1
$\begingroup$

I'm fairly new to functional calculus but and posting here since the question seems more appropriate than for MSE. When coming across this post I could not help but wonder the following.

Let $H$ be the reproducing-kernel Hilbert space obtained by completing the set of all $C^1(\mathbb{R}^n,\mathbb{R}^m)$ with finite norm finite: $$ \|f(x)\|_H:= \|f(0)\|_{\mathbb{R}^m} + \int_{x \in \mathbb{R}^n} \|(\nabla f)(x)\|_{\mathbb{R}^m} e^{-\|x\|} dx. $$

If $w:C^1(\mathbb{R}^n,\mathbb{R}^m)\times \mathbb{R}^n\rightarrow [0,\infty)$ is the functional $$ (f,x) \mapsto \|f(x)\|_{\mathbb{R}^m}, $$ what is its Fréchet derivative? Thinking analogously to the linked post and appealing to the chain-rule for Fréchet derivatives, I would guess it is $$ Dw(f,x) (g,y)= \frac1{\|g(x)+J_f(y)\|}\left(g(x) + (J_f)(y)\right). $$ However, I don't know how to show more than this (if even it is a correct ansatz).

$\endgroup$

1 Answer 1

1
$\begingroup$

The general procedure for the identification of a Fréchet derivative is the following

  1. Calculate the functional derivative of the given functional, then
  2. verify its linearity and
  3. verify its continuity respect to the topology that is considered on the domain of the given functional i.e., for a Banach or Hilbert space topology, verify that the norm of the derivative does not depend on the structure of the variation but only on its size (norm).

The functional $w$ is defined on a vector space structure defined on $C^1(\Bbb R^n, \Bbb R^m)\times \Bbb R^n$, since we should be able to give a meaning to the word "linear", and the topology considered on this vector space is the product topology between the Banach space topology on $\Bbb R^n$ and the topology by $H$ on $C^1(\Bbb R^n,\Bbb R^m)$: following the above list we have $$ \begin{split} Dw\big[(f,x)\big]\big((g,y)\big) & \triangleq \frac{\mathrm{d}}{\mathrm{d}\varepsilon}w\big[(f,x)+\varepsilon(g,y)\big]\bigg{|}_{\varepsilon = 0}\\ &=\frac{\mathrm{d}}{\mathrm{d}\varepsilon}w\big[(f+\varepsilon g, x+\varepsilon y)\big]\bigg|_{\varepsilon = 0}\\ &=\frac{\mathrm{d}}{\mathrm{d}\varepsilon}\big\|f(x+\varepsilon y)+\varepsilon g(x+\varepsilon y)\big\|\bigg|_{\varepsilon = 0}\\ &=\left.\frac{\mathrm{d}}{\mathrm{d}\varepsilon}\bigg[\sum_{i=1}^m \Big(f_i(x+\varepsilon y)+\varepsilon g_i(x+\varepsilon y)\Big)^2\bigg]^\frac{1}{2}\right|_{\varepsilon = 0}\\ &=\frac{1}{2}{\big\|f(x+\varepsilon y)+\varepsilon g(x+\varepsilon y)\big\|}^{-1} \left.\frac{\mathrm{d}}{\mathrm{d}\varepsilon}\bigg[\sum_{i=1}^m \Big(f_i(x+\varepsilon y)+\varepsilon g_i(x+\varepsilon y)\Big)^2\bigg]\right|_{\varepsilon = 0}\\ &=\frac{1}{2}{\big\|f(x+\varepsilon y)+\varepsilon g(x+\varepsilon y)\big\|}^{-1}\\ &\qquad\cdot \left.\frac{\mathrm{d}}{\mathrm{d}\varepsilon}\bigg[\sum_{i=1}^m f_i^2(x+\varepsilon y)+ 2\varepsilon f_i(x+\varepsilon y)g_i(x+\varepsilon y) +\varepsilon^2 g_i^2(x+\varepsilon y)\bigg]\right|_{\varepsilon = 0}\\ &={\big\|f(x+\varepsilon y)+\varepsilon g(x+\varepsilon y)\big\|}^{-1}\\ &\qquad\cdot \bigg[\sum_{i=1}^m \langle\nabla f_i(x+\varepsilon y),y\rangle+ f_i(x+\varepsilon y)g_i(x+\varepsilon y) \\ &\qquad\qquad +\varepsilon\langle\nabla f_i(x+\varepsilon y),y\rangle g(x+\varepsilon y) +\varepsilon f_i(x+\varepsilon y) \langle \nabla g_i(x+\varepsilon y),y\rangle \\ &\qquad\qquad\qquad + \varepsilon g_i^2(x+\varepsilon y) + \left.\varepsilon^2 g_i(x+\varepsilon y)\langle \nabla g_i(x+\varepsilon y),y\rangle\bigg]\right|_{\varepsilon = 0}\\ &={\big\|f(x)\big\|}^{-1}\bigg[\sum_{i=1}^m \langle\nabla f_i(x),y\rangle +f_i(x)g_i(x)\bigg]=\frac{\langle 1, \mathbf{J}_f(x)y\rangle+\langle f(x),g(x)\rangle }{\big\|f(x)\big\|} \end{split} $$ Thus, apart from errors, we have done step 1 and checked the linearity as required by step 2. Regarding step 3, we see that that if $$ \|f(x)\|>0 \iff f(x)\neq 0 $$ for the given $x\in\Bbb R^n$, then the functional derivative norm depend only on the value $\|g(x)\|_H+\|y\|_{\Bbb R^n}$ and not on the structure of the element $(g,y)$. Thus the functional derivative of $w$ is a Fréchet derivative.

$\endgroup$
9
  • $\begingroup$ Great answer Daniele. May I ask, have you ever come across this space/ is it a well-studied object? $\endgroup$
    – ABIM
    Aug 24, 2020 at 10:14
  • $\begingroup$ Thank you, @AnnieLeKatsu. Regarding the linear product space $C^1(\Bbb R^n, \Bbb R^m)\times\Bbb R^n$ I am not aware of other uses and I haven’t seen it used in the literature. It’s a very particular space since it is the product between a infinite dimensional Hilbert space and a finite dimensional one, and I constructed (or better guessed)the structure of its topology from the one you gave to $C^1$ and the standard $\Bbb R^n$ one. $\endgroup$ Aug 24, 2020 at 10:43
  • $\begingroup$ Oh I meant $C^1$ with topology induced by the inner product I provided. $\endgroup$
    – ABIM
    Aug 24, 2020 at 11:12
  • 1
    $\begingroup$ @AnnieLeKatsu, better "Weighted Hilbert space". In general Hörmander spaces are (weighted) Hilbert spaces only if their associated Lebesgue space is $L^2$: otherwise they are only Banach. By googling "Weighted Hilbert space" you should be able to choose from a wide range of constructions and find what you need. $\endgroup$ Aug 25, 2020 at 4:32
  • 1
    $\begingroup$ @AnnieLeKatsu I suspect it is not so: as a matter of fact $$ (f+\varepsilon g)\circ (f +\varepsilon g)\neq f\circ f+\varepsilon g\circ g.$$ However, despite the presumably higher complexity of calculations, the same method for the calculation of functional derivative is applicable. Just the computations are more tedious. $\endgroup$ Sep 11, 2020 at 19:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.