Strong Variational Sufficiency

• Strong Variational Sufficiency is a local optimality condition defined by the variational strong convexity that aligns the subgradient graph of a function with that of a strongly convex surrogate.
• It offers a complete second-order characterization through Moreau-envelope convexity, positive-definiteness of second-order subdifferentials, and generalized Hessian analysis.
• The framework underpins stability, tilt stability, and rapid convergence of augmented Lagrangian methods in diverse settings including composite, nonpolyhedral, and semidefinite programs.

Strong variational sufficiency is a sufficient condition for local optimality formulated through variational strong convexity of a function or of an augmented composite model. For a lower semicontinuous function $f$, the strong variational sufficient condition at $\bar x$ with modulus $\sigma>0$ means that $0\in\partial f(\bar x)$ and $f$ is variationally strongly convex at $\bar x$ for $0$; for composite optimization, at a KKT pair $(\bar x,\bar u)$ of $\min_x f_0(x)+g(F(x))$, it means that there exists $\sigma>0$ such that

$\bar x$0

is variationally strongly convex at $\bar x$1 (Khanh et al., 2022, Mordukhovich et al., 13 Jul 2025). In current variational analysis, the notion is treated as a primal–dual sufficiency property: locally, the subgradient mapping behaves like that of a strongly convex function. It has been characterized by Moreau-envelope convexity, positive-definiteness of second-order subdifferentials, strong second-order sufficient conditions, and positive-definiteness of generalized Hessians of augmented Lagrangians in finite-dimensional, infinite-dimensional, semidefinite, and manifold settings (Wang et al., 2022, Zhou et al., 2023, Khanh et al., 2023).

1. Variational formulation and basic definition

The modern notion originates in Rockafellar’s theory of variational convexity and variational strong convexity. For an l.s.c. function $\bar x$2, a point $\bar x$3, and $\bar x$4, $\bar x$5 is variationally strongly convex at $\bar x$6 for $\bar x$7 if there exist a convex neighborhood $\bar x$8 of $\bar x$9, an l.s.c. strongly convex function $\sigma>0$0 on $\sigma>0$1, and $\sigma>0$2 such that

$\sigma>0$3

with

$\sigma>0$4

and $\sigma>0$5 at the common graph points (Khanh et al., 2022). In this sense, variational strong convexity is not ordinary strong convexity of the original objective; it is a graph-based local equivalence between the subgradient graph of the given function and that of a strongly convex function.

This leads directly to variational sufficiency and strong variational sufficiency for unconstrained minimization. For

$\sigma>0$6

the strong variational sufficient condition for local optimality holds at $\sigma>0$7 with modulus $\sigma>0$8 if $\sigma>0$9 and $0\in\partial f(\bar x)$0 is variationally strongly convex at $0\in\partial f(\bar x)$1 for $0\in\partial f(\bar x)$2 with modulus $0\in\partial f(\bar x)$3 (Khanh et al., 2022). In composite optimization, the same idea is applied to an augmented perturbation model. The 2025 composite-optimization formulation uses

$0\in\partial f(\bar x)$4

equivalently

$0\in\partial f(\bar x)$5

with $0\in\partial f(\bar x)$6, $0\in\partial f(\bar x)$7 of class $0\in\partial f(\bar x)$8, and $0\in\partial f(\bar x)$9 proper, l.s.c., convex. At a KKT pair $f$0,

$f$1

where $f$2, strong variational sufficiency holds at $f$3 relative to $f$4 if there exists $f$5 such that $f$6 is variationally strongly convex at $f$7 (Mordukhovich et al., 13 Jul 2025).

2. Complete second-order characterization in composite optimization

The broad composite model studied in "Characterizations of Strong Variational Sufficiency in General Models of Composite Optimization" (Mordukhovich et al., 13 Jul 2025) is

$f$8

with nonconvexity entering through the composition $f$9. The theory does not assume $\bar x$0 is linear nor any constraint qualification. Instead, it uses two assumptions on $\bar x$1: parabolic regularity together with parabolic epi-differentiability, and a proximal Jacobian structure formulated through a special matrix $\bar x$2. These assumptions are constructively implemented for nonpolyhedral problems that involve the nuclear norm function and the indicator function of the positive-semidefinite cone without any constraint qualifications.

