1 Definition and basic setup

1.1 Fenchel–Legendre transform (convex conjugate)

Let \(f\) be a function on a real vector space. The convex conjugate of \(f\), also called the Fenchel–Legendre transform, is the function \(f^*\) defined by \[ f^*(y)=\sup_x \{\langle y,x\rangle - f(x)\}. \] It is constructed as the pointwise supremum over all affine functions of \(x\) parameterized by \(y\). In optimization settings, \(f^*\) can be interpreted as encoding the “best possible linear underestimates” of \(-f\), and it often turns constrained or difficult primal problems into dual problems with different structure.

1.2 Conjugate in vector spaces and pairing \(\langle y,x\rangle\)

The definition uses a pairing \(\langle y,x\rangle\) between elements \(y\) and \(x\). In finite-dimensional spaces, this pairing is frequently the Euclidean dot product \(y^\top x\). More generally, one may take \(x\) in a topological vector space and \(y\) in its (topological) dual so that \(\langle y,x\rangle\) is well-defined and continuous. The choice of pairing determines which dual objects and geometric notions correspond to the transform.

1.3 Extended-real-valued functions and the role of \(+\infty\)

To make the transform robust, \(f\) is typically allowed to take values in the extended real line \((-\infty,+\infty]\). In this setting, assigning \(f(x)=+\infty\) for infeasible points is convenient: it forces those points to contribute nothing to the supremum. This convention also allows constraints to be absorbed into the function itself, simplifying both theoretical statements and duality computations.

2 Properties of the convex conjugate

2.1 Convexity and closedness

For an arbitrary extended-real-valued \(f\), the conjugate \(f^*\) is always convex (or identically \(+\infty\) in degenerate cases). The conjugate “regularizes” the original function in the sense that it interacts naturally with the closed convex hull. A related refinement appears through biconjugation, where \(f^{**}\) recovers the appropriate closed convex envelope of \(f\).

2.2 Order-reversing and monotonicity aspects

The transform is order-reversing: if \(f\le g\) pointwise, then \(f^*\ge g^*\) pointwise. This monotonicity direction follows directly from the definition involving \(\langle y,x\rangle-f(x)\). Such relations are useful for comparing dual bounds derived from different primal relaxations.

2.3 Homogeneity and translation rules

Conjugation interacts predictably with scaling and shifting. For instance, scaling the argument of \(f\) by a linear factor can induce a reciprocal scaling in the conjugate under suitable conditions. Similarly, adding an affine term to \(f\) corresponds to a shift of the conjugate’s input; adding a constant to \(f\) shifts \(f^*\) by the same constant with sign changes reflecting the “\(-f(x)\)” structure in the definition. These rules are frequently applied to simplify expressions in derivations.

2.4 Conjugation of sums and infimal convolution

Conjugation converts certain sums into infimal convolutions. While the conjugate of a sum is not generally the sum of conjugates, there is a consistent dual relationship: \[ (f+g)^* = f^* \,\square\, g^*, \] where \(\square\) denotes infimal convolution. This identity clarifies how separate regularizers or components of an energy function combine on the dual side and is central in splitting methods for optimization.

2.5 Support functions and indicator functions as examples

Two standard families of examples illustrate the geometry of conjugation.

  • The indicator function of a set \(C\),

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

  • Conversely, support functions have conjugates that encode the closed convex set associated with the original geometry.

These correspondences make the transform a practical bridge between optimization constraints and dual variables.

3 The biconjugate and duality theorems

3.1 Fenchel–Moreau theorem (biconjugation)

The Fenchel–Moreau theorem states that, under standard assumptions, the biconjugate \(f^{}\) equals the lower semicontinuous convex hull of \(f\). More precisely, if \(f\) is proper and convex and lower semicontinuous, then \(f=f^{}\). If \(f\) is not convex or not closed, then \(f^{**}\) provides the closest “convex closed” representative below \(f\) in an appropriate sense.

3.2 Tightness: when \(f = f^{**}\)

Equality \(f=f^{**}\) is not automatic. It holds when \(f\) is already convex and lower semicontinuous (and proper). When it fails, the biconjugate can be understood as removing nonconvex features and filling “gaps” created by lack of closure. In optimization language, this means that dual representations derived via conjugates become exact precisely when the primal function has the right convex-analytic regularity.

3.3 Primal–dual gap and reformulations of optimization problems

For optimization problems of the form \[ \inf_x \{f(x)+g(Ax)\}, \] conjugation often yields a dual formulation whose value is related to the primal value. The difference between the two is the primal–dual gap. Under constraint qualifications or regularity conditions, the gap may vanish, giving strong duality. Even when strong duality does not hold, conjugates still provide computable dual bounds.

