1 Statement of the Fenchel–Moreau theorem

1.1 Convex conjugate and biconjugate definitions

Let \(X\) be a real topological vector space and \(X^\*\) its continuous dual. For a function \(f:X\to(-\infty,+\infty]\), the convex conjugate (also called the Fenchel conjugate) is defined by \[ f^\*(x^\*)=\sup_{x\in X}\{\langle x^\*,x\rangle-f(x)\},\qquad x^\*\in X^\*, \] where \(\langle x^\*,x\rangle\) denotes the dual pairing. The biconjugate is the conjugate of the conjugate: \[ f^{\*\*}(x)=\sup_{x^\*\in X^\*}\{\langle x^\*,x\rangle-f^\*(x^\*)\},\qquad x\in X. \] The construction produces from \(f\) a new function \(f^{\*\*}\) determined entirely by \(f^\*\), which in turn encodes the “best affine minorants” of \(f\).

1.2 Assumptions: properness, convexity, and lower semicontinuity

The theorem’s standard conclusion requires three properties of the original function \(f\):

  • Properness: \(f\) never takes the value \(-\infty\), and it is not identically \(+\infty\). Equivalently, its effective domain \(\{x\in X: f(x)<+\infty\}\) is nonempty.
  • Convexity: for all \(x,y\in X\) and \(\theta\in[0,1]\),

\[ f(\theta x+(1-\theta)y)\le \theta f(x)+(1-\theta)f(y). \]

  • Lower semicontinuity (lsc): with respect to the topology of \(X\), the epigraph of \(f\) is closed, or equivalently,

\[ f(x)\le \liminf_{x_k\to x} f(x_k) \] for every net (or sequence in metrizable spaces) \(x_k\to x\).

These conditions are the natural regularity requirements for the conjugation operation to recover the original function without “loss.”

1.3 Equality with the biconjugate and interpretation

Under the assumptions above, the Fenchel–Moreau theorem states: \[ f(x)=f^{\*\*}(x)\quad\text{for all }x\in X. \] Interpretively, the biconjugate can be viewed as a canonical convex and lower-semicontinuous “closure” of \(f\). When \(f\) already possesses those two structural features, the closure does not change it, so equality holds. When \(f\) fails lower semicontinuity or convexity, biconjugation modifies it in a predictable way, producing the greatest closed convex minorant compatible with the conjugate.

2 Variants and equivalent formulations

2.1 Closed convex envelope perspective

A common reformulation uses the closed convex envelope (often described as the “largest closed convex function below \(f\)” in appropriate contexts). For general \(f\), \(f^{\*\*}\) is always convex and lower semicontinuous, and one has the universal inequality \[ f^{\*\*}\le f. \] When \(f\) is proper, convex, and lsc, this envelope coincides with \(f\) itself, yielding the theorem as an exact recovery statement.

2.2 Order-theoretic view via greatest lower bounds

From an order-theoretic standpoint, the conjugation process selects the maximal function—under the pointwise order—that can be represented through dual affine functionals. Concretely, one can characterize \(f^{\*\*}(x)\) as a supremum of affine functions of the form \[ x\mapsto \langle x^\*,x\rangle - f^\*(x^\*), \] and these affine maps are precisely those that lie below \(f\) in the sense enforced by the conjugate construction. This viewpoint highlights that biconjugation is a “greatest lower bound” operation within the class of closed convex functions dominated by \(f\).

2.3 Fenchel inequality and its role in the theorem

A central ingredient in all formulations is the Fenchel inequality: \[ f(x)+f^\*(x^\*)\ge \langle x^\*,x\rangle\quad\text{for all }x\in X,\ x^\*\in X^\*. \] Rearranging gives \[ f(x)\ge \langle x^\*,x\rangle - f^\*(x^\*). \] Taking the supremum over \(x^\*\) yields \(f(x)\ge f^{\*\*}(x)\), establishing the general inequality \(f^{\*\*}\le f\). The theorem’s content is that under properness, convexity, and lsc, no strict gap remains.

3 Geometric and functional-analytic intuition

3.1 Supporting hyperplanes and subgradients

In convex geometry, conjugates are intimately connected to supporting affine functions. If \(x^\*\) is a subgradient of \(f\) at \(x\) (notation \(x^\*\in\partial f(x)\)), then \[ f(y)\ge f(x)+\langle x^\*,y-x\rangle\quad\text{for all }y. \] Such a supporting hyperplane corresponds to equality cases in Fenchel-type inequalities. The biconjugate’s representation as a supremum of affine minorants can thus be understood as piecing together all valid supporting hyperplanes, and the lsc/convex assumptions ensure these supports are sufficiently rich to reconstruct \(f\).

