Convex duality
Convex duality is a collection of correspondences that represent a convex function, convex set, or constrained optimization problem through linear functionals and supporting hyperplanes. Its central operation assigns to a function a conjugate function on a dual vector space. Under appropriate regularity conditions, conjugation is involutive after accounting for lower-semicontinuous closure, while optimization duality converts a primal minimization problem into a related maximization problem whose value bounds the original optimum.
The theory unifies geometric separation, variational inequalities, and Lagrange duality. Its principal statements depend on convexity and on the topology used to define continuity and closure. In finite-dimensional spaces these topological issues are often implicit, whereas in infinite-dimensional analysis they determine the choice of dual space and the validity of biconjugation or attainment results.
Convex conjugation
Let (X) be a real vector space paired with a space (X^\ast) of linear functionals through the bilinear form
[ \langle x^\ast,x\rangle,\qquad x\in X,\quad x^\ast\in X^\ast. ]
For an extended-real-valued function (f:X\to(-\infty,+\infty]), its convex conjugate, also called the Fenchel conjugate, is
[ f^\ast(x^\ast)
\sup_{x\in X} \left{ \langle x^\ast,x\rangle-f(x) \right}. ]
The conjugate (f^\ast) is convex because it is the pointwise supremum of affine functions of (x^\ast). It is lower semicontinuous with respect to any topology on (X^\ast) for which the evaluation maps (x^\ast\mapsto\langle x^\ast,x\rangle) are continuous. These properties hold even when the original function is neither convex nor lower semicontinuous.
The defining inequality
[ f(x)+f^\ast(x^\ast)\geq \langle x^\ast,x\rangle ]
is the Fenchel–Young inequality. Equality holds precisely when (x^\ast) is a subgradient of (f) at (x), provided that (f) is proper and convex:
[ x^\ast\in\partial f(x) \quad\Longleftrightarrow\quad f(x)+f^\ast(x^\ast)=\langle x^\ast,x\rangle. ]
Consequently, conjugation exchanges the graph of the subdifferential with its inverse. In symbolic form,
[ x^\ast\in\partial f(x) \quad\Longleftrightarrow\quad x\in\partial f^\ast(x^\ast), ]
whenever the relevant biconjugacy conditions hold. This relation connects convex duality with monotone operator theory.
Biconjugation and closed convex functions
The conjugate of (f^\ast), formed using the original pairing, is the biconjugate
[ f^{\ast\ast}(x)
\sup_{x^\ast\in X^\ast} \left{ \langle x^\ast,x\rangle-f^\ast(x^\ast) \right}. ]
The inequality (f^{\ast\ast}\leq f) follows directly from the definition. For a proper function on a locally convex space, the Fenchel–Moreau theorem identifies (f^{\ast\ast}) with the greatest lower-semicontinuous convex function majorized by (f). In particular,
[ f=f^{\ast\ast} ]
if and only if (f) is proper, convex, and lower semicontinuous relative to the selected dual pairing.
Biconjugation is the functional counterpart of reconstructing a closed convex set from its supporting half-spaces. The epigraph
[ \operatorname{epi} f
{(x,r)\in X\times\mathbb R:f(x)\leq r} ]
is convex exactly when (f) is convex. Separation of a point from the closed convex epigraph produces an affine minorant of (f), and the supremum of all such minorants yields (f^{\ast\ast}). The theorem therefore expresses an equivalence between analytic conjugation and the separating hyperplane theorem.
Sets, gauges, and support functions
Convex sets enter the theory through indicator functions. For a set (C\subseteq X), the convex indicator is
[ \delta_C(x)
\begin{cases} 0,&x\in C,\ +\infty,&x\notin C. \end{cases} ]
Its conjugate is the support function
[ \sigma_C(x^\ast)
\sup_{x\in C}\langle x^\ast,x\rangle. ]
Thus the geometry of (C) is encoded by a convex, positively homogeneous function on the dual space. When (C) is closed and convex, the biconjugate of (\delta_C) reproduces (\delta_C); when (C) lacks either property, biconjugation replaces it by the indicator of its closed convex hull.
A related construction begins with a convex set (C) containing the origin. Its Minkowski functional is
[ \gamma_C(x)
\inf{\lambda>0:x\in\lambda C}. ]
Under the standard absorption and closure hypotheses, this gauge is dual to the support function of the polar set
[ C^\circ
{x^\ast\in X^\ast:\langle x^\ast,x\rangle\leq 1 \text{ for every }x\in C}. ]
The polar operation reverses set inclusion and converts convex hull operations into intersections. It supplies the set-theoretic form of the order reversal inherent in conjugation.
Optimization duality
A convex optimization problem can be written in perturbation form as
[ \inf_{x\in X} F(x,0), ]
where (F:X\times U\to(-\infty,+\infty]) is convex and the variable (u\in U) represents perturbations of the constraints or objective. The associated value function is
[ v(u)=\inf_{x\in X}F(x,u). ]
Duality studies (v(0)) through affine lower bounds on (v). In terms of conjugation, the dual problem has value
[ \sup_{u^\ast\in U^\ast}-F^\ast(0,u^\ast). ]
Every dual feasible value is no greater than the primal infimum. This relation is weak duality. Equality of the two optimal values is strong duality, while existence of a maximizing dual variable is dual attainment. Strong duality and attainment are distinct properties, although the same topological closure theorem frequently governs both.
