The following Wikipedia page uses $x_\varepsilon (t) = x(t) + \varepsilon (t)$ in the proof.


But in my mechanics book (by David Morin), the author uses something similar to $x_\varepsilon (t) = x(t) + \varepsilon \eta (t), \ \varepsilon \in \mathbb{R}$.

Since the proof using the concept of $\delta \mathcal S$ is written on the Wikipedia page I won't go through that.

Consider the following partial derivative of the action functional (whichever epsilon it is).

$$\frac{\partial \mathcal S}{\partial \varepsilon} = \int_{t_1}^{t_2} \frac{\partial \mathscr L}{\partial \varepsilon} \ dt = \int_{t_1}^{t_2} \frac{\partial \mathscr L}{\partial x_\varepsilon}\frac{\partial x_\varepsilon}{\partial \varepsilon} + \frac{\partial \mathscr L}{\partial x'_\varepsilon}\frac{\partial x'_\varepsilon}{\partial \varepsilon} \ dt $$

Note that I used the notation $x'$ to indicate the time derivative.

If I put $x_\varepsilon (t) = x(t) + \varepsilon \eta (t)$ here then it works and I get the Euler-Lagrange equation.

But if I substitute $x_\varepsilon (t) = x(t) + \varepsilon (t)$ into the integral, I obtain

$$ \frac{\partial x_\varepsilon}{\partial \varepsilon} = 1 $$ and $$ \frac{\partial x'_\varepsilon}{\partial \varepsilon} = \frac{\partial \varepsilon '}{\partial \varepsilon} = 0$$

which lead to an absurd result.


  • 1
    $\begingroup$ Are you familiar with functional derivatives? $\endgroup$
    – fqq
    Jul 18, 2020 at 9:47
  • $\begingroup$ Why do you think $\displaystyle\frac{\partial \varepsilon'}{\partial \varepsilon}=0$? $\endgroup$
    – user258881
    Jul 18, 2020 at 10:13
  • $\begingroup$ @FakeMod Because that is the way we treat time derivatives when dealing with partial derivatives? For example, the partial derivative of the Lagrangian $\frac{1}{2}mv^2 - V(x)$ with respect to x is $- dV/dx$ because, obviously, the kinetic energy term is treated as a constant. $\endgroup$ Jul 18, 2020 at 13:44

1 Answer 1


Recall that we are looking for a function $x(t)$, such that first order changes in it will induce second order changes in $S[x(t)]$. These changes in $x(t)$ can be expressed in two different ways, as you have identified:

  1. $$x_{\epsilon}(t) = x(t) + \epsilon \eta(t)$$ where $\epsilon$ tracks the "smallness" and $\eta(t)$ is any (suitable) arbitrary function. In this set-up, we seek a condition on $x(t)$ such that $\delta S = 0 + \mathcal{O}(\epsilon^2)$ for all $\eta(t)$.

  2. $$x_{\epsilon}(t) = x(t) + \epsilon(t)$$ where now $\epsilon(t)$ is any suitable arbitrary function and is used to track the "smallness" (i.e. the size of the variation of $x'(t)$ from $x(t)$). Here we also seek a condition on $x(t)$ such that $\delta S = 0 + \mathcal{O}(\epsilon^2)$ but the "for all $\eta$" is incorporated into the nature of $\epsilon$.

Viewing $S$ as a function of $\epsilon$ for the purpose of variation, $S(\epsilon) = S(0) + \frac{\mathrm{d}S}{\mathrm{d}\epsilon}\bigr|_{\epsilon = 0} \epsilon + \mathcal{O}(\epsilon^2)$ implies that we require $\frac{\mathrm{d}S}{\mathrm{d}\epsilon}\bigr|_{\epsilon = 0} = 0$.

  1. $$S(\epsilon) = S(0) + \int \mathrm{d}t \left(\frac{\partial L}{\partial x} \eta + \frac{\partial L}{\partial x'} \eta' \right) \epsilon + \mathcal{O}(\epsilon^2)$$. Therefore, we require:

$$ \frac{\mathrm{d}S}{\mathrm{d}\epsilon} = \int \mathrm{d}t \left(\frac{\partial L}{\partial x} \eta + \frac{\partial L}{\partial x'} \eta' \right) = 0$$ as in Morin's derivation.

  1. $$S(\epsilon) = S(0) + \int \mathrm{d}t \left(\frac{\partial L}{\partial x} \epsilon + \frac{\partial L}{\partial x'} \epsilon' \right) + \mathcal{O}(\epsilon^2)$$

Like before, we also require the integrand in the second term to vanish, but in its current form, $\frac{\mathrm{d}S}{\mathrm{d}\epsilon}$ doesn't naturally drop out like in case 1, since we have an $\epsilon'$. However, using integration by parts (like we would need to in case 1) will reduce this integrand to a term proportional to $\epsilon$ again, so like in case 1, we'll find that $\frac{\mathrm{d}S}{\mathrm{d}\epsilon}\bigr|_{\epsilon = 0} = 0$ gives us the Euler-Lagrange equations.

Side Note:

I'm not quite sure here how to justify why I've written $\mathcal{O}(\epsilon^2)$ over $\mathcal{O}(\epsilon^2, \epsilon^{'2})$ - i.e. I've made the assumption that terms of order $\epsilon'$ are of order $\epsilon$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.