1 Definition and basic properties

1.1 Convex conjugate via supremum (Fenchel transform)

Let \(X\) be a real topological vector space and \(Y\) its (continuous) dual, paired by \(\langle y,x\rangle\). For an extended-real-valued function \(f:X\to \overline{\mathbb{R}}:=\mathbb{R}\cup\{+\infty,-\infty\}\), the (convex) Fenchel–Legendre transform (often called the convex conjugate) is the function \[ f^*(y)=\sup_{x\in X}\bigl(\langle y,x\rangle - f(x)\bigr). \] The transform converts a function into another defined by a pointwise supremum of affine forms in \(x\) (parameterized by \(y\)). When \(f\) is convex, this supremum construction encodes the tightest affine minorants that “touch” the epigraph of \(f\).

1.2 Relationship to the classical Legendre transform

In Euclidean settings, the classical Legendre transform is typically written for sufficiently smooth, strictly convex functions and can be expressed using the gradient map. The Fenchel–Legendre transform extends this idea to arbitrary convex (possibly nonsmooth or nondifferentiable) functions and uses an intrinsic supremum definition rather than requiring differentiability. When \(f\) is closed, proper, and convex and satisfies suitable regularity, the two notions coincide after identifying the dual variable with the classical slope parameter.

1.3 Domain, effective domain, and extended-real-valued functions

Because \(f\) may take the values \(+\infty\) or \(-\infty\), it is useful to distinguish:

  • The domain of \(f\): the set where \(f(x)\) is finite (in many texts restricted to \(+\infty\) allowed but \(-\infty\) excluded for properness).
  • The effective domain: \(\{x\in X: f(x)<+\infty\}\).

The conjugate \(f^*\) is also extended-real-valued. For example, if \(f\) grows slower than linearly in some direction, then \(\langle y,x\rangle-f(x)\) can be unbounded above, yielding \(f^*(y)=+\infty\). Conversely, if \(f\) imposes strong penalties via \(+\infty\) values, then some dual points \(y\) may give finite results while others do not.

1.4 Monotonicity and order-reversing behavior

The transform reverses inequalities in a controlled manner. If \(f\le g\) pointwise, then for each \(y\), \[ f^*(y)=\sup_x(\langle y,x\rangle-f(x))\ \ge\ \sup_x(\langle y,x\rangle-g(x))=g^*(y). \] Thus, the conjugation operator is order-reversing: larger primal functions correspond to smaller conjugates (and vice versa), reflecting how taking a supremum of “affine minus function” reacts to increased penalties in the original objective.

2 Examples and computations

2.1 Transform of affine and constant functions

If \(f(x)=\langle a,x\rangle+b\) is affine, then \[ f^*(y)=\sup_x\bigl(\langle y-a,x\rangle-b\bigr). \] If \(y\neq a\), the supremum is \(+\infty\) because \(\langle y-a,x\rangle\) can be made arbitrarily large. If \(y=a\), the expression equals \(-b\) for all \(x\), so \(f^*(a)=-b\). Hence \(f^*\) is the indicator-like object concentrated at the matching slope: \[ f^*(y)= \begin{cases} -b, & y=a,\\ +\infty, & y\ne a. \end{cases} \] For a constant \(f(x)=c\), this is the special case \(a=0\), producing \(f^*(0)=-c\) and \(f^*(y)=+\infty\) for \(y\ne 0\).

2.2 Transform of norms and gauge functions

Norms and gauge (Minkowski) functionals are central because they yield conjugates expressed via dual norms and indicator/support functions.

2.2.1 Euclidean norm example

Consider \(f(x)=\|x\|_2\) on \(\mathbb{R}^n\). Compute

\[

f^*(y)=\sup_x\bigl(\langle y,x\rangle-\|x\|_2\bigr).

\]

Using Cauchy–Schwarz, \(\langle y,x\rangle\le \|y\|_2\|x\|_2\), so the expression is at most \((\|y\|_2-1)\|x\|_2\). If \(\|y\|_2>1\), taking \(\|x\|_2\to\infty\) makes the supremum \(+\infty\). If \(\|y\|_2\le 1\), the best choice is \(x=0\), giving value \(0\). Thus

\[

(\|\cdot\|_2)^*(y)=

\begin{cases}

0, & \|y\|_2\le 1,\\
+\infty, & \|y\|_2>1.

