Lagrangian Dual Transform in Optimization

• Lagrangian Dual Transform is a framework that establishes a precise mapping between primal perturbation functions and dual Lagrangians via inf-sup transformations.
• It embeds minimization problems into families of perturbed setups using Rockafellians and recovers classical dual functions through convex conjugation.
• Under convexity and lower semicontinuity, the method guarantees strong duality and enables exact two-way reconstruction between primal and dual optimization representations.

The Lagrangian Dual Transform (LDT) provides a bidirectional dictionary between Lagrangians and Rockafellians, generalizing Fenchel duality to settings involving perturbation functions and dual vector spaces. The method embeds minimization problems into families of perturbed problems via Rockafellians and connects them to classical duality theory using inf-sup transformations and convex conjugates. Under convexity and lower semicontinuity, LDT enables exact two-way reconstruction between primal and dual representations, extending classical duality techniques.

1. Formal Definitions and Structural Elements

Let $X$ denote the decision or uncertainty set, $\mathcal U$ the primal perturbation space, and $\mathcal V$ its dual, paired by a bilinear form $\langle u,v\rangle$.

Rockafellian (Perturbation Function):

Given an objective $f: X \to \overline{\RR}$, the Rockafellian is defined as

$R: X \times \mathcal U \to \overline{\RR}, \qquad R(x, 0) = f(x), \,\, \forall x \in X,$

so that $R(x,u)$ extends $f(x)$ to a perturbed family. The associated value function or perturbation function is

$\varphi: \mathcal U \to \overline{\RR}, \qquad \varphi(u) = \inf_{x \in X} R(x, u),$

with $\varphi(0) = \inf_x f(x)$.

Bilinear Coupling:

The spaces $\mathcal U, \mathcal V$ are paired by a bilinear map $\langle u,v\rangle : \mathcal U \times \mathcal V \to \RR$.

Lagrangian:

From the Rockafellian, the Lagrangian is built as

$L: X \times \mathcal V \to \overline{\RR}, \qquad L(x,v) = \inf_{u \in \mathcal U} \{ R(x,u) - \langle u,v\rangle \}.$

Each $L(x,\cdot)$ is concave and upper semicontinuous in $v$.

Dual Function and Problem:

The dual function is

$\psi: \mathcal V \to \overline{\RR}, \qquad \psi(v) = \inf_{x \in X} L(x,v),$

yielding the dual problem: $$\sup_{v \in \mathcal V} \psi(v) = \sup_v \inf_x L(x, v).$$ Fenchel–Moreau theory guarantees weak duality,

$\sup_v \psi(v) \leq \varphi(0) = \inf_x f(x).$

2. Infimum–Supremum Transformations and Mutual Recoverability

Bidirectional transforms relate the Rockafellian and Lagrangian via the Fenchel–Moreau biconjugation principle. Specifically,

$$R(x, u) = \sup_{v \in \mathcal V} \{ L(x,v) + \langle u, v \rangle \},$$

$$L(x, v) = \inf_{u \in \mathcal U} \{ R(x, u) - \langle u, v \rangle \},$$

for all $(x, u) \in X \times \mathcal U$. These formulas establish that $R$ and $L$ are mutually recoverable through complementary inf-sup transforms, forming a complete “dictionary” between perturbation-based and Lagrangian dual perspectives.

3. Strong Duality and Two-Way Reconstruction under Convexity

A central result is the exact two-way LDT under convexity (Theorem 3.1 in (Lara, 2022)): Suppose $R: X \times \mathcal U \to \overline{\RR}$ is convex, lower semicontinuous (lsc), and proper in $u$ for fixed $x$. Then:

• Dual Equality:

$$\sup_{v \in \mathcal V} \psi(v) = \varphi^{**}(0) = \varphi(0) = \inf_{x \in X} f(x),$$

where $\varphi^{**}$ is the lsc-convex closure (biconjugate) of $\varphi$. If $\varphi$ is convex lsc, strong duality holds.

• Two-Way Reconstruction:

$$L(x,v) = \inf_{u} \{ R(x,u) - \langle u,v\rangle \}, \qquad R(x,u) = \sup_{v} \{ L(x,v) + \langle u,v\rangle \},$$

establishing that each object uniquely determines the other under these regularity hypotheses.

Applying the convex-conjugate theorem for $u \mapsto R(x,u)$ leads to

$$\left(R(x,\cdot)\right)^*(v) = \sup_u \{ \langle u, v \rangle - R(x,u) \} = -L(x,v)$$

and

$R(x,u) = \left(-L(x,\cdot)\right)^{**}(u).$

Global duality follows through conjugation of the value function $\varphi(u) = \inf_x R(x, u)$.

4. Forward and Backward Transform Table

The primary relationships between Rockafellians and Lagrangians can be succinctly presented as follows:

From $R$ to $L$	From $L$ to $R$
$L(x,v) = \inf_u \{ R(x,u) - \langle u,v \rangle \}$	$R(x,u) = \sup_v \{ L(x,v) + \langle u,v \rangle \}$
$\varphi(u) = \inf_x R(x,u)$	$\psi(v) = \inf_x L(x,v)$
$-L(x,\cdot) = (R(x,\cdot))^*$	$R(x,\cdot) = (-L(x,\cdot))^{**}$
$\psi = -\varphi^*$	$\varphi = (-\psi)^{**}$

All equalities hold under the convexity and lower semicontinuity assumptions specified in Theorem 3.1.

5. Illustrative Examples of the Lagrangian Dual Transform

Two canonical situations illustrate the machinery:

Example 1 (Shift Model):

Given $f: X \to \RR \cup \{+\infty\}$, convex and lsc,

$R(x,u) = f(x+u).$

Computation yields

$$L(x,v) = \inf_u \{ f(x+u) - \langle u,v\rangle \} = f^*(v) - \langle x,v\rangle.$$

The backward transform,

$$R(x,u) = \sup_v \{ f^*(v) - \langle x, v \rangle + \langle u, v \rangle \} = f(x+u).$$

The dual function is $\psi(v) = f^*(v)$, illustrating the classical Fenchel conjugate structure.

Example 2 (Equality-Constraint Lagrangian):

For $f(x)$ with constraint $A x = b$, $A \in \RR^{m \times n}$,

$$R(x, u) = f(x) + \delta_{\{Ax+u = b\}}(0) = \begin{cases} f(x) &\text{if } Ax + u = b, \ +\infty &\text{otherwise}, \end{cases}$$

yielding

$L(x,v) = f(x) + \langle b - Ax, v \rangle,$

and

$\sup_v \inf_x L(x,v) = \inf_{\{x: Ax = b\}} f(x).$

The reverse transform recovers the indicator form in the primal Rockafellian.

6. Significance and Generalizations

The Lagrangian Dual Transform, as unified by De Lara (Lara, 2022), equates the process of building a Lagrangian from a Rockafellian with the inverse direction, establishing the concept of Lagrangian-Rockafellian couples. These couples are characterized in terms of dual functions with respect to the bilinear pairing and also in generalized convex analytic terms. The LDT approach extends both classical Fenchel duality and Rockafellar's perturbation method, clarifying bidirectional transformations between primal and dual problems and supporting generalized convexity frameworks.

While duality between perturbation and dual functions is not always as clear cut as between the Lagrangian and Rockafellian constructs, the methodology provides a universal and explicit mechanism for recovering all principal dual objects from their primal counterparts in optimization theory.