3.2 Relationship to epigraphs and convex hulls

The epigraph of \(f\), \[ \operatorname{epi} f=\{(x,t)\in X\times\mathbb{R}: t\ge f(x)\}, \] provides a geometric lens. Conjugation corresponds to moving between \(f\) and dual descriptions that, in effect, replace the epigraph by a structure that is both convex and closed. In this setting, biconjugation can be interpreted as transforming \(\operatorname{epi} f\) into the smallest closed convex set compatible with the original data, and then reading off the resulting boundary function.

3.3 Duality between primal epigraphs and conjugate structure

Conjugate functions encode “dual certificates” of lower bounds on \(f\). The theorem asserts that, when \(f\) is closed and convex, those certificates are complete: the primal function’s graph is exactly recovered from its dual-implied lower supporting structure. This duality is a recurring theme in convex optimization: primal feasibility and objective properties are mirrored by conjugate representations on the dual side.

4 Conjugate calculus used with the theorem

4.1 Conjugates of standard functions (norms, indicators, linear forms)

Conjugation becomes most useful when paired with explicit formulas. Typical examples include:

  • Indicator functions: for a convex set \(C\subset X\),

\[ \delta_C(x)=\begin{cases} 0,&x\in C,\\ +\infty,&x\notin C, \end{cases} \] has conjugate \[ \delta_C^\*(x^\*)=\sup_{x\in C}\langle x^\*,x\rangle, \] which is the support function of \(C\).

- Norms: if \(\|\cdot\|\) is a norm, then the conjugate of \(\|\cdot\|\) is closely tied to the dual norm \(\|\cdot\|_\*\). Precise identities depend on whether the norm is used directly or squared, but the general mechanism is that conjugation converts growth conditions into constraints in the dual space.
  • Linear (affine) functions: for \(x\mapsto \langle a,x\rangle+b\), the conjugate is typically a shifted indicator of a singleton in the dual variables, reflecting that affine functions have conjugates concentrated on the corresponding dual point.

These elementary conjugates often appear as building blocks in optimization models.

4.2 Sum and infimal convolution rules (basic forms)

Conjugate calculus also includes rules connecting sums in the primal with operations in the dual:

  • The conjugate of a sum frequently becomes an operation involving conjugates, most commonly the infimal convolution:

\[ (f+g)^\* = f^\* \,\square\, g^\*,\quad f^\*\square g^\*(x^\*)=\inf_{u^\*+v^\*=x^\*}\{f^\*(u^\*)+g^\*(v^\*)\}. \]

  • Conversely, products and other nonlinear expressions generally do not enjoy comparably simple rules, which is why convex analysis often reformulates models in terms of sums of simpler convex components.

These rules are essential for applying Fenchel–Moreau systematically: one computes conjugates and then uses biconjugation to return to a primal description.

4.3 Scaling and translation properties

Conjugates respond predictably to affine transformations:

  • Scaling: if \(f_\alpha(x)=\alpha f(x)\) with \(\alpha>0\), then

\[ (f_\alpha)^\*(x^\*)=\alpha f^\*\!\left(\frac{x^\*}{\alpha}\right) \] under standard conventions.

  • Translation in the primal: replacing \(f(x)\) by \(f(x-x_0)\) affects the conjugate by a phase factor in the dual pairing, typically yielding

\[ (f(\cdot-x_0))^\*(x^\*)=f^\*(x^\*)+\langle x^\*,x_0\rangle \] (again, subject to sign conventions and whether one writes the translation in terms of \(x-x_0\) or \(x_0-x\)).

Such identities enable quick derivations of biconjugates for shifted or rescaled models.

5 Applications to convex optimization and duality

5.1 Deriving dual problems from biconjugation

Many duality derivations in convex optimization hinge on the ability to replace a primal term \(f(x)\) by its biconjugate: \[ f(x)=f^{\*\*}(x)=\sup_{x^\*}\{\langle x^\*,x\rangle-f^\*(x^\*)\}. \] When an optimization problem involves a convex objective plus linear constraints, this substitution often permits moving the supremum outside or exchanging optimization order (subject to regularity conditions). The result is a dual problem whose objective involves \(f^\*\) and the conjugates of other terms.

5.2 Optimality conditions via subgradients

The Fenchel–Moreau framework also provides an efficient way to express first-order optimality. In broad terms, optimality can be characterized by the existence of dual variables \(x^\*\) such that equality holds in the Fenchel inequality: \[ f(x)+f^\*(x^\*)=\langle x^\*,x\rangle. \] This equality corresponds to the subgradient relation \(x^\*\in\partial f(x)\) (under appropriate assumptions). Hence, the theorem links primal-dual correspondence to geometric tangency: the dual variable identifies a supporting hyperplane at the optimal primal point.