The central second-order object is the second-order variational function

$\bar x$3

defined from the generalized Jacobian of the proximal mapping. For the composite problem, the generalized strong second-order sufficient condition is

$\bar x$4

Under the proximal Jacobian structure, $\bar x$5 reduces to the explicit quadratic form

$\bar x$6

and $\bar x$7 otherwise.

The same paper proves a complete equivalence theorem. At a KKT pair $\bar x$8, under the two assumptions on $\bar x$9, the following are equivalent: strong variational sufficiency at $0$0; the generalized SSOSC just displayed; and positive-definiteness of all matrices in the generalized Hessian of the augmented Lagrangian

$0$1

for sufficiently large $0$2. Here

$0$3

This yields a complete second-order characterization of strong variational sufficiency in a broad composite, nonpolyhedral setting, without any constraint qualification (Mordukhovich et al., 13 Jul 2025).

3. Moreau envelopes, second-order subdifferentials, and classical strong SOSC

A decisive structural result is that variational strong convexity is equivalent to local strong convexity of the Moreau envelope. For an l.s.c. proper $0$4 and $0$5,

$0$6

If $0$7 is l.s.c. and prox-bounded, then $0$8 is variationally strongly convex at $0$9 for $(\bar x,\bar u)$0 with modulus $(\bar x,\bar u)$1 if and only if $(\bar x,\bar u)$2 is prox-regular at $(\bar x,\bar u)$3 for $(\bar x,\bar u)$4 and, for all sufficiently small $(\bar x,\bar u)$5, $(\bar x,\bar u)$6 is locally strongly convex around $(\bar x,\bar u)$7 with modulus

$(\bar x,\bar u)$8

The same theory gives second-order criteria: variational strong convexity is equivalent to uniform positive-definiteness of the generalized Hessian in a graph neighborhood, and also to the pointwise condition

$(\bar x,\bar u)$9

under prox-regularity and subdifferential continuity (Khanh et al., 2022).

These results place strong variational sufficiency in direct contact with tilt stability. For continuously prox-regular functions, variational strong convexity at $\min_x f_0(x)+g(F(x))$0, tilt stability of the local minimizer, and local strong monotonicity of the subdifferential are equivalent. In nonlinear programming, the composite theory collapses to the classical strong second-order sufficient condition. Under LICQ and KKT conditions, strong variational sufficiency at $\min_x f_0(x)+g(F(x))$1 is equivalent to

$\min_x f_0(x)+g(F(x))$2

where $\min_x f_0(x)+g(F(x))$3 is the critical subspace (Khanh et al., 2022). The 2025 composite paper states the same reduction more broadly: in nondegenerate settings it reduces exactly to the SSOSC characterization of strong variational sufficiency obtained earlier for composite problems, and, in the NLP case, to the classical SSOSC (Mordukhovich et al., 13 Jul 2025).

4. Nonpolyhedral models, explicit formulas, and absence of constraint qualifications

A major development is the transfer of strong variational sufficiency from polyhedral and nondegenerate settings to nonpolyhedral composite models without constraint qualifications. In the 2025 composite theory, the assumptions on $\min_x f_0(x)+g(F(x))$4 are verified in the appendix for the indicator of the positive-semidefinite cone and for the nuclear norm. This produces explicit SSOSC formulas for nonlinear semidefinite programs and for composite nuclear-norm problems, including rank-regularized matrix problems, again without CQs on $\min_x f_0(x)+g(F(x))$5 (Mordukhovich et al., 13 Jul 2025).

For nonlinear semidefinite programming,

$\min_x f_0(x)+g(F(x))$6

the earlier paper "Strong Variational Sufficiency for Nonlinear Semidefinite Programming and its Implications" proves that strong variational sufficiency at a KKT pair $\min_x f_0(x)+g(F(x))$7 is equivalent to a semidefinite strong SOSC and also equivalent to positive-definiteness of all matrices in the Hessian bundle

$\min_x f_0(x)+g(F(x))$8

with $\min_x f_0(x)+g(F(x))$9. The same paper emphasizes that this equivalence does not require the uniqueness of multiplier or any other constraint qualifications (Wang et al., 2022).

A closely related strong second-order theory for decomposable nonsmooth problems studies

$\sigma>0$0

with $\sigma>0$1 and $\sigma>0$2 prox-regular and $\sigma>0$3-strictly decomposable. There, SSOSC is expressed as positive-definiteness of $\sigma>0$4 on the affine hull of the critical cone, and is shown to be equivalent to strong metric regularity of the subdifferential, the normal map, and the natural residual, as well as to CD- and BD-regularity type conditions. That paper describes this constellation as a form of strong variational sufficiency in a composite setting with decomposable structure (Ouyang et al., 2023).

