Loading field map
The calculus of variations asks a deceptively simple question: among all possible functions satisfying certain boundary conditions, which one minimizes or maximizes a given quantity—a functional? Unlike ordinary calculus, which finds stationary points of a function of finitely many variables, the calculus of variations seeks stationary functions in an infinite-dimensional space. This shift from points to functions forced mathematicians to develop entirely new ways of thinking about existence, smoothness, and constraints. The history of the subfield is a story of successive expansions: from local differential conditions to global existence arguments, from smooth to nonsmooth problems, and from unconstrained to dynamically constrained optimization.
The classical framework, forged in the eighteenth and nineteenth centuries, treated variational problems as a natural extension of ordinary calculus. The brachistochrone problem—finding the curve of fastest descent under gravity—motivated the Bernoulli brothers, Euler, and Lagrange to derive the first necessary condition for a minimizer: the Euler–Lagrange equation. This differential equation, satisfied by any sufficiently smooth extremal function, became the workhorse of the field. Jacobi later added a second-variation condition to distinguish minima from other stationary points, and Weierstrass sharpened the analysis with necessary conditions involving the corner condition and the field theory of extremals.
Classical calculus of variations assumed that minimizers were smooth and that the functional was sufficiently differentiable. Its methods were local: they examined small perturbations around a candidate solution and derived conditions that any interior minimizer must satisfy. The Dirichlet principle—that a minimizer of the Dirichlet integral exists because the integral is bounded below—was accepted on physical intuition long before it was rigorously justified. By the end of the nineteenth century, the classical framework had produced a powerful toolkit for solving problems in geometry (minimal surfaces, geodesics) and physics (least action principles), but it could not guarantee existence when the minimizer was not smooth or when the functional lacked differentiability.
Around 1900, Hilbert and others recognized that the classical approach could not answer the most basic question: does a minimizer exist at all? The classical Euler–Lagrange equation only gave candidates; it did not prove that a minimizer existed in the first place. Hilbert's 1900 Paris lecture on the Dirichlet principle revived the idea that existence could be proved by global arguments rather than by solving differential equations. This shift marked the birth of the calculus of variations "in the large," a framework that studies variational problems using topological and geometric methods.
Calculus of variations in the large does not rely on local differential conditions. Instead, it uses global properties of the space of admissible functions—such as compactness, connectedness, and the topology of level sets—to prove that minimizers exist or to count their number. Morse theory, developed by Morse in the 1920s and 1930s, relates the critical points of a functional to the topology of the underlying manifold, providing a global picture of the variational landscape. This framework coexists with the classical one: the classical Euler–Lagrange equation remains a tool for characterizing critical points once existence is established, but the emphasis shifts from solving differential equations to understanding the global structure of the problem.
Concurrent with the rise of global topological methods, a second new framework emerged: the direct methods. Tonelli, in the early twentieth century, developed a functional-analytic approach that proved existence directly, without passing through differential equations. The core idea is to construct a minimizing sequence, show that it has a convergent subsequence in a suitable weak topology, and then prove that the limit is indeed a minimizer. This requires two properties of the functional: coercivity (the functional grows fast enough to prevent minimizers from escaping to infinity) and lower semicontinuity (the functional does not jump upward under weak limits).
The direct methods absorbed the existence goals of the calculus of variations in the large but replaced topological arguments with functional-analytic ones. They also transformed the role of the classical Euler–Lagrange equation: instead of being the primary existence tool, it became a regularity tool. Once the direct methods prove that a minimizer exists in a weak sense (for example, in a Sobolev space), one can ask whether that minimizer is smooth enough to satisfy the Euler–Lagrange equation classically. This separation of existence from regularity is a hallmark of the modern approach. The direct methods remain the dominant framework for proving existence in variational problems today, especially in partial differential equations and geometric measure theory.
By the mid-twentieth century, variational methods faced a new challenge: how to handle problems where the admissible functions are constrained by differential equations and control variables. Classical calculus of variations could treat simple constraints via Lagrange multipliers, but it could not handle inequality constraints on the control or state variables that arise in engineering and economics. Pontryagin's maximum principle (1956) provided the answer. It extended the Euler–Lagrange formalism to systems governed by ordinary differential equations, introducing an adjoint variable (the costate) and a Hamiltonian function. The maximum principle gives necessary conditions for optimal control, analogous to the Euler–Lagrange equation but capable of handling control constraints.
Optimal control theory did not replace the earlier frameworks; it coexists with them as a specialized extension. When the control problem is smooth and unconstrained, the maximum principle reduces to the classical Euler–Lagrange equation. The bridge between optimal control and core variational methods is explicit: the maximum principle can be derived from the calculus of variations by considering variations in the control, and many modern textbooks treat optimal control as a chapter within the calculus of variations. However, optimal control theory developed its own research program—focused on dynamic programming, the Hamilton–Jacobi–Bellman equation, and numerical methods—that goes beyond the traditional variational toolkit.
The direct methods assume that the functional is at least lower semicontinuous and often convex. But many important problems—friction, contact mechanics, plasticity, and certain economic models—involve functionals that are not differentiable or even convex. In the 1960s and 1970s, Moreau, Rockafellar, and others developed nonsmooth analysis to handle these cases. The key innovation was the concept of the subdifferential: a set-valued generalization of the derivative for convex functions. For nonconvex problems, Clarke's generalized gradient extended the idea further.
Nonsmooth analysis and variational inequalities form a framework that both extends and challenges the direct methods. It extends them by providing existence and optimality conditions for problems where the direct methods' convexity assumptions fail. It challenges them by showing that the smooth regularity theory—the step that follows existence in the direct methods—may not apply: minimizers can have corners, jumps, or fractal-like structures. Variational inequalities, introduced by Lions and Stampacchia, reformulate many nonsmooth problems as inequalities rather than equations, linking the calculus of variations to partial differential equations with obstacles or free boundaries.
This framework remains active and is now integrated with the direct methods: many modern existence proofs use nonsmooth analysis to handle lower-order terms or constraints, then apply direct methods to the smooth part of the functional. The classical Euler–Lagrange equation survives as a heuristic: even when a minimizer is not smooth, one can often derive a differential inclusion or a variational inequality that plays the same role.
Today, the calculus of variations is a pluralistic field. The direct methods are the default toolkit for proving existence in most variational problems, from minimal surfaces to nonlinear elasticity. Calculus of variations in the large provides the geometric and topological perspective needed for problems with multiple critical points or nontrivial topology. Optimal control theory handles constrained dynamical systems and is a separate but overlapping discipline. Nonsmooth analysis supplies the tools for problems where differentiability fails.
What the leading frameworks agree on is the functional-analytic foundation: existence is proved via weak convergence and lower semicontinuity, not by solving differential equations. The classical Euler–Lagrange equation is understood as a regularity result, not an existence result. What they disagree on is the level of generality: direct methods typically require convexity or quasiconvexity, while nonsmooth analysis relaxes those assumptions at the cost of weaker conclusions. Calculus of variations in the large and optimal control theory each emphasize different structures—topology versus dynamics—and their methods do not always transfer easily. This division of labor is productive: each framework is best at the problems it was designed for, and the boundaries between them are increasingly porous as researchers combine techniques from multiple frameworks to solve new problems.