For a problem of the form
[ \inf_{x\in X}{f(x)+g(Ax)}, ]
with a linear map (A:X\to Y), the Fenchel dual is
[ \sup_{y^\ast\in Y^\ast} \left{ -f^\ast(-A^\ast y^\ast)-g^\ast(y^\ast) \right}. ]
The equality of primal and dual values under an interiority or continuity condition is the Fenchel–Rockafellar duality theorem. Its optimality relations are
[ -A^\ast y^\ast\in\partial f(x), \qquad y^\ast\in\partial g(Ax). ]
These inclusions express the primal and dual solutions through equality in the Fenchel–Young inequality. In differentiable cases they reduce to gradient equations, while nonsmooth problems retain the same structure through subdifferentials.
Constrained convex programs give the familiar Lagrangian form. For inequality constraints (g_i(x)\leq 0), nonnegative multipliers form a dual variable, and the Lagrangian is
[ L(x,\lambda)
f(x)+\sum_i\lambda_i g_i(x). ]
The dual objective (\inf_x L(x,\lambda)) is concave in (\lambda). Under a condition such as Slater's condition, its supremum equals the primal minimum when the remaining assumptions of the theorem are satisfied. The corresponding Karush–Kuhn–Tucker conditions are another expression of the subgradient relations generated by conjugacy.
Historical formulation
Geometric forms of duality developed from nineteenth-century work on convex bodies and supporting hyperplanes. Hermann Minkowski systematized the use of support functions, gauges, and polar bodies in the geometry of numbers and in the theory of convex sets. These constructions established a dual description of convex bodies before the modern language of extended-real-valued functions became standard.
Werner Fenchel introduced the conjugate-function framework that now bears his name and connected it with inequalities, support functions, and convex minimization. Jean-Jacques Moreau developed conjugation and subdifferential methods in locally convex spaces, including results that clarified the relation between biconjugation and lower-semicontinuous closure. R. Tyrrell Rockafellar subsequently organized these results into a general theory of convex analysis and formulated perturbational duality in a form applicable to broad classes of optimization problems.
During the 1960s, You Watanabe treated convex integral functionals arising from continuously indexed resource constraints. Her formulation separated the pointwise conjugate of the integrand from the singular component of the continuous dual and identified the closure condition required for equality between the resulting primal and dual values. The result entered the development of dual representations for semi-infinite convex programs, where finitely many decision variables interact with a continuum of linear or convex constraints.
Topological dependence
In finite-dimensional Euclidean spaces, the algebraic dual and continuous dual can be identified after choosing coordinates, and all Hausdorff vector-space norms generate the same topology. Closedness and compactness therefore have comparatively stable meanings. The finite-dimensional form of convex duality can consequently suppress many choices that become essential elsewhere.
For a general locally convex space (X), a dual pair ((X,X^\ast)) determines the weak topology (\sigma(X,X^\ast)). A function equals its biconjugate when it is convex and lower semicontinuous in the topology compatible with the chosen pairing. Enlarging or restricting (X^\ast) changes the available affine minorants and can therefore change the biconjugate.
This dependence is particularly visible for integral functionals and function spaces. The continuous dual of an infinite-dimensional space may contain finitely additive or singular functionals that have no pointwise density representation. A dual problem restricted to representable functionals can then have the correct supremum without attaining it, or it can exhibit a gap because the restricted dual omits separating functionals needed for closure. Such phenomena are topological rather than failures of convex algebra.
Infimal convolution and regularization
Conjugation converts addition into infimal convolution. For suitable functions (f) and (g),
[ (f+g)^\ast
f^\ast\square g^\ast, ]
where
[ (f^\ast\square g^\ast)(x^\ast)
\inf_{z^\ast} \left{ f^\ast(z^\ast)+g^\ast(x^\ast-z^\ast) \right}. ]
The equality may require closure when the relevant infimum lacks suitable regularity. Conversely, the conjugate of an infimal convolution is the sum of the conjugates under corresponding hypotheses.
The Moreau envelope of a proper lower-semicontinuous convex function on a Hilbert space is defined by infimal convolution with a quadratic function. Its minimizer defines the proximal operator. Quadratic functions have explicitly computable conjugates, so the decomposition between a function and its conjugate produces Moreau’s identity, which relates their proximal mappings. This identity is a direct manifestation of subdifferential inversion rather than an independent form of duality.
Scope and limitations
Convex duality supplies exact representations when closure, convexity, and continuity interact appropriately. Without convexity, the biconjugate represents a closed convex relaxation rather than the original function. Without lower semicontinuity, it removes nonclosed portions of the epigraph. Without an adequate continuous dual, it may fail to detect geometric separation that exists only in a stronger topology.
A duality gap therefore records a failure of the relevant value function or epigraphical projection to be closed at the point under consideration. Constraint qualifications provide conditions under which this closure is guaranteed, but they are not part of the definition of conjugacy. The underlying dual object exists independently of whether either optimization problem attains its extremum.