Fenchel–Moreau Theorem Overview

• The Fenchel–Moreau theorem is a fundamental result in convex analysis that shows every proper, convex, and lsc function equals its biconjugate via the Legendre–Fenchel transform.
• It underpins duality in optimization problems and provides the basis for primal–dual methodologies in settings ranging from Euclidean spaces to manifolds and stochastic modules.
• Generalizations extend the theorem to nonlinear domains such as Wasserstein spaces, pointed cones, and random modules, highlighting its broad applicability.

The Fenchel–Moreau theorem characterizes the biconjugate of an extended-real-valued function via duality, establishing that for proper, convex, lower semicontinuous (lsc) functions on a topological vector space, the function coincides with its biconjugate defined through the Legendre–Fenchel transform. This result holds under classical convexity hypotheses but admits a range of generalizations to nonlinear settings, metric spaces, cones, manifolds, and stochastic modules, each requiring an adapted notion of convex conjugacy and biconjugacy. The theorem underlies primal–dual frameworks and is central in convex optimization, duality theory, and variational analysis.

1. Classical Fenchel–Moreau Theorem

Let $X$ be a locally convex real topological vector space with dual $X^*$. For a function $f\colon X \to (-\infty, +\infty]$ that is proper (not identically $+\infty$), convex, and lsc, the Fenchel conjugate is

$$f^*(x^*) = \sup_{x \in X} \big\{ \langle x^*, x \rangle - f(x) \big\},$$

and the biconjugate is

$$f^{**}(x) = \sup_{x^* \in X^*} \big\{ \langle x^*, x \rangle - f^*(x^*) \big\}.$$

The Fenchel–Moreau theorem asserts the equality

$f^{**}(x) = f(x), \qquad \forall x \in X.$

This equivalence expresses that a proper, convex, lsc function can be reconstructed exactly as the supremal envelope of all affine minorants below it (Chen et al., 2020, Laschos et al., 2016, Chancelier et al., 2018).

2. Key Proof Techniques and Separating Hyperplane Arguments

The proof of the classical theorem has two principal steps:

• Biconjugate Minorization: $f^{**}(x) \leq f(x)$ is immediate by definition and Fenchel’s inequality.
• Supremality via Separation: For any $x$ and $r < f(x)$, the separation theorem provides an affine minorant passing below $(x, r)$ and the closed convex epigraph of $f$. Taking suprema over all such minorants yields $f(x) \leq f^{**}(x)$ (Chen et al., 2020, Laschos et al., 2016, Chancelier et al., 2018).

Extension to non-vector settings requires adapted separation results: for example, in Hadamard manifolds, separation through metric projection onto the epigraph and existence of supporting cotangent vectors replaces Hahn–Banach arguments (Louzeiro et al., 2021).

3. Generalizations Beyond Linear Spaces

3.1. Pointed Convex Cones

Let $C \subset H$ be a closed, convex, pointed cone in a Hilbert space. The partial order $x \leq_C y$ is induced, and one defines the monotone conjugate and biconjugate restricted to $C$ and its dual cone. Under self-duality and interiority of every face (i.e., for perfect cones), the biconjugation identity holds for all proper, convex, lsc, and cone-nondecreasing $f$ on $C$ (Chen et al., 2020).

3.2. Wasserstein and Metric Spaces

For the Kantorovich–Wasserstein space $(\mathcal{P}_1(X), W_1)$, the conjugate and biconjugate employ duality with Lipschitz functions. The Fenchel–Moreau–Rockafellar theorem holds for proper, convex, $W_1$-lsc functionals, enabling duality-based variational representations, e.g., for the Donsker–Varadhan formula in stochastic control (Laschos et al., 2016).

3.3. Nonlinear and Nonstructured Domains

In the nonlinear Fenchel conjugate framework, given any set $X$ and a linear space of "test functions" $\mathcal{F}$, the conjugate is defined by

$$f^*(\varphi) = \sup_{x \in X} \{ \varphi(x) - f(x) \}, \quad \varphi \in \mathcal{F}.$$

The biconjugate is

$$f^{**}(x) = \sup_{\varphi \in \mathcal{F}} \{ \varphi(x) - f^*(\varphi) \}.$$

Equality with the $\mathcal{F}$-regularization (the supremal envelope of affine minorants from $\mathcal{F}$) generalizes Fenchel–Moreau without assumption on the domain structure, subsuming convex, smooth manifold, and group-convolution cases (Schiela et al., 2024).