\end{cases} \] So the conjugate of the norm is an indicator of the unit ball in the dual space.

2.2.2 \(p\)-norm / \(\ell^p\) conjugacy example

For \(1&lt;p&lt;\infty\), let \(q\) be the Hölder conjugate, \(1/p+1/q=1\). With \(f(x)=\frac{1}{p}\|x\|_p^p\), one has the classical identity

\[

f^*(y)=\frac{1}{q}\|y\|_q^q.

\] This equality is a standard manifestation of Hölder’s inequality and the tightness conditions for equality in Hölder. It illustrates a general theme: powers of norms conjugate to corresponding dual powers.

2.3 Transform of indicator and support functions

For a set \(C\subseteq X\), the indicator function is \[ \delta_C(x)= \begin{cases} 0, & x\in C,\\ +\infty, & x\notin C. \end{cases} \] Its conjugate is the support function: \[ \delta_C^*(y)=\sup_{x\in C}\langle y,x\rangle=: \sigma_C(y). \] Conversely, support functions conjugate back to indicators of closed convex hulls (under standard closure conditions). This duality turns geometric constraints on \(x\) into linear-growth expressions on \(y\).

2.4 Piecewise-linear convex functions

Piecewise-linear convex functions are convenient because the supremum in the definition reduces to checking finitely many candidate slopes or faces. For example, if \(f\) is the maximum of finitely many affine functions, \[ f(x)=\max_{i=1,\dots,m}\bigl(\langle a_i,x\rangle+b_i\bigr), \] then \(f^*\) can often be expressed using convex hulls of \(\{a_i\}\) and the constants \(b_i\), reflecting how taking a supremum over \(x\) becomes a supremum over active affine pieces. The result is typically a function that is \(+\infty\) outside a polyhedral region and affine (or piecewise affine) inside.

3 Convex-analytic interpretation

3.1 Supporting hyperplanes and epigraph geometry

The epigraph of \(f\), \[ \operatorname{epi} f=\{(x,t)\in X\times\mathbb{R}: t\ge f(x)\}, \] is a geometric object. For convex \(f\), the conjugate \(f^*\) arises from hyperplanes that support the epigraph. The inequality \[ f(x)+f^*(y)\ge \langle y,x\rangle \] states that every pair \((x,y)\) yields a separating affine underestimator of \(f\) relative to its conjugate. When equality holds, the corresponding hyperplane “touches” the epigraph at \(x\) with slope \(y\).

3.2 Subgradients and optimality points

A vector \(y\) is a subgradient of \(f\) at \(x\) if \[ f(z)\ge f(x)+\langle y,z-x\rangle \quad \text{for all }z. \] Equivalently, \(y\in \partial f(x)\) precisely when \[ f^*(y)=\langle y,x\rangle - f(x). \] Thus, equality in the Fenchel inequality corresponds to subgradient relations. The conjugate provides a way to locate supporting slopes without requiring differentiability.

3.3 Conjugacy in terms of separation theorems

Separation theorems in convex analysis assert that disjoint convex sets can be separated by hyperplanes. Conjugation can be derived by applying separation to sets formed from \(\operatorname{epi} f\) and affine regions. Informally, \(f^*(y)\) measures how far an affine functional can be shifted upward while still remaining below \(f\) in a global sense. This perspective links conjugacy to geometric separation and dual certificates.

3.4 Young–Fenchel inequality

For all \(x\in X\) and \(y\in Y\), the Young–Fenchel inequality states \[ f(x)+f^*(y)\ge \langle y,x\rangle. \] It follows directly from the definition of \(f^*(y)\) as a supremum: for any \(x\), \[ f^*(y)\ge \langle y,x\rangle - f(x). \] Equality holds exactly under the subgradient condition described above. The inequality is the fundamental estimate underpinning many duality results.

