The calculus of variations is the branch of analysis concerned with finding functions that minimize or maximize a quantity that depends on an entire function, rather than on a finite set of variables. In ordinary calculus, one seeks a number \(x\) that minimizes a function \(f(x)\). In the calculus of variations, one seeks a function \(y(x)\) that minimizes an integral of the form
\[ J[y] = \int_a^b F(x, y(x), y'(x))\,dx, \]
where the input is the whole curve \(y\), and the output \(J\) is a single number. Such an object, a function whose argument is itself a function, is called a functional. The central problem of the field is to determine which functions make a given functional stationary—that is, where small perturbations of the function produce no first-order change in the value of the functional—and then to decide whether such stationary functions are minimizers, maximizers, or neither.
The subject is not a collection of isolated tricks but a coherent mathematical theory with a central equation, a set of rigorous existence and regularity results, and a web of connections to geometry, physics, and optimization. Its development has been driven by two interacting traditions: one rooted in the geometric and physical questions that motivated the field, and another rooted in the analytic machinery needed to make those questions precise.
The foundational result of the calculus of variations is the Euler–Lagrange equation. Suppose \(y(x)\) is a smooth function defined on an interval \([a,b]\), with fixed endpoint values \(y(a)=ya\) and \(y(b)=yb\). Consider the functional
\[ J[y] = \int_a^b F(x, y, y')\,dx, \]
where \(F\) is a given smooth function of three variables. If \(y\) is a minimizer of \(J\) among all smooth functions with those fixed endpoints, then \(y\) must satisfy the second-order ordinary differential equation
\[ \frac{\partial F}{\partial y} - \frac{d}{dx}\left(\frac{\partial F}{\partial y'}\right) = 0. \]
This is the Euler–Lagrange equation. The derivation is a direct analogue of setting a derivative to zero in ordinary calculus. One considers a one-parameter family of competitor functions \(y(x) + \epsilon \eta(x)\), where \(\eta\) vanishes at the endpoints. Substituting into \(J\), differentiating with respect to \(\epsilon\), and setting the result to zero at \(\epsilon=0\) yields the equation after an integration by parts. The condition that \(\eta\) is arbitrary forces the integrand of the resulting integral to vanish identically.
The Euler–Lagrange equation is necessary but not sufficient for a minimum. A solution of the equation is called an extremal; it may be a minimizer, a maximizer, or a saddle point. Determining which requires additional conditions, analogous to the second-derivative test in ordinary calculus. The second variation of the functional—the quadratic term in the expansion of \(J[y+\epsilon\eta]\)—plays this role. A condition called the Legendre condition (the positivity of \(\partial^2 F/\partial y'^2\) along the extremal) is necessary for a minimum, and a strengthened version, together with the absence of conjugate points, is sufficient. These ideas were developed in the nineteenth century, most notably by Carl Gustav Jacob Jacobi, and they form the classical local theory.
The Euler–Lagrange equation is the workhorse of the subject. Many physical laws can be expressed as the statement that some quantity is stationary: the path of a light ray minimizes travel time (Fermat's principle), the shape of a hanging chain minimizes gravitational potential energy, and the trajectory of a mechanical system makes the action integral stationary (Hamilton's principle). In each case, the Euler–Lagrange equation derived from the relevant functional reproduces the known governing equation.
The Euler–Lagrange equation gives a necessary condition, but it does not by itself guarantee that a minimizer exists. A functional may have extremals that are not minimizers, or it may have no minimizer at all. The classical approach—write down the equation, solve it, and check the second variation—works when solutions are smooth and the domain is well-behaved, but it fails in many natural problems.
The direct method in the calculus of variations, initiated by David Hilbert around 1900 and developed rigorously by Leonida Tonelli in the 1920s and 1930s, addresses existence directly. The idea is to avoid the Euler–Lagrange equation entirely and instead prove that a minimizer exists by compactness and lower semicontinuity arguments. One takes a minimizing sequence of functions—a sequence whose functional values approach the infimum—and shows that a subsequence converges to a limit function, and that the functional is lower semicontinuous, meaning that the limit's functional value is no larger than the limit of the values along the sequence. If both hold, the limit is a genuine minimizer.
