Fenchel–Moreau Conjugate: Duality in Convex Analysis

• Fenchel–Moreau conjugate is a duality transformation in convex analysis that generalizes the Legendre transform to produce exact dual representations for proper, convex, and lower semicontinuous functions.
• Its extensions to Banach modules, manifolds, and convex cones broaden its applicability in optimization, variational analysis, and partial differential equations.
• The geometric interpretation using support functions of epigraphs underpins advanced subdifferential calculus and unifies duality frameworks across diverse mathematical settings.

The Fenchel–Moreau conjugate is a central construct in convex analysis, functional analysis, and optimization theory, representing a duality transformation that generalizes the classical Legendre transform. For proper, convex, and lower semicontinuous functions, the conjugate provides an exact dual representation and underpins the principle that the closed convex hull of the epigraph coincides with the function itself. Its generalization from real-valued functions on vector spaces to various non-classical settings—including Banach modules, manifolds, cones, abstract sets, and measure spaces—has led to nuanced duality and biconjugation identities crucial for modern variational analysis, optimization, and partial differential equations.

1. Classical Definition and Fundamental Theorem

Given a real vector space $X$ and an extended real-valued function $f : X \to \mathbb{R} \cup \{+\infty\}$, the Fenchel (convex) conjugate of $f$ is

$$f^* : X^* \to \mathbb{R} \cup \{+\infty\}, \quad f^*(y) = \sup_{x \in X} \{\langle y, x \rangle - f(x)\}$$

where $X^*$ is the topological dual of $X$.

The biconjugate is defined by

$$f^{**} : X \to \mathbb{R} \cup \{+\infty\}, \quad f^{**}(x) = \sup_{y \in X^*} \{\langle y, x \rangle - f^*(y)\}$$

The Fenchel–Moreau theorem states that $f^{**} = f$ if and only if $f$ is proper, convex, and lower semicontinuous. This identity establishes that every such $f$ can be represented as a supremum over its affine minorants, providing the geometric foundation for duality theory (Mordukhovich et al., 2016, Schiela et al., 2024).

2. Extension to Generalized Settings

2.1. $f : X \to \mathbb{R} \cup \{+\infty\}$0-Valued Functions (Bochner Modules)

For functions $f : X \to \mathbb{R} \cup \{+\infty\}$1, where $f : X \to \mathbb{R} \cup \{+\infty\}$2 denotes the space of extended real-valued measurable functions on a $f : X \to \mathbb{R} \cup \{+\infty\}$3-finite measure space $f : X \to \mathbb{R} \cup \{+\infty\}$4, and $f : X \to \mathbb{R} \cup \{+\infty\}$5 is a dual pair of Banach spaces, the Fenchel–Moreau conjugate is

$f : X \to \mathbb{R} \cup \{+\infty\}$6

where $f : X \to \mathbb{R} \cup \{+\infty\}$7 is the space of strongly measurable $f : X \to \mathbb{R} \cup \{+\infty\}$8-valued functions (Drapeau et al., 2017, Drapeau et al., 2016). The notion of stable (os-) lower semicontinuity is essential, requiring semicontinuity along stable nets and under almost-everywhere concatenations in the $f : X \to \mathbb{R} \cup \{+\infty\}$9-module topology.

The corresponding biconjugacy theorem: $f$0 holds if and only if $f$1 is os-lsc, convex, and proper. This result subsumes the classical theorem as a special case and is foundational for random-valued convex analysis.

2.2. Manifolds, Groups, and Nonlinear Test Functions

Generalizations to smooth manifolds and abstract sets replace the dual pairing by families of test functions (possibly nonlinear). For a set $f$2 and a family $f$3 of “test functions,” the nonlinear Fenchel conjugate is defined as

$f$4

with the biconjugate characterized by

$f$5

This framework recovers classical biconjugacy when $f$6 is the space of affine functions (Schiela et al., 2024).

On smooth manifolds, test functions can be taken as $f$7-functions, leading to Fenchel–Young principles and viscosity subdifferential characterizations. On Lie groups, using homomorphisms, the conjugate encodes infimal convolution structures natural to the group.

2.3. Convex Cones and Monotone Conjugacy

For a pointed convex cone $f$8 (e.g., $f$9, $$f^* : X^* \to \mathbb{R} \cup \{+\infty\}, \quad f^*(y) = \sup_{x \in X} \{\langle y, x \rangle - f(x)\}$$0), define the monotone Fenchel conjugate for $$f^* : X^* \to \mathbb{R} \cup \{+\infty\}, \quad f^*(y) = \sup_{x \in X} \{\langle y, x \rangle - f(x)\}$$1 as

$$f^* : X^* \to \mathbb{R} \cup \{+\infty\}, \quad f^*(y) = \sup_{x \in X} \{\langle y, x \rangle - f(x)\}$$2

