A version 1 arXiv preprint develops a mathematical framework for examining how an optimal value changes when a problem’s parameter moves in a chosen direction. Its central result gives upper estimates for both the directional limiting and directional singular subdifferentials of the value function. Those estimates are conditional: they rely on compactness, flatness, differentiability and regularity assumptions stated in the theorems.
The document is dated 20 Aug 2026 and is presented as an arXiv preprint. It is a theoretical methods paper, not an empirical study: the analyzed unit is a mathematical value function and its directional subdifferentials, rather than a participant sample or dataset. The design therefore does not produce a population estimate or a causal finding.
An optimization problem without a fixed dimension
At the center is a value function: the mathematical record of the best objective value associated with each parameter in a parametric optimization problem. A directional subdifferential adds information about possible changes along a specified direction, including cases in which ordinary smooth calculus is not enough. The paper asks how that directional information can be bounded when the spaces involved may be infinite-dimensional.
The formal model uses a parameter space X, a decision space Y and a constraint space Z. It assumes locally Lipschitz objective and constraint mappings and a closed constraint set, with the spaces treated as Asplund spaces unless otherwise specified. In plain terms, the work is concerned with an optimization problem whose underlying geometry may not fit the familiar finite-dimensional setting.
That makes the study different from a statistical sensitivity analysis. There are no groups to compare and no observed outcomes to aggregate. The evidence consists of formal definitions, propositions, theorems and proofs for the general model, together with an illustrative quadratic-program calculation.
Rules for tracing directional change
Before turning to the value function, the paper extends directional normal-cone calculus—the machinery used to describe the geometry of constraint sets—for preimage sets in infinite-dimensional spaces. The extension is stated under metric subregularity and Hadamard differentiability, conditions that provide the regularity and directional differentiability needed by the argument. This gives the main framework a way to carry directional information through constraints.
Two other pieces handle familiar mathematical constructions. A directional chain rule is established for subdifferentials in a composition setting. A directional sum rule covers a locally Lipschitz function added to a lower semicontinuous function, meaning a function whose value at a point does not sit above the limiting values approached nearby. These rules let the analysis follow directional behavior through composed and combined expressions.
The paper then establishes lower semicontinuity of the value function under a set of compactness and closure conditions. With restricted inf-compactness, a norm–τ-closed feasible graph and a norm–τ lower semicontinuous objective, it obtains lower semicontinuity of the value function; directional restricted inf-compactness gives the corresponding conclusion along a direction.
These preliminary results are not separate empirical findings. They are the structural conditions that support the later inclusion for the value function. The paper proves each implication within its stated function classes and regularity assumptions; it does not report how frequently those assumptions hold in applied problems.
Two routes to the main estimate
One route to the main estimate assumes strong directional inner-semicompactness. Under that condition, the theorem places the selected solution direction in the critical cone and gives an upper estimate for the directional subdifferential of the value function. The result connects the behavior of the solution mapping with the directional information sought about the optimum.
To reach a broader class of problems, the paper introduces directional V-flatness as a weaker assumption. A separate proposition says that, under its stated conditions, directional V-flatness holds for the pair (f, Γ). The conclusion is conditional: satisfying those requirements is enough for the property in that setting.
The main theorem still requires a specific package of assumptions: directional V-flatness and directional restricted inf-compactness, Hadamard differentiability and weak* strict Lipschitzianity of P—the latter a Lipschitz-type regularity condition—and directional metric subregularity of the constraint mapping Ψ. These conditions set the boundaries of the theorem rather than decorating it.
Within that package, the framework derives upper estimates for both the directional limiting and directional singular subdifferentials. Here, “upper estimate” means an inclusion: the theorem places the object of interest inside a mathematically described set instead of proving that the two sets are equal. That is why the result is best read as a way to narrow the possibilities, not as a universal exact formula.
Where the guarantee ends
The clearest boundary appears in the general Asplund-space result. Under the main V-flatness theorem, the selected direction is guaranteed to belong to the linearization cone, one of the geometric sets used to describe allowable first-order movement, but not to the critical cone. The finite-dimensional counterpart gives the stronger critical-cone membership.
That distinction is important because the paper is designed to extend directional analysis beyond finite-dimensional settings. The general result retains its upper estimates for the value-function subdifferentials, but it does not carry over the stronger cone conclusion under the stated level of compactness. In an applied model, identifying which cone conclusion is justified would depend on the assumptions that can actually be verified.
The same caution applies to the sensitivity estimates themselves. The principal conclusions are upper estimates or inclusions, not general equality characterizations. The evidence scope is formal variational analysis, not a numerical benchmark, computational algorithm or empirical validation, so the preprint does not quantify how closely the bounds match a particular applied problem.
Applicability depends on problem-specific conditions that include restricted inf-compactness or directional V-flatness, differentiability, weak* strict Lipschitzianity and metric subregularity. The paper supplies the theoretical conditions in its framework, but checking them in a hierarchical optimization model remains a separate question.
A worked example, not a benchmark
The supplied analysis also includes an illustrative quadratic-program example. In it, the reported directional derivative is 1, and V′(0;1) equals the objective’s joint directional derivative. The calculation shows how the abstract framework can produce a concrete directional-derivative statement for a specified problem.
But the example is not validation across a sample of optimization problems. There is no empirical sample in the study, and the calculation cannot establish that the upper estimates are equalities in general or that they automatically become a numerical sensitivity-analysis procedure.
The preprint’s contribution is therefore a conditional analytical framework: it extends several directional-calculus tools, derives upper estimates for value-function subdifferentials and makes explicit one place where the general infinite-dimensional theory is weaker than its finite-dimensional counterpart. For readers working on hierarchical or parametric optimization, the open issue is practical verification—whether the compactness and regularity assumptions can be established in the particular model under study.
Paper data and sources
Original title: Directional Subdifferentials of the Value Function in Asplund Spaces
Authors: Weihao Mao, Jane J. Ye
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text