4.1 Relationship between subgradients \(\partial f\) and \(\partial f^*\)

Subgradients provide the differential-like objects for nonsmooth convex functions. A key dual relationship is: \[ y \in \partial f(x) \quad \Longleftrightarrow \quad x \in \partial f^*(y), \] under standard properness conditions. This symmetry explains why optimality conditions expressed through subgradients can be translated between primal and dual variables.

4.2 Equality conditions in Fenchel’s inequality

Fenchel’s inequality states that for all \(x\) and \(y\), \[ f(x)+f^*(y)\ge \langle y,x\rangle. \] Equality holds exactly when the pair \((x,y)\) satisfies a subgradient condition, typically \(y\in \partial f(x)\) (equivalently \(x\in \partial f^*(y)\)). Thus, the inequality not only provides bounds but also characterizes when primal and dual candidates are optimally consistent.

4.3 Characterizing maximizers/minimizers via conjugates

The supremum in the definition of \(f^*(y)\) may be attained for suitable \(x\). When maximizers exist, they satisfy first-order optimality conditions that can be written using subgradients. Likewise, minimizers of expressions involving \(f\) plus other convex terms can often be characterized by stationarity conditions on the dual. Conjugation therefore serves as a systematic way to translate “argmax/argmin” questions into subgradient inclusion statements.

4.4 Young’s inequality and its consequences

Young’s inequality is another name often used for Fenchel’s inequality: \[ \langle y,x\rangle \le f(x)+f^*(y). \] It has immediate consequences in analysis: it yields estimates in variational inequalities, supports stability bounds, and can be used to derive energy inequalities. In many settings, the inequality is tight at optimality and becomes a tool for proving convergence of iterative schemes in convex problems.

5 Computation of conjugates

5.1 Conjugate of common convex functions (templates)

Many conjugates can be computed using standard patterns: completing the square for quadratic forms, leveraging known conjugates of norms, or applying rule-based transformations such as scaling and translation. A practical approach is to start from elementary conjugates (linear, absolute value, indicator functions) and build more complex expressions using algebraic identities for conjugation.

5.2 Quadratic, norm, and absolute-value examples

- For a quadratic function \(f(x)=\frac12\|x\|^2\) (in a Hilbert space setting with the induced norm), the conjugate has a similar quadratic form: \(f^*(y)=\frac12\|y\|^2\).
- For absolute-value functions on \(\mathbb{R}\), conjugation produces indicator functions of intervals, reflecting that \(x\) is dual to the unit ball in \(\ell_\infty\)-type geometry.
  • For norms, conjugates often become indicators of dual norm balls or related support functions. This is frequently used in regularized optimization, where one regularizer’s conjugate directly describes feasible dual variables.

5.3 Indicator functions and constraints (duality interpretation)

Indicator functions turn constraints into conjugate objects that are easier to interpret geometrically. If a primal problem restricts \(x\) to a convex set \(C\), then including \(\delta_C(x)\) in the objective leads to a conjugate \(\delta_C^*(y)\) equal to the support function. Consequently, feasibility constraints become dual objective terms that measure alignment with direction \(y\).

5.4 Conjugates under linear transformations

When \(f\) is composed with a linear operator \(A\), conjugates transform accordingly. Identities of the type \[ (f\circ A)^*(y) \] are expressed using adjoints and appropriate domain restrictions. These formulas are central in deriving duals for problems with linear constraints or in modeling where \(Ax\) represents features, measurements, or coupling between variables.

6 Regularity and geometric interpretation

6.1 Smoothness vs. strong convexity duality

Conjugation exchanges curvature properties. In broad terms, smoothness of \(f\) relates to strong convexity of \(f^*\), and strong convexity of \(f\) corresponds to smoothness of \(f^*\). This duality is reflected in how Lipschitz constants of gradients map to strong convexity moduli across the transform. Such connections guide the choice of regularizers in algorithms and help predict stability.

6.2 Lipschitz continuity and bounded domain effects

If \(f\) grows at least linearly at infinity (or if it has bounded effective domain), then \(f^*\) may be differentiable or Lipschitz on its effective domain. Conversely, boundedness properties of \(f^*\) correspond to coercivity or growth of \(f\). These correspondences are often used to bound dual variables and to ensure well-posedness of saddle-point formulations.

6.3 Geometric viewpoint via supporting hyperplanes

Geometrically, \(f^*(y)\) represents the largest gap between \(\langle y,x\rangle\) and \(f(x)\). When \(f\) is convex, each subgradient \(y\in \partial f(x)\) defines a supporting hyperplane to the epigraph of \(f\). The dual variable thus indexes a family of supporting hyperplanes, and the transform can be visualized as encoding the epigraph’s tangent structure in dual coordinates.

6.4 Gauge functions, norms, and polarity

Gauge functions generalize norms by measuring the minimal scaling needed to reach a set. In many convex-analytic frameworks, gauges are closely tied to polarity: the polar set of a convex set determines the conjugate of its indicator or the support function of that polar. This yields a geometric interpretation of dual norms and explains why conjugation naturally expresses “polar” relationships between primal feasibility regions and dual constraints.