where $$f^* : X^* \to \mathbb{R} \cup \{+\infty\}, \quad f^*(y) = \sup_{x \in X} \{\langle y, x \rangle - f(x)\}$$3 is the dual cone. The biconjugation $$f^* : X^* \to \mathbb{R} \cup \{+\infty\}, \quad f^*(y) = \sup_{x \in X} \{\langle y, x \rangle - f(x)\}$$4 equals $$f^* : X^* \to \mathbb{R} \cup \{+\infty\}, \quad f^*(y) = \sup_{x \in X} \{\langle y, x \rangle - f(x)\}$$5 provided $$f^* : X^* \to \mathbb{R} \cup \{+\infty\}, \quad f^*(y) = \sup_{x \in X} \{\langle y, x \rangle - f(x)\}$$6 is perfect (self-dual with regular faces) and $$f^* : X^* \to \mathbb{R} \cup \{+\infty\}, \quad f^*(y) = \sup_{x \in X} \{\langle y, x \rangle - f(x)\}$$7 is proper, convex, l.s.c., and $$f^* : X^* \to \mathbb{R} \cup \{+\infty\}, \quad f^*(y) = \sup_{x \in X} \{\langle y, x \rangle - f(x)\}$$8-monotone. This biconjugacy supports variational principles for cone-constrained optimization and Hamilton–Jacobi PDEs on cones (Chen et al., 2020).

3. Geometric and Variational Approaches

An influential interpretation due to Moreau and others expresses the conjugate in terms of support functions of epigraphs: $$f^* : X^* \to \mathbb{R} \cup \{+\infty\}, \quad f^*(y) = \sup_{x \in X} \{\langle y, x \rangle - f(x)\}$$9 where $X^*$0 is the support function (Mordukhovich et al., 2016). Many calculus rules—for sums, compositions, pointwise maxima, and infimal convolutions—are derived from the geometric intersection rules for support functions, yielding unified and general proofs of duality and subdifferential formulas.

The approach also facilitates subdifferential calculus, with $X^*$1 naturally linked to normal cones to epigraphs.

4. Abstract and Multi-Coupling Frameworks

Fenchel–Moreau conjugacy extends to settings with arbitrary coupling functions (not necessarily bilinear). Given primal and dual sets $X^*$2 and a coupling $X^*$3, the conjugate is

$X^*$4

Results include inequalities and equality conditions for three-coupling settings, underpinning duality for infimal convolutions with general kernels and providing dual forms for stochastic dynamic programming value functions, including for Bellman equations in the presence of uncertainty or stochasticity (Chancelier et al., 2018).

Notably, precise duality and biconjugation identities hinge on the geometry of the coupling, convexity in the associated structures, and appropriate semicontinuity or face-regularity assumptions.

5. Illustrative Examples and Special Cases

Setting	Primal Space	Dual Space / Test Functions
Classical vector spaces	$X^*$5	$X^*$6
$X^*$7-spaces	$X^*$8	$X^*$9
Probability measures	$X$0	$X$1
Manifolds	$X$2, $X$3	$X$4, $X$5
Cones	$X$6	$X$7
Abstract sets	$X$8	General $X$9

Illustrative examples include:

• The absolute value function on $$f^{**} : X \to \mathbb{R} \cup \{+\infty\}, \quad f^{**}(x) = \sup_{y \in X^*} \{\langle y, x \rangle - f^*(y)\}$$0, with conjugate and biconjugate structures exactly mirroring the subdifferential $$f^{**} : X \to \mathbb{R} \cup \{+\infty\}, \quad f^{**}(x) = \sup_{y \in X^*} \{\langle y, x \rangle - f^*(y)\}$$1 and inner envelopes of smooth minorants (Schiela et al., 2024).
• On cones, $$f^{**} : X \to \mathbb{R} \cup \{+\infty\}, \quad f^{**}(x) = \sup_{y \in X^*} \{\langle y, x \rangle - f^*(y)\}$$2 over $$f^{**} : X \to \mathbb{R} \cup \{+\infty\}, \quad f^{**}(x) = \sup_{y \in X^*} \{\langle y, x \rangle - f^*(y)\}$$3 yields biconjugation identity holding precisely with the cone-restricted conjugate (Chen et al., 2020).
• On manifolds, squared distance functions and affine-invariant log-det potentials on the symmetric positive definite cone admit classical conjugate calculations via the Riemannian exponential/log maps (Louzeiro et al., 2021, Bergmann et al., 2019).
• Dual representation of convex functionals on Wasserstein-1 space utilizing AE-Lip duality, key for variational analysis on spaces of probability measures (Laschos et al., 2016).

6. Applications, Impact, and Future Perspectives

The Fenchel–Moreau conjugate, with its generalizations, is foundational in convex and variational analysis, optimization, stochastic control, and PDE theory. Applications span:

• Primal-dual algorithms (with explicit geometric and Riemannian generalization),
• Variational representations of value functions for POMDPs and optimal transport,
• Duality principles in conic optimization, risk measures, and stochastic programming,
• Hamilton–Jacobi equations over cones and manifolds.

Ongoing research extends conjugacy theory to settings lacking linear or even metric structure, employing abstract test function families, and exploring further the role of duality and separation in nonclassical geometries (Schiela et al., 2024, Chancelier et al., 2018). Characterizations of biconjugacy in infinite-dimensional and partially ordered settings hinge on sophisticated face geometry and conditional set-theoretic analysis (Chen et al., 2020, Drapeau et al., 2016). These avenues shape future directions in both theory and computational methodology.