Interpreting the Euler-Lagrange Equations as the Gradient of the Action Functional

We study the smooth path spaces of Euclidean spaces $\mathbb{R}^N$, as diffeological spaces. We show that the tangent spaces of the free path space $\mathscr{P}$ are isomorphic to $\mathscr{P}$ itself, and that the tangent spaces of the space $\mathscr{P}_{\mathbf{p}, \mathbf{q}}$ of paths with fixed endpoints $\mathbf{p}$ and $\mathbf{q}$ are isomorphic to the smooth loop space of $\mathbb{R}^N$ based at the origin. We also define cotangents and gradients of smooth maps from these path spaces, and then show that, in the case of the action functional which arises in the calculus of variations, the gradient is precisely the path formed out of the terms of the Euler-Lagrange equations. We show that solutions of the Euler-Lagrange equations correspond precisely to the zeros of the gradient, and also provide analogous interpretations for the constrained Euler-Lagrange equations. This gives an illuminating geometric perspective on these equations. Finally, we illustrate the theory with several concrete examples from geometry, mechanics and machine learning.