4 Duality results

4.1 Involution property for closed convex functions

Taking the conjugate twice does not always return the original function, but for well-behaved functions there is a reversal of the failure: \[ (f^*)^*=\operatorname{cl}\, f \] for functions that are closed (lower semicontinuous and convex) in the standard sense. This is an “involution” statement: conjugation becomes an involution on the class of proper, convex, lower semicontinuous functions.

4.2 Biconjugate theorem

The biconjugate \(f^{}\) is always a convex, lower semicontinuous function satisfying \(f^{}\le f\). The biconjugate theorem states that \(f^{**}\) equals the greatest lower semicontinuous convex minorant of \(f\). Therefore, conjugation regularizes \(f\): if \(f\) is already closed and convex, no change occurs; otherwise, the theorem describes precisely what information is lost and replaced by its closed convex envelope.

4.3 Fenchel–Moreau theorem (closedness and properness conditions)

The Fenchel–Moreau theorem formalizes when the conjugation recovers \(f\):

  • If \(f\) is proper, convex, and lower semicontinuous (closed), then \(f^{**}=f\).
  • If \(f\) is not lower semicontinuous, then \(f^{**}\) yields its lower semicontinuous hull; if it is not convex, the convex envelope appears.

Properness rules out degenerate cases where the conjugate becomes trivial because \(f\) takes \(-\infty\) values or is identically \(+\infty\).

4.4 Primal–dual formulations in convex optimization

In convex optimization, a primal problem is often written as minimizing \(f(x)\) (possibly plus linear terms). The conjugate then produces a dual formulation that typically has the form \[ \sup_{y} \bigl(-f^*(A^*y) - g^*(y)\bigr) \] for suitable compositions with linear maps. Strong duality may follow under constraint qualifications (e.g., existence of a point in the relative interior of certain domains). Even when strong duality fails, the conjugate provides computable bounds and dual certificates for optimality.

5 Subdifferential calculus and conjugation

5.1 Subgradient mapping between \(f\) and \(f^\*\)

The conjugate relationship is tightly linked to subdifferentials. Under standard convexity assumptions, \[ y\in \partial f(x) \quad \Longleftrightarrow \quad x\in \partial f^*(y). \] Moreover, when this holds, Fenchel–Young equality occurs: \[ f(x)+f^*(y)=\langle y,x\rangle. \] This bidirectional correspondence allows one to transfer information about optimality from the primal to the dual and back.

5.2 Differentiable case and gradient relationships

If \(f\) is differentiable at \(x\), then \(\partial f(x)=\{\nabla f(x)\}\), and the conjugate satisfies a gradient swap: \[ \nabla f(x)=y \quad \Longleftrightarrow \quad \nabla f^*(y)=x, \] when the relevant regularity and invertibility conditions hold (commonly ensured by strict convexity and suitable smoothness). The conjugate then acts like a dual coordinate change between gradients.

5.3 Strict convexity and smoothness correspondence

A classic duality in regularity states:

  • If \(f\) is strictly convex, then \(f^*\) is essentially differentiable on the interior of its domain.
  • If \(f\) is differentiable with a Lipschitz gradient (or other smoothness properties), then \(f^*\) exhibits strong convexity or related curvature bounds.

While exact formulations depend on norms and topologies, the overarching idea is that curvature and regularity exchange roles under conjugation.

5.4 Essential smoothness and essential strict convexity (overview)

In nonsmooth analysis, “essential” variants accommodate behavior near the boundary of the effective domain. Essential smoothness typically requires that \(f\) becomes steep as it approaches boundary points, ensuring that subgradients do not blow up without control. Essential strict convexity forbids flat segments on the effective domain except possibly at the boundary. Under these conditions, conjugation yields a correspondence: essential smoothness of \(f\) parallels essential strict convexity of \(f^*\), providing a robust framework for non-Euclidean and extended-real-valued settings.

6 Operations under the transform

6.1 Transform of sums and infimal convolution

Conjugation converts sums into infimal convolutions. Roughly, for suitable \(f\) and \(g\), \[ (f+g)^* = f^* \,\square\, g^*, \] where the infimal convolution is \[ (h_1\square h_2)(y)=\inf_{y_1+y_2=y}\bigl(h_1(y_1)+h_2(y_2)\bigr). \] This identity reflects how taking a supremum of an expression containing \(-f-g\) can be reorganized into an infimum over how the dual variable splits between conjugate parts.