The direct method requires choosing a function space in which the minimizing sequence has a convergent subsequence. The natural setting is a Sobolev space, a space of functions that are integrable together with their weak derivatives. Sobolev spaces are complete, and bounded sets in them are compact in weaker topologies, which makes them ideal for compactness arguments. The functional must satisfy a coercivity condition—growing at least linearly in the derivative—to ensure that minimizing sequences are bounded, and a convexity condition in the derivative variable to ensure lower semicontinuity.
The direct method transformed the field. It shifted the focus from solving differential equations to proving existence and regularity of minimizers in broad classes of problems. It also revealed that minimizers need not be smooth: for many natural functionals, the minimizer is only piecewise smooth, or even discontinuous. The study of the regularity of minimizers—under what conditions a minimizer of a variational problem is necessarily smooth—became a major research program in its own right, associated with the work of Ennio De Giorgi, Jürgen Moser, and others in the mid-twentieth century. The celebrated De Giorgi–Nash theorem, which established Hölder regularity for solutions of certain elliptic equations, grew directly out of this program.
The direct method and the regularity theory together created a productive tension. On one hand, the classical theory assumed smoothness and derived equations. On the other hand, the direct method showed that minimizers often exist in spaces of nonsmooth functions, and the regularity theory asked when those minimizers are in fact smooth. This tension is not a historical accident but a structural feature of the subject.
A central example is the least gradient problem, where the functional is
\[ J[u] = \int_\Omega |\nabla u|\,dx, \]
with prescribed boundary values. The direct method proves existence of a minimizer in the space of functions of bounded variation, but the minimizer may have jump discontinuities along curves. The Euler–Lagrange equation, which would be \(\nabla \cdot (\nabla u / |\nabla u|) = 0\), is only meaningful where the gradient does not vanish, and it says nothing about the behavior of the minimizer on the set where the gradient is zero. Understanding such problems required the development of a theory of nonsmooth minimizers, including the notion of variational inequalities and the study of free boundary problems, where the boundary of the region on which the solution is smooth is itself unknown and must be determined as part of the problem.
This line of work connects the calculus of variations to the theory of partial differential equations in a deep way. Many elliptic and parabolic equations can be viewed as the Euler–Lagrange equations of appropriate functionals, and the regularity theory for minimizers translates directly into regularity for solutions of those equations. Conversely, the need to handle nonsmooth data and nonsmooth solutions in PDE theory pushed the calculus of variations toward more general function spaces and more delicate compactness arguments.
A second major tradition within the calculus of variations comes from mechanics. The Euler–Lagrange equation is a second-order ordinary differential equation, but it can be transformed into a system of first-order equations through the Legendre transform. Define the momentum
\[ p = \frac{\partial F}{\partial y'}, \]
and the Hamiltonian
\[ H(x, y, p) = p\,y' - F(x, y, y'), \]
where \(y'\) is expressed in terms of \(p\) by inverting the defining relation. Then the Euler–Lagrange equation is equivalent to Hamilton's equations:
\[ \frac{dy}{dx} = \frac{\partial H}{\partial p}, \qquad \frac{dp}{dx} = -\frac{\partial H}{\partial y}. \]
This reformulation, developed by William Rowan Hamilton in the 1830s, reveals a deep symmetry structure. The Hamiltonian formulation is the natural language for classical mechanics, and it connects the calculus of variations to symplectic geometry, the geometry of phase space. The flow of Hamilton's equations preserves a symplectic form, and this preservation is the geometric content of the variational principle.
