Problem: minimise $F(x,y,y')$ over $x$, constrained by $G(x,y,y')=0$.

$$J_1(x,y,y')=\large \int_{x_0}^{x_1}F(x,y,y')+ \lambda (x) G(x,y,y')dx$$

I understand the Euler-Lagrange equation and Lagrange multipliers in multivariable (i.e. not variational) calculus, but am having a hard time putting them together, I don't understand the logic behind this equation.

  1. Is this the same $\lambda$ that appears in the multivariable $\nabla f (x,y)= \lambda \nabla g (x,y) $? If so, why is it not also a function of $y$?

  2. By the fundamental lemma of the calculus of variations, doesn't $ \int_{x_0}^{x_1}\lambda (x) G(x,y,y')dx=\int_{x_0}^{x_1}\lambda (x) 0dx=0$, thus $J=J_1$, and so the constraint has had no effect on the integral?

  3. More importantly, what is the proof that if the minimum of $J_1$ is found (i.e. the Euler-Lagrange equations are satisfied with this new integrand) $G$ will be constrained properly?

    In short: what is the logic behind that integrand?

Side-note: Without constraints, the problem is simply to minimise $$J(x,y,y')=\large \int_{x_0}^{x_1}F(x,y,y')dx$$ i.e. solve the Euler-Lagrange equation for the functional $F$ and the variable $y$. (Side-question: why is $G$ usually given as a function of only $x$ and $y$, and $J$ of $y$? Surely in $J$'s case this severely limits the number of possible functionals $F$ [as the integral must not be a function of $x$ or $y'$], and in $G$'s case there are many possible constraints involving $y'$ that are overlooked?).

  • $\begingroup$ On 3.: That's not the case; you need to solve the resulting Euler-Lagrange equations together with the constraint equation; the constraint equation isn't automatically satisfied. You've introduced one more unknown, $\lambda$, and you have one more equation to determine it, just as in the multivariable case. $\endgroup$
    – joriki
    Jan 15 '13 at 20:31
  • $\begingroup$ @joriki Thank you. Why not just solve the 'normal' (nonconstrained) Euler-Lagrange equation and the constraint together (what is special about the 'new' Euler- Lagrange equation?)? I suppose that may be one of the cruces of my bafflement. $\endgroup$
    – Meow
    Jan 15 '13 at 20:37
  • 1
    $\begingroup$ You can't -- without $\lambda$, you don't have the additional freedom required to satisfy them both. If you just solve the normal Euler-Lagrange equation, the solution will generally not fulfil the constraint. It's exactly analogous to the multivariable case. $\endgroup$
    – joriki
    Jan 15 '13 at 20:40

I think you have some problems, because you use an incorrect notation. Let me rewrite your original problem: \begin{align*} \text{minimize}\quad & J(y) = \int_{x_0}^{x_1} F(x, y(x), y'(x)) \, \mathrm{d} x \\ \text{subject to}\quad & G(x, y(x), y'(x)) = 0 \quad\text{for all } x \in [x_0,x_1]. \end{align*} Here, $F : \mathbb{R} \times \mathbb R \times \mathbb R \to \mathbb R$ and $G : \mathbb R \times \mathbb R \times \mathbb R \to \mathbb R^n$. Do you see the differences? $J$ only depends on the function $y$, whereas the integrand $F$ and the constraint $G$ depend on real numbers.

Now (if a constraint qualification is satisfied), you get a multiplier $\lambda : [x_0, x_1] \to \mathbb R^n$ (compare with section 6.2 in your link: you get a multplier for each constraint, that is, for each $x$), such that the derivative of the Lagrangian $$J(y) + \int_{t_0}^{t_1} G(x, y(x), y'(x)) \, \lambda(x) \, \mathrm{d} x$$ with respect to $y$ is zero (that is, the derivative of your lagrangian w.r.t. the optimization variable). Now, you can continue like for the derivation of the euler-lagrange equation.

@2.: Yes, this is correct, but the derivative of $J$ w.r.t. $y$ does not equals the derivative of $J_1$ w.r.t. $y$.

@3.: As joriki already said, you have to solve the resulting Euler-Lagrange equation together with the constraint. In other words: The Euler-Lagrange equation depends on $\lambda$. Once you have fixed $\lambda$, they have a unique solution $y$ (depending on $\lambda$). It remains to choose $\lambda$, such that the corresponding $y$ satisfies the constraints (this is reasonable, since you have as many constraints as degrees of freedom in $\lambda$).

  • $\begingroup$ Thanks for the thorough answer, but I have one further question: if we know $G(x,y(x),y'(x))=0$, why don't we just use this knowledge: insert it into $J_1$ and make out lives easier? This doesn't seem incorrect, but implies that $J=J_1$, which is obviously untrue as the unconstrained problem is very different from the constrained one. $\endgroup$
    – Meow
    Jan 22 '13 at 18:54
  • 2
    $\begingroup$ Your $J_1$ is the Lagrangian and $J$ is the objective. Note that $J_1$ depends on your optimization variable and additionally on your multiplier $\lambda$. You stated correctly $J_1(y, \lambda) = J(y)$ iff $y$ satisfies the constraints. However, the multiplier rule tells you to compute the zeros of the derivative of $J_1$. But this differs from the derivative of $J$. $\endgroup$
    – gerw
    Jan 23 '13 at 11:33

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.