6.2 Infimal convolution and conjugate identities

The reverse identity also appears under appropriate hypotheses, enabling computations of conjugates for functions expressed via infimal convolution. Infimal convolution is particularly useful in regularization theory, where it models combining penalties or merging constraints. In many cases, it yields conjugates with tractable forms, especially when one of the functions is an indicator or a norm-based term.

6.3 Scaling and translation rules

Two basic rules help with algebraic manipulations:

  • Scaling: If \(f_\alpha(x)=\alpha f(x/\alpha)\) for \(\alpha>0\) in appropriate vector spaces, then \(f_\alpha^*\) scales correspondingly in \(y\).
  • Translation by a linear term: If \(f(x)+\langle a,x\rangle\) is formed, then the conjugate shifts by the dual variable: \((f(\cdot)+\langle a,\cdot\rangle)^*(y)=f^*(y-a)\).

These rules follow by substituting variables in the defining supremum.

6.4 Composition with linear maps (pushforward/pullback rules)

If \(A:X\to Z\) is linear and one studies \(f\circ A\), conjugation interacts with the adjoint \(A^*:Z^*\to X^*\). A typical relationship is \[ (f\circ A)^*(y)=f^*(u)\quad \text{with a constraint depending on }y \text{ through }A^*. \] In practice, dual constraints emerge naturally: composing with \(A\) “pushes” the primal variable through \(A\), and the dual variable appears through \(A^*\). The resulting expressions often combine \(f^*\) with indicator functions enforcing feasibility of the dual representation.

7 Regularity, existence, and edge cases

7.1 Proper vs improper functions

Conjugacy theory is cleanest for proper functions, typically meaning \(f\not\equiv +\infty\) and \(f(x)>-\infty\) for all \(x\). If \(f\) is improper (for instance, identically \(+\infty\)), then \(f^*\) becomes degenerate (often identically \(-\infty\) or \(+\infty\), depending on conventions). Properness ensures the supremum defining \(f^*\) is meaningful.

7.2 Lower semicontinuity requirements

Even when \(f\) is convex, lack of lower semicontinuity can prevent \(f^{}\) from matching \(f\). The biconjugate theorem explains the remedy: \(f^{}\) replaces \(f\) by its closed convex hull. This is essential in applications, because optimization problems often require lower semicontinuity to guarantee existence of minimizers and stability under limits.

7.3 Cases with empty supremum and infinities

The supremum in the definition may yield \(+\infty\) if the affine functional \(\langle y,x\rangle-f(x)\) is unbounded above. Conversely, if \(f(x)\) is \(+\infty\) on the effective domain relevant to the supremum, then the supremum may be taken over an empty effective region in effect, producing \(f^*(y)=-\infty\) under extended-value conventions. Many frameworks avoid \(-\infty\) outcomes by restricting to proper convex functions.

7.4 When conjugates fail to recover the original function

If \(f\) is not convex, or is convex but not closed, then \(f^{**}\) will typically differ from \(f\). The discrepancy can be interpreted as:

  • convexification: replacing \(f\) by its convex envelope,
  • closure: taking the lower semicontinuous hull.

In short, conjugation does not preserve arbitrary functions; it preserves the maximal closed convex structure compatible with \(f\).

8 Applications in analysis and optimization

8.1 Variational principles and energy methods

Many energy functionals in analysis can be written in terms of convex terms plus linear couplings. Conjugation then produces dual energies and yields inequalities that help estimate minimizers. This is frequently used to derive bounds, show existence of solutions, and establish stability with respect to perturbations.

8.2 Maximum entropy / duality viewpoints (conceptual)

In statistical mechanics and information theory, optimization problems can be framed as maximizing entropy subject to constraints. Conjugate duality offers a way to transform such constrained maximization into an unconstrained or lower-dimensional minimization involving conjugate functions (often log-partition functions). Conceptually, the conjugate captures how constraints in the primal translate into potentials in the dual.