The Hamiltonian perspective also leads to the Hamilton–Jacobi equation, a partial differential equation for a function \(S(x,y)\) that generates the solution of the variational problem. The Hamilton–Jacobi equation is central to the modern theory of optimal control and dynamic programming, where the value function—the minimum cost achievable from a given state—satisfies a Hamilton–Jacobi equation. In the calculus of variations, the Hamilton–Jacobi equation provides a way to characterize the value of the functional along extremals, and it is the bridge between the variational problem and the theory of first-order PDEs.
A further development in this direction is the theory of viscosity solutions, introduced by Michael Crandall and Pierre-Louis Lions in the 1980s. The Hamilton–Jacobi equation does not in general have smooth solutions, even when the variational problem is well-posed. Viscosity solutions provide a notion of weak solution that is unique and stable, and they are now the standard framework for Hamilton–Jacobi equations. The connection to the calculus of variations is direct: the value function of a variational problem is often the viscosity solution of the associated Hamilton–Jacobi equation.
The calculus of variations in the present day is a broad and interconnected subject. Several research programs, while not replacing the classical core, have reshaped the field's boundaries and tools.
Gamma-convergence, introduced by Ennio De Giorgi in the 1970s, is a notion of convergence for functionals that is designed to preserve minimizers. If a sequence of functionals \(J\varepsilon\) gamma-converges to a limit functional \(J\), then minimizers of \(J\varepsilon\) converge (along a subsequence) to a minimizer of \(J\). This framework is the standard tool for studying variational problems with a small parameter, such as thin films, homogenization, and phase transitions. It provides a rigorous way to pass from a microscopic description to a macroscopic effective model.
Optimal transport is a variational problem with a long history—Gaspard Monge posed it in 1781—that has undergone a renaissance since the 1990s. The problem is to move mass from one distribution to another at minimal cost, where the cost depends on the distance moved. The modern theory, developed by Yann Brenier, Luigi Ambrosio, Cédric Villani, and others, connects the calculus of variations to measure theory, partial differential equations, and geometry. The key object is the Wasserstein distance between probability measures, which is defined through an optimal transport problem. The theory has found applications in economics, image processing, and machine learning, and it has generated new variational techniques, such as the use of gradient flows in the space of measures.
Calculus of variations on manifolds extends the classical theory to problems where the unknown function takes values in a manifold rather than in Euclidean space. The prototype is the harmonic map problem, where one minimizes the energy
\[ E[u] = \int_\Omega |\nabla u|^2\,dx \]
among maps \(u: \Omega \to M\) into a Riemannian manifold \(M\). Harmonic maps are central in geometry and mathematical physics, and their regularity theory—which is substantially more delicate than in the Euclidean case—has been a major research area since the 1960s, associated with the work of Karen Uhlenbeck, Richard Schoen, and others.
The field also continues to interact with the calculus of variations in the classical sense through the study of minimal surfaces, surfaces that locally minimize area. The Plateau problem—finding the surface of least area spanning a given boundary curve—was one of the original motivations for the subject, and it remains an active area. The modern theory, developed by Herbert Federer, Wendell Fleming, Enrico Bombieri, and others, uses geometric measure theory to handle surfaces that may have singularities, and it has deep connections to the regularity theory of variational problems.
The calculus of variations is unified by a single question—what functions make a given integral stationary or minimal—and by a single method—perturb a candidate function and examine the resulting change in the functional. The Euler–Lagrange equation, the direct method, the Hamiltonian formulation, and the modern theories of gamma-convergence and optimal transport are all elaborations of this basic idea. The field's history is not a succession of rival schools but a progressive broadening of the class of problems that can be treated: from smooth functions on intervals, to nonsmooth functions on domains, to measures on manifolds. Each new development has extended the reach of the variational principle without invalidating the earlier results, which remain true within their original scope.
The subject's enduring importance lies in its role as a bridge. It translates geometric and physical questions into analytic ones, and it provides a common language for problems that arise in mechanics, geometry, economics, and data science. The central questions—does a minimizer exist, is it unique, is it smooth, and how does it depend on the data—are the same questions that motivated the field's founders, and they continue to drive its current research.