist(x_n,\Omega)\to 0\) [2508.16462]. The paper states that Levitin-Polyak well-posedness generalizes Tykhonov well-posedness in this sense, and that if the feasible set \(\Omega\) is bounded and \(f\) is uniformly continuous on bounded sets, the two notions are equivalent [2508.16462].

The same source introduces the infeasibility-measure-based extension. The implication chain is summarized as
\[
\text{Levitin-Polyak WP (w.r.t. \(\varphi\))} \implies
\text{Levitin-Polyak WP (extended)} \implies
\text{Tykhonov WP (extended)},
\]
but not conversely; specifically, Tykhonov well-posedness does not imply Levitin-Polyak well-posedness [2508.16462].

A further structural property appears in Proposition 2 and Proposition 6 of that paper: Levitin-Polyak well-posedness in the extended sense is equivalent to the weak version plus compactness of \(\argmin(\mathcal P)\), and the same is true for the \(\varphi\)-based extension [2508.16462]. This is closely aligned with the convex bilevel result that generalized and strong generalized LP well-posedness are equivalent to compactness of the optimal set \(X^*\) [2509.18304]. Across these settings, compactness is not incidental; it is the mechanism that upgrades weak asymptotic closeness to subsequential convergence.

The elliptic variational inequality paper adds a distinct perspective. There, the authors introduce T-well-posedness through T-approximating sequences satisfying
\[
d(u_n,K)<\epsilon_n,\quad
(Au_n,v-u_n)_x + j(v)-j(u_n)+\epsilon_n(1+|v-u_n|_x)\ge (f,v-u_n)_x\quad \forall v\in K,
\]
and show that this coincides exactly with the necessary and sufficient condition of Theorem 8 [2309.04805]. In that treatment, classical LP and Tykhonov notions are only sufficient, whereas the new criterion is necessary and sufficient [2309.04805]. A plausible implication is that, in some variational inequality settings, LP well-posedness is best viewed as one member of a larger taxonomy of approximation-sensitive convergence notions rather than as the maximal possible characterization.

## 6. Applications and broader significance

The source materials attach LP well-posedness to several applied and algorithmic contexts. In elliptic variational inequalities, two applications are highlighted: a heat transfer problem and an elastic frictionless contact problem [2309.04805]. In the contact setting, the general criterion is used to show that solutions to perturbed approximating problems converge strongly to the solution of the original frictionless problem [2309.04805]. In the heat transfer setting, approximating the boundary condition yields a family of problems whose solutions converge to the original as prescribed by the general criterion, and LP well-posedness is guaranteed [2309.04805].

In split equilibrium problems, the metric characterizations imply that any iterative process whose outputs form generalized approximating sequences converges, or has convergent subsequences, to true solutions under LP well-posedness [2208.07126]. The paper explicitly interprets this as convergence and stability under perturbations from discretization, errors, or modeling [2208.07126]. Similarly, in split multivalued variational inequalities, LP well-posedness is specialized to split feasibility, split minimization, split variational inequalities, and split mixed variational inequalities [2311.18060]. The theory is designed to cover exact and approximate computation in Hilbert spaces with multivalued operators and split constraints [2311.18060].

In constrained optimization, LP well-posedness with respect to an infeasibility measure \(\varphi\) is used to formulate conditions for exact penalty functions and global saddle points of augmented Lagrangians [2508.16462]. If the problem is weakly Levitin-Polyak well-posed with respect to \(\varphi\), then the penalty function
\[
F_c(x)=f(x)+c\varphi(x)
\]
is globally exact if and only if three conditions hold: zero duality gap,
\[
\sup_{c\ge 0}\inf_{x\in Q} F_c(x)=f_*,
\]
existence of global minimizers of \(F_{c_0}\) over \(Q\) for all \(c\ge c_0\), and uniform local exactness at global minimizers [2508.16462]. Under additional assumptions, analogous localization principles characterize the global exactness of penalty functions and the existence of global saddle points of augmented Lagrangians [2508.16462]. Here LP well-posedness functions as a stability hypothesis that allows local information to propagate to global conclusions in infinite-dimensional settings.

## 7. Common themes, limitations, and points of interpretation

Several common themes recur across the cited problem classes.

| Theme | Manifestation | Source |
|---|---|---|
| Approximate feasibility | Measured by \(\dist(x_n,\Omega)\to 0\), \(d(u_n,K)\to 0\), or analogous split constraints | [2508.16462], [2309.04805], [2208.07126], [2311.18060] |
| Approximate optimality or residual control | Function values approach optima or residual inequalities vanish | [2509.18304], [2309.04805], [2208.07126], [2311.18060] |
| Compactness as decisive property | Compactness of optimal/solution set is equivalent to generalized LP variants in several settings | [2509.18304], [2508.16462] |
| Metric characterization | Vanishing diameter, Hausdorff convergence, or vanishing noncompactness of approximate solution sets | [2208.07126], [2311.18060] |

A central limitation emphasized in convex bilevel optimization is that strong duality and convergence of function values do not by themselves guarantee convergence to the optimal solution set in distance [2509.18304]. Condition 4.1 ensures strong duality and \(f(x_t)\to f^*\), but not \(\operatorname{Dist}(x_t,X^*)\to 0\) [2509.18304]. Only the stronger compactness-based condition yields both. This addresses a common misconception that asymptotically vanishing primal gaps are sufficient for full convergence of iterates in bilevel problems; the paper explicitly exhibits failure due to super-optimal solutions and lack of approach to \(X^*\) [2509.18304].

Another limitation is dimensional. For split equilibrium problems, the equivalence between LP well-posedness and uniqueness is fully established in finite-dimensional Banach spaces; for infinite-dimensional spaces, the paper states that this remains an open question [2208.07126]. This indicates that some of the neat metric equivalences rely materially on finite-dimensional compactness mechanisms.

The overall picture is that Levitin-Polyak well-posedness is a convergence-stability concept for inexact solution processes. Its strongest formulations combine three ingredients: approximate feasibility, approximate optimality or residual decay, and compactness of the exact solution set. In convex bilevel optimization, these ingredients connect directly to strong duality and distance convergence [2509.18304]. In variational inequalities and split equilibrium models, they yield necessary and sufficient metric convergence criteria and perturbation-stable solution behavior [2309.04805; 2208.07126; 2311.18060]. In constrained optimization in infinite-dimensional spaces, extended and \(\varphi\)-based versions support exact penalization and augmented Lagrangian theory [2508.16462]. Taken together, these results position Levitin-Polyak well-posedness as a unifying stability principle across contemporary nonlinear analysis and optimization.

Source: https://www.emergentmind.com/topics/levitin-polyak-well-posedness