8.3 Large deviations intuition (conceptual connection)

In large deviations theory, rate functions are often convex and arise as conjugates of cumulant-generating functions. The Fenchel–Legendre transform appears because the asymptotic exponential scaling turns moment generating properties into a dual variational problem. This provides intuition: rare-event probabilities correspond to specific slopes in the conjugate representation.

8.4 Deriving Euler–Lagrange-type conditions via duality

For convex problems, optimality conditions can be expressed through subgradient inclusions rather than classical derivatives. Using conjugates and the Fenchel equality condition, one obtains stationarity relations resembling Euler–Lagrange equations but written in dual form. These conditions identify when a candidate pair \((x,y)\) simultaneously minimizes the primal and maximizes the dual.

9.1 Relation to support functions and Minkowski functionals

Support functions \(\sigma_C\) and Minkowski functionals (gauges) are closely connected through conjugation. For convex sets containing the origin, the gauge of a set and the support function of its polar set can be paired via the Fenchel–Legendre transform. This builds a dictionary between geometric objects (polars, balls, convex hulls) and analytic ones (norms, conjugates).

In modern analysis, the subdifferential \(\partial f\) of a convex function is a canonical example of a monotone operator. Conjugation links monotone operators via duality: subgradients of \(f\) correspond to subgradients of \(f^*\). This connection is a cornerstone in splitting methods and operator-theoretic formulations of optimization.

9.3 Comparison with other dual transforms (overview)

Other transforms, such as the Legendre transform in classical mechanics, the Laplace transform in asymptotic analysis, or the convex conjugate in different conventions, share structural similarities but differ in domain, assumptions, and interpretation. The Fenchel–Legendre transform is distinguished by its generality for nonsmooth convex functions and its compatibility with extended-real values and lower semicontinuity.

10 Illustrative worked problems

10.1 Computing \(f^\*\) for a quadratic function

Let \(f(x)=\frac{1}{2}\|x\|_2^2\) on \(\mathbb{R}^n\). Then

\[

f^*(y)=\sup_x\left(\langle y,x\rangle-\frac{1}{2}\|x\|_2^2\right).

\] Complete the square: \[

\langle y,x\rangle-\frac{1}{2}\|x\|_2^2
= -\frac{1}{2}\|x-y\|_2^2+\frac{1}{2}\|y\|_2^2.

\] The supremum is attained at \(x=y\), giving \[

f^*(y)=\frac{1}{2}\|y\|_2^2.

\] This self-conjugacy reflects the symmetry of the quadratic energy.

10.2 Conjugate of an indicator function set

Let \(f(x)=\delta_C(x)\). Then \[ f^*(y)=\sup_{x\in X}\bigl(\langle y,x\rangle-\delta_C(x)\bigr) =\sup_{x\in C}\langle y,x\rangle=\sigma_C(y). \] So computing the conjugate of an indicator reduces to finding the support function of the set \(C\), a geometric quantity.

10.3 Using subgradients to identify optimizer

Suppose \(f\) is closed and convex and consider the maximization defining \(f^*(y)\): \[ f^*(y)=\sup_x\bigl(\langle y,x\rangle-f(x)\bigr). \] If an optimizer \(x^\star\) exists and achieves the supremum, then equality in the Fenchel inequality holds at \((x^\star,y)\), yielding \[ y\in \partial f(x^\star). \] Conversely, any \(x^\star\) for which \(y\in \partial f(x^\star)\) attains the supremum. In computations, one can therefore find the dual maximizer by solving a subgradient inclusion rather than performing a global supremum.

10.4 Infimal convolution example

Let \(f=\delta_{C}\) and \(g=\delta_{D}\). Their conjugates are support functions: \[ f^*=\sigma_C,\qquad g^*=\sigma_D. \] Since \(f+g=\delta_{C\cap D}\) when the intersection is nonempty (and is \(+\infty\) outside), one finds \[ (f+g)^*=\sigma_{C\cap D}. \] On the other hand, the sum identity gives \[ (f+g)^* = f^*\square g^* = \sigma_C \square \sigma_D, \] which translates into a dual statement about how support functions combine through infimal convolution. The equivalence illustrates how operations in the primal (intersection/indicator addition) become nontrivial but structured operations in the dual.