4. Fenchel–Moreau Duality on Manifolds and Random Modules

4.1. Hadamard and Riemannian Manifolds

For a Hadamard manifold $M$ (complete, simply connected, nonpositive curvature), given $F: M \rightarrow \overline{\mathbb{R}}$, the Fenchel conjugate is

$$F^*(p, \xi) = \sup_{q \in M} \{ \langle \xi, \log_p q \rangle - F(q) \}.$$

The biconjugate is defined analogously. For proper, lsc, geodesically convex $F$, the manifold Fenchel–Moreau theorem yields $F^{**}(p) = F(p)$ for all $p \in M$. The proof leverages geodesic convexity and metric projections (Louzeiro et al., 2021, Bergmann et al., 2019).

4.2. Random Locally Convex Modules and $L^0$-Valued Functions

In the probabilistic module setting, functions map into $\bar{L}^0(\mathcal{F})$, the extended real-valued measurable functions. $L^0$-convexity and lower semicontinuity are localized: sublevel and epigraph analysis exploit measure-theoretic partitions, with random duals being $L^0(\mathcal{F})$-linear and continuous functionals. The Fenchel–Moreau theorem for both $(\epsilon,\lambda)$- and locally $L^0$-convex topologies characterizes the biconjugate as the supremum of affine minorants, patching together local information via countable concatenation (Guo et al., 2015, Drapeau et al., 2017).

For functions $f: X \to \bar{L}^0$ (where $X$ is a Banach space), the dual is the space of strongly measurable dual random variables ($L^0(Y)$). Proper, convex, stable lsc $f$ satisfy

$f(x) = \esssup_{y \in L^0(Y)} \{ \langle x, y \rangle - f^*(y) \}.$

Here, the "stability" and "conditional" lower semicontinuity are essential to maintain the identity almost everywhere (Drapeau et al., 2017).

5. Extension: General Couplings and Primal–Dual Algorithms

In a more abstract setting, Fenchel–Moreau conjugation can be built via arbitrary couplings (not just bilinear pairings). Chancelier & De Lara introduce conjugation relative to couplings $$c: X \times Y \to \overline{\mathbb{R}}, d: U \times V \to \overline{\mathbb{R}}$$ to derive three-coupling inequalities and extend the biconjugation theory to generalized infimal convolutions, partial conjugates, and Bellman equations in stochastic dynamic programming (Chancelier et al., 2018). The presence of regularity or saddle-point structures determines the existence of equality in such multi-coupling dualities.

In computational contexts, the Fenchel–Moreau theorem underlies primal–dual and proximal algorithms. For example, on Hadamard manifolds, the equality $F = F^{**}$ ensures no duality gap, and the subdifferential calculus extends to Riemannian settings via exponential and logarithm maps, providing necessary tools for nonsmooth manifold optimization (Bergmann et al., 2019, Louzeiro et al., 2021).

6. Illustrative Examples and Applications

Setting	Function/Class	Biconjugate Description
Euclidean space	Proper, convex, lsc	Lower-semicontinuous convex hull
Perfect cones	Cone-monotone convex	Supremum of affine cone-minorants
Hadamard manifold	Geodesic convex, lsc	Manifold envelope via cotangent vectors
Wasserstein-1 space	Convex, $W_1$-lsc	Supremum over linear functionals/Lipschitz
$L^0$-modules	$L^0$-convex, lsc	Essential supremum of random minorants

Standard examples include:

• The distance and squared distance functions on Euclidean and manifold domains, where the conjugate recovers norm-indicator functions and quadratic forms respectively (Louzeiro et al., 2021).
• The log-determinant over symmetric positive-definite matrices with the manifold structure, essential in statistical design (Louzeiro et al., 2021, Bergmann et al., 2019).
• Dual representations of conditional risk measures and value functions in stochastic optimization (Drapeau et al., 2017, Laschos et al., 2016).
• Stochastic Bellman equations with multi-coupling duality (Chancelier et al., 2018).

7. Significance and Unifying Principle

The Fenchel–Moreau theorem, across its variants, operationalizes the principle that proper, lower semicontinuity and convexity/lsc minorant structure suffice to reconstruct the original function from duality data, without further restriction to linear domains. The theorem’s unifying property underpins the viability of duality and optimization theory in nonlinear, stochastic, and geometric contexts, enabling extension of primal–dual methodologies and convex analysis far beyond the classical linear regime (Schiela et al., 2024, Drapeau et al., 2017, Louzeiro et al., 2021).