7 Legendre transform as a special case

7.1 Differentiable strictly convex functions

The classical Legendre transform typically assumes differentiability and strict convexity (on an open domain). Under these hypotheses, the supremum in the Fenchel–Legendre transform is achieved at a unique point where the gradient condition holds, making the mapping between variables resemble a change of coordinates via \(\nabla f\). In this smooth setting, the Legendre transform becomes a special case of the Legendre–Fenchel transform.

7.2 Domain restrictions and classical mechanics connection

In classical mechanics, the Legendre transform links a Lagrangian formulation to a Hamiltonian formulation through a change from velocity variables to momentum variables. Mathematically, this corresponds to transforming a function whose derivative is invertible on a suitable domain. Domain restrictions ensure that the gradient map is well-defined and that the relevant convexity assumptions hold locally.

7.3 Comparison of Legendre and Legendre–Fenchel transforms

The Legendre–Fenchel transform extends the classical construction to nonsmooth and non-strictly convex functions by using a supremum rather than a derivative-based change of variables. It remains applicable for extended-real-valued functions and naturally incorporates constraints through indicator functions. As a result, it is the standard tool in modern convex analysis, while the classical Legendre transform fits within the smoother subset where optimization and calculus coincide.

8 Applications in optimization and analysis

8.1 Fenchel duality in convex optimization

Fenchel duality provides a template for turning primal convex problems into dual ones by applying conjugation to terms in the objective. A common workflow is to rewrite the primal as \(\inf_x \{f(x)+g(Ax)\}\), then apply conjugates to obtain a dual \(\sup\)-problem in variables associated with \(f^*\) and \(g^*\). Duality theory clarifies when dual feasible points provide certified lower bounds (or upper bounds depending on sign conventions) for the primal.

8.2 Lagrangians, dual variables, and saddle-point structure

The dual variables in Fenchel duality can often be recognized as Lagrange multipliers. By expressing constraints through indicator functions and forming Lagrangians, conjugation yields saddle-point representations: \[ \inf_x \sup_y \; \langle y, Ax\rangle - f(x) - g^*(y), \] or closely related forms. Saddle-point structure underlies many algorithms, including primal–dual methods, and it connects convex conjugation to variational principles.

8.3 Proximal methods and Moreau envelopes (connections)

Moreau envelopes and proximal operators are frequently described using conjugates. The Moreau envelope regularizes a function by infimal convolution with a quadratic term, and the proximal operator can be expressed in terms of minimizing that regularized objective. Conjugation plays a role through identities that relate the proximal map of a function to the proximal map of its conjugate (with parameter adjustments). This connection helps analyze convergence and derive efficient update rules.

8.4 Variational problems and energy functionals

Many variational models in analysis and applied mathematics involve minimizing energy functionals that may include nonsmooth components, such as norms or constraints. By expressing these energies with convex functions and employing conjugates, one can derive dual variational principles. These dual formulations often simplify analysis, enable stronger a priori estimates, and provide alternative characterizations of minimizers.

9 Advanced topics

9.1 Infimal convolution and generalized duality constructions

Infimal convolution extends the idea of adding functions by minimizing over intermediate arguments: \[ (f \square g)(x) = \inf_{u}\{ f(u)+g(x-u)\}. \] Through conjugation, infimal convolution becomes equivalent to pointwise multiplication in an appropriately dual sense. This machinery supports generalized duality constructions for composite structures, including problems with multiple regularizers or layered constraints.

9.2 Conjugates of nonconvex or nonclosed functions (via closure)

Even if \(f\) is not convex, its conjugate depends only on the convexified closed version of \(f\). Specifically, biconjugation yields the greatest lower semicontinuous convex function not exceeding the original function. This allows one to work with \(f^*\) and \(f^{**}\) as a principled “convex regularization” mechanism, clarifying which features of a nonconvex model are retained in the dual view.

9.3 Measure-theoretic or infinite-dimensional settings

In infinite-dimensional spaces, conjugation requires careful attention to topology and dual pairings. One typically works with locally convex spaces and continuous duals, ensuring that the pairing \(\langle y,x\rangle\) is meaningful and that lower semicontinuity is taken with respect to the appropriate topology. With these precautions, Fenchel–Moreau and conjugation-based duality extend beyond finite-dimensional optimization.

9.4 Computational considerations and numerical approximations

Computing conjugates exactly may be difficult for complex functions. Numerical approaches often approximate the supremum or use discretization strategies, sometimes leveraging structure such as separability, piecewise linearity, or known proximal operators. In practical optimization software, conjugates are frequently not computed explicitly; instead, the algorithm uses equivalent dual operations (e.g., via proximal maps or support-function evaluations) that are more stable and efficient.