5.3 Strong duality heuristics in convex settings

While the mere equality \(f=f^{\*\*}\) does not automatically guarantee “strong” duality (e.g., equality of optimal values in a specific primal/dual pair), it supplies the structural reason such results can hold. When constraints and functions satisfy convexity and closedness requirements, biconjugation tends to ensure that dual representations are not missing any information. Additional assumptions—often expressed via constraint qualifications—then determine whether the supremum or infimum is attained and whether optimal values coincide.

6 Examples and worked cases

6.1 Recovery of a lower semicontinuous convex function from conjugates

Consider a proper, convex, lower semicontinuous function \(f\). Compute \(f^\*\) via \[ f^\*(x^\*)=\sup_x \{\langle x^\*,x\rangle-f(x)\}. \] Then form \(f^{\*\*}\): \[ f^{\*\*}(x)=\sup_{x^\*}\{\langle x^\*,x\rangle-f^\*(x^\*)\}. \] The Fenchel–Moreau theorem asserts that this reconstructed expression equals \(f(x)\) at every \(x\). In practice, the equality is often verified by showing:

1 Statement of the Fenchel–Moreau theorem

2 Variants and equivalent formulations

6.2 Example where biconjugation “closes” a non-lower-semicontinuous function

Let \(f\) be convex but not lower semicontinuous. A prototypical phenomenon is that \(f\) may lie strictly above its closed convex envelope at some points. For such an \(f\), one still always has \(f^{\*\*}\le f\). The biconjugate corrects the “defect” by lowering values at points where lower semicontinuity fails, effectively producing the largest lower semicontinuous convex minorant consistent with the same conjugate transform. Thus, biconjugation functions as a closure operator in the convex-analysis setting.

6.3 Example using indicator functions and convex sets

Take \(f=\delta_C\) for a nonempty convex set \(C\). Then \(f^\*\) is the support function: \[ \delta_C^\*(x^\*)=\sup_{x\in C}\langle x^\*,x\rangle. \] Now compute the biconjugate: \[ \delta_C^{\*\*}(x)=\sup_{x^\*}\{\langle x^\*,x\rangle-\sup_{y\in C}\langle x^\*,y\rangle\}. \] Geometrically, this expression evaluates to \(0\) on the closed convex hull of \(C\) and to \(+\infty\) outside that closed convex hull. Therefore, \(\delta_C^{\*\*}\) recovers the indicator of the closed convex hull of \(C\). This illustrates how biconjugation converts a set indicator into the closed convex set dictated by dual support behavior.

7.1 Connections to separation theorems

Separation theorems in convex analysis—statements that disjoint convex sets can be separated by a continuous affine functional—underpin the Fenchel–Moreau result. Roughly, closed convexity ensures the existence of supporting hyperplanes to the epigraph at every boundary point. Those supports translate into dual variables achieving the supremum in the biconjugate formula, which yields the equality \(f=f^{\*\*}\).

A common proof strategy proceeds through a functional-analytic route: Hahn–Banach-type arguments are used to construct continuous linear functionals that separate a point from a closed convex set. Since epigraphs of convex lsc functions are closed convex sets, applying separation yields the supporting affine minorants needed to show \(f\le f^{\*\*}\). In this way, Hahn–Banach serves as a mechanism for deriving the dual representations required by Fenchel–Moreau.

7.3 Relation to Legendre–Fenchel transforms

Fenchel conjugation is often discussed alongside the Legendre transform, especially in settings where the function is differentiable or defined on Euclidean spaces with additional regularity. The Legendre–Fenchel transform broadens the classical Legendre transform by allowing nondifferentiability and extended-real-valued functions. Fenchel–Moreau can then be interpreted as a bidirectional recovery principle for the Legendre–Fenchel transform: taking the transform twice returns the closed convex function.

8 References and further reading

8.1 Classic sources in convex analysis

Classic references include foundational books and lecture notes that develop conjugacy, duality, and separation in a unified framework. These texts typically present Fenchel–Moreau early as a cornerstone theorem, then reuse it repeatedly in duality proofs and in the study of convex sets.

8.2 Modern expositions and textbook treatments

Modern treatments often emphasize:

  • geometric interpretations via epigraphs and supporting hyperplanes,
  • computational examples using conjugate calculus rules, and
  • applications to optimization problems, including primal-dual algorithms and variational models.

Look for chapters on conjugate functions, duality, and functional transforms, where Fenchel–Moreau is frequently paired with infimal convolution and subdifferential calculus.

8.3 Typical prerequisites and learning path

A typical learning sequence begins with:

  • convex sets and convex functions,
  • extended-real-valued functions and epigraph methods,
  • basic duality and separation principles,
  • then conjugates and subgradients.

Prerequisites often include linear functional analysis at the level of topological vector spaces (or Banach spaces in common courses), along with familiarity with how lower semicontinuity is expressed through closed epigraphs.