Bilateral facial reduction: qualification-free subdifferential calculus and exact duality

Qualification conditions (also termed constraint qualifications) help avoid pathological behavior at domain boundaries in convex analysis. By generalizing facial reduction from conic programming to general convex programs of the form $f(x) + g(Ax)$, we provide qualification-free generalizations of several key results: an exact Fenchel-Rockafellar dual, KKT optimality conditions, an attained infimal convolution for the conjugate of a sum, subdifferential sum and chain rules, and normal cones of intersections. All our results reduce seamlessly to their original formulations when qualification conditions hold. The core insight is that for a sum of two convex functions, there is an affine subspace$\unicode{x2014}$the joint supporting subspace$\unicode{x2014}$that contains the feasible region, and such that qualification conditions hold when restricting the effective domain of each function to it. We offer a number of characterizations for the joint supporting subspace, including one that obtains the affine subspace via iterative, bilateral reduction between the two domains. In our proofs, which are self-contained, we develop a structured induction on faces where inductive steps are associated with normal vectors nested in supporting subspaces (a generalization of supporting hyperplanes). With this tool, we characterize the facial structure of the difference of two convex sets from the facial structures of the individual convex sets.

💡 Research Summary

This paper extends the classical convex analysis toolkit—Fenchel‑Rockafellar duality, Karush‑Kuhn‑Tucker (KKT) optimality conditions, subdifferential sum and chain rules, and normal‑cone formulas—to settings where the usual qualification (or constraint) conditions fail. The authors achieve this by introducing a geometric construction called the “joint supporting subspace” (also referred to as the joint supporting affine subspace) for a pair of convex functions f and g. Given the effective domains C = dom f and D = dom g, the joint supporting subspace T(C,D) is defined as the linear span of the minimal face of the Minkowski difference C − D that contains the origin; its affine counterpart Tₐ(C,D) is T(C,D) translated by the intersection C ∩ D. This subspace has two crucial properties: (i) it always contains the feasible region C ∩ D, and (ii) within Tₐ(C,D) the two domains are never properly separable, which is equivalent to the satisfaction of the usual interior‑intersection qualification. Consequently, by restricting each function to Tₐ(C,D) (i.e., defining f′ = f + ι_{Tₐ} and g′ = g + ι_{Tₐ}), the qualification conditions become automatic, and all classical results hold verbatim for f′ and g′. Because f + g ≡ f′ + g′, the original problem inherits these exact formulas without any additional assumptions.

The paper provides several equivalent characterizations of the joint supporting subspace. One representation uses generated supporting subspaces: T(C,D) = H_C(C∩D) + H_D(C∩D), where H_C(S) = span(F_C(S) − F_C(S)) and F_C(S) denotes the smallest face of C containing S. Another, more algorithmic, description is given by a bilateral facial reduction procedure. Starting from any point x ∈ C∩D, one iteratively intersects the current affine subspace with the orthogonal complement of the sum of the normal cones N_C(x) and −N_D(x). This process reduces the ambient dimension at each step and terminates in at most n steps (the ambient space dimension), yielding exactly T(C,D). The authors prove that the resulting subspace is independent of the initial point, establishing a well‑defined reduction operator.

Armed with this geometric machinery, the authors derive the following qualification‑free results:

1. Exact Fenchel‑Rockafellar Duality – The dual problem constructed from f′ and g′ always attains its optimum and satisfies strong duality, regardless of the original domains’ relative interiors.

2. Qualification‑Free KKT Conditions – For problems of the form min { f(x) + g(Ax) }, the KKT system involving the subgradients of f′ and g′ holds at any primal‑dual optimal pair, without any Slater‑type assumptions.

3. Exact Subdifferential Sum Rule – The subdifferential of the sum satisfies ∂(f+g)(x) = ∂f′(x) + ∂g′(x) for all x in the domain, even when relint dom f ∩ relint dom g = ∅. The paper revisits a classic counterexample (f(x)=−√x, g = indicator of (−∞,0]) and shows that the joint supporting subspace reduces to {0}, restoring the correct sum rule.

4. Exact Subdifferential Chain Rule – For compositions g∘A, the chain rule holds with the reduced functions, again without any interior‑intersection requirement.

5. Normal‑Cone Intersection Formula – The normal cone to the intersection C∩D is expressed as the sum of the normal cones to C and D restricted to the joint supporting subspace, i.e., N_{C∩D}(x) = N_C(x) + N_D(x) when x ∈ C∩D.

A novel conceptual tool introduced is the notion of nested normals. A nested normal is a normal vector that belongs to a sequence of faces, each contained in the next, forming a lattice structure. By organizing faces via nested normals, the authors develop a structured induction that traverses all supporting subspaces, allowing them to prove the facial characterizations of differences of convex sets and to establish the reduction algorithm’s correctness. This approach generalizes the lexicographic face theory previously used in conic programming, offering a more flexible and geometric perspective.

The paper also discusses computational implications. In conic programming, facial reduction is known to reduce problem size and improve numerical stability; the authors argue that their bilateral facial reduction, being applicable to general convex programs, should provide similar benefits. By projecting the problem onto the joint supporting subspace, one can potentially solve a lower‑dimensional problem with better conditioning, which is especially valuable when the original domains lie on lower‑dimensional manifolds or have intricate boundary structures.

In summary, the authors present a unified, qualification‑free framework for convex subdifferential calculus and duality. The central insight—that a single affine subspace can simultaneously regularize both functions—allows them to recover all classical results without any Slater‑type assumptions. The paper supplies both analytic characterizations and an explicit reduction algorithm, supported by a robust geometric theory based on nested normals. This work opens avenues for applying exact duality and subdifferential calculus in settings previously considered pathological, and suggests practical algorithmic improvements for large‑scale convex optimization.