0

Let $f:\mathbb{R}^2 \rightarrow \mathbb{R}$ twice differentiable and for $t \in \mathbb{R}$ and for $x,y$ fixed define the function $$F(t) = f(x+t,y+t)+f(x-t,y-t)-f(x+t,y-t)-f(x-t,y+t)$$

In order to solve another problem I need to calculate $F''(t)$. I found (if my calculation is right) that $$F'(t) = f'(x+t,y+t)(1,1)+f'(x-t,y-t)(-1,-1)+f'(x+t,y-t)(-1,1)+f'(x-t,y+t)(1,-1).$$ I don't know how to proceed and calculate the second derivative, I appreciate any detailed help on how to calculate $F''(t)$

In the above we are using the notion of Fréchet derivative found in the wikipedia article https://en.wikipedia.org/wiki/Fr%C3%A9chet_derivative

Victor
  • 289
  • 2
    Since $;f;$ is a function in two variables, so $;f';$ doesn't really mean anything meaningful here, (unless you define something). You perhaps mean the partial derivatives of $;f;$ wrt $;x,y;$ and then apply some chain rule... – DonAntonio Jul 05 '19 at 20:01
  • In this exemple I'm using this notion of derivative https://en.wikipedia.org/wiki/Fr%C3%A9chet_derivative. – Victor Jul 05 '19 at 20:04
  • 2
    Oh, my! Then write so explicitly in your question ! – DonAntonio Jul 05 '19 at 20:12
  • 1
    Sorry, I thought that Fréchet derivative was the standard derivative notion for more general spaces :) – Victor Jul 05 '19 at 20:18
  • $F$ is a function of one real variable ($F:\mathbb{R} \rightarrow \mathbb{R}$). – Victor Jul 05 '19 at 20:28
  • If $F(t)$ is function of 1 variable, than derivative must be a real number (for every t). I wonder why did you get a vector ) – Eugene Sirkiza Jul 05 '19 at 20:29
  • Yes, in the above $F'(t)$ is a real number because for each $x$, $f'(x)$ is a linear transformation from $\mathbb{R}^2$ to $\mathbb{R}$ – Victor Jul 05 '19 at 20:32
  • Via isomorphism $f''(x)$ is a bilinear form from $\mathbb{R}^2 \times \mathbb{R}^2$ to $\mathbb{R}$, so the problem is to calculate $F''(t)$ in function of $f''$ (which is a bilinear form in each $x$). – Victor Jul 05 '19 at 20:35

2 Answers2

2

For a twice differentibale mapping $f:V \to W$ between Banach spaces, if $a\in V$, then by $df_a$ I shall mean the differential of $f$ at $a$ (Frechet derivative), which is a linear transformation from $V$ into $W$. Then $df_a(h)$ shall mean the value of the linear map $df_a$ on the vector $h \in V$. In general, the chain rule says the following: if $g:U \to V$ is another twice differentiable mapping between (open subsets of) Banach spaces, then the chain rule says that \begin{align} d(f \circ g)_a = df_{g(a)} \circ dg_a \end{align} Notice that composition $(\circ)$ on the RHS is a bounded bilinear map from $L(V,W) \times L(U,V) \to L(U,W)$, defined by $(T,S) \mapsto T \circ S$, so we can apply a "generalised product rule": for all $h \in U$, \begin{align} d^2(f \circ g)_a[h] = \left(d^2f_{g(a)}[dg_a(h)] \right) \circ dg_a + df_{g(a)} \circ d^2g_a(h) \end{align} (The above is an equality of elements in $L(U,W)$). If we now use an isomorphism (which I suppress in the notation) to think of second derivatives as bilinear maps, then fully evaluating everything, we get that for all $h,k \in U$, \begin{align} d^2(f \circ g)_a[h,k] = d^2f_{g(a)}[dg_a(h), dg_a(k)] + df_{g(a)}[d^2g_a(h,k)] \end{align} (This is now an equality of elements in $W$).


I wrote the above formula of the chain rule just so you know how things work. In your specific case, they're slightly easier. For convenience, I'll only do one term; the rest are similar. Let $\phi: \Bbb{R} \to \Bbb{R}$ be defined by $\phi(t) = f(x+t,y+t)$. Then, by the chain rule, \begin{align} \phi'(t) = df_{(x+t,y+t)}[(1,1)] \end{align} Now, we can either use the second derivative formula above (in which case the second term involving $d^2g_a$ will be zero because $(1,1)$ is just a constant vector) or we can argue directly as follows. Define the "evaluation at $(1,1)$ map" $\varepsilon: L(\Bbb{R}^2, \Bbb{R}) \to \Bbb{R}$ by $\varepsilon(T) = T[(1,