5. Stability, local duality, and algorithmic consequences

Strong variational sufficiency has become a stability concept, not only a local-optimality criterion. In the parametric framework

$\sigma>0$5

with multiplier set

$\sigma>0$6

the 2026 paper "Variational Sufficiency and Solution Stability in Optimization" defines strong variational sufficiency at $\sigma>0$7 by requiring that an elicited function

$\sigma>0$8

be variationally $\sigma>0$9-convex at $\bar x$00 for every multiplier $\bar x$01. Under the basic constraint qualification on horizon subgradients, strong variational sufficiency implies full substability: the localized solution map is Lipschitz in the tilt variable $\bar x$02, continuous in the perturbation variable $\bar x$03, and the localized value function is $\bar x$04 in $\bar x$05 with

$\bar x$06

The same paper also shows that full stability is stronger: it is equivalent to strong variational sufficiency, or full substability, together with a uniform bound on the inner norms of the graphical derivatives of the partial solution maps (Benko et al., 21 Feb 2026).

On the algorithmic side, the semidefinite paper establishes local convergence of the augmented Lagrangian method under strong SOSC, because strong variational sufficiency implies the strong convexity-like local behavior needed for a local dual problem and for a dual proximal point interpretation of ALM subproblems. It also shows that under strong SOSC the generalized Hessian of the augmented Lagrangian is positive-definite, which is the condition used to apply semismooth Newton methods to the ALM subproblems (Wang et al., 2022). The later parametric-stability theory suggests the same overall pattern more abstractly: strong variational sufficiency gives a structured local model of optimal values, solutions, and multipliers, but it does not by itself collapse all stability notions into full stability (Benko et al., 21 Feb 2026).

6. Infinite-dimensional and manifold extensions, and scope of the term

The analytic core of strong variational sufficiency extends beyond finite-dimensional Euclidean problems. In Banach spaces, strong variational convexity is defined by the same local coincidence of the subgradient graph with that of a strongly convex function, now using a value-filtered neighborhood. The infinite-dimensional theory shows, under suitable geometric assumptions, that strong variational convexity is equivalent to $\bar x$07-local strong maximal monotonicity of the subdifferential, to strong convexity of local Moreau envelopes, and to tilt stability of local minimizers. This provides the infinite-dimensional analytic backbone for strong variational sufficiency, even though the latter is usually stated for optimization models rather than for standalone functions (Khanh et al., 2023).

On Riemannian manifolds, the notion is transferred to tangent spaces through a retraction $\bar x$08. For

$\bar x$09

the manifold strong variational sufficient condition at $\bar x$10 means that the localized augmented perturbed objective on $\bar x$11 is variationally strongly convex. This is equivalent to positive-definiteness of all matrices in the $\bar x$12-Hessian bundle of the localized augmented Lagrangian, and, for polyhedral, second-order cone, and semidefinite structures, to the manifold strong second-order sufficient condition. The same equivalence yields local duality, augmented tilt stability, local linear convergence of the Riemannian augmented Lagrangian method, and nonsingularity of the generalized Hessian needed for semismooth Newton methods (Zhou et al., 2023).

A recurrent misconception is that strong variational sufficiency is merely a restatement of strong convexity of the original objective. The current literature treats it differently: it is a local graph property of subdifferentials, typically for an augmented objective or perturbation model, and its strongest concrete forms are second-order characterizations at KKT pairs. Another source of ambiguity is terminological rather than mathematical. In one discussion of regret and divergence functionals, “strong variational sufficiency” is used as shorthand for a sufficiency-like invariance property of a regret functional, while the paper itself does not use that phrase; that usage is distinct from Rockafellar’s optimization notion (Harremoës, 2017).

Taken in its established optimization meaning, strong variational sufficiency now denotes a second-order, primal–dual, and stability-oriented sufficient condition for local optimality. In finite dimensions it subsumes classical strong SOSC in nonlinear programming; in composite nonpolyhedral models it admits complete characterizations through second-order variational functions and augmented-Lagrangian Hessian bundles; and in stability and algorithmic theory it underpins tilt stability, local duality, local convergence of augmented Lagrangian schemes, and Newton-type fast local rates (Mordukhovich et al., 13 Jul 2025).