Lagrange multipliers on Banach spaces

In the field of calculus of variations in mathematics, the method of Lagrange multipliers on Banach spaces can be used to solve certain infinite-dimensional constrained optimization problems. The method is a generalization of the classical method of Lagrange multipliers as used to find extrema of a function of finitely many variables.

The Lagrange multiplier theorem for Banach spaces

Let X and Y be real Banach spaces. Let U be an open subset of X and let f : U → R be a continuously differentiable function. Let g : U → Y be another continuously differentiable function, the constraint: the objective is to find the extremal points (maxima or minima) of f subject to the constraint that g is zero.

Suppose that u₀ is a constrained extremum of f, i.e. an extremum of f on

$g^{-1}(0)=\{x\in U\mid g(x)=0\in Y\}\subseteq U.$

Suppose also that the Fréchet derivative Dg(u₀) : X → Y of g at u₀ is a surjective linear map. Then there exists a Lagrange multiplier λ : Y → R in Y^∗, the dual space to Y, such that

$\mathrm{D}f(u_0)=\lambda \circ \mathrm{D}g(u_0).\text{(L)}$

Since Df(u₀) is an element of the dual space X^∗, equation (L) can also be written as

$\mathrm{D}f(u_0)={(\mathrm{D}g(u_0))}^{*}(\lambda ),$

where (Dg(u₀))^∗(λ) is the pullback of λ by Dg(u₀), i.e. the action of the adjoint map (Dg(u₀))^∗ on λ, as defined by

${(\mathrm{D}g(u_0))}^{*}(\lambda )=\lambda \circ \mathrm{D}g(u_0).$

Connection to the finite-dimensional case

In the case that X and Y are both finite-dimensional (i.e. linearly isomorphic to R^m and Rⁿ for some natural numbers m and n) then writing out equation (L) in matrix form shows that λ is the usual Lagrange multiplier vector; in the case n = 1, λ is the usual Lagrange multiplier, a real number.

Application

In many optimization problems, one seeks to minimize a functional defined on an infinite-dimensional space such as a Banach space.

Consider, for example, the Sobolev space $X=H_0^1([-1,+1];\mathbb{ℝ})$ and the functional $f:X\to \mathbb{ℝ}$ given by

$f(u)=\underset{-1}{\overset{+1}{\int }}{u}^{\prime }(x{)}^2\mathrm{d}x.$

Without any constraint, the minimum value of f would be 0, attained by u₀(x) = 0 for all x between −1 and +1. One could also consider the constrained optimization problem, to minimize f among all those u ∈ X such that the mean value of u is +1. In terms of the above theorem, the constraint g would be given by

$g(u)=\frac{1}{2}\underset{-1}{\overset{+1}{\int }}u(x)\mathrm{d}x-1.$

However this problem can be solved as in the finite dimensional case since the Lagrange multiplier $\lambda$ is only a scalar.

See also

• Pontryagin's minimum principle, Hamiltonian method in calculus of variations

References

• Luenberger, David G. (1969). "Local Theory of Constrained Optimization". Optimization by Vector Space Methods. New York: John Wiley & Sons. pp. 239–270. ISBN 0-471-55359-X. 
• Zeidler, Eberhard (1995). Applied functional analysis: Variational Methods and Optimization. Applied Mathematical Sciences 109. 109. New York, NY: Springer-Verlag. doi:10.1007/978-1-4612-0821-1. ISBN 978-1-4612-0821-1. https://link.springer.com/book/10.1007/978-1-4612-0821-1.  (See Section 4.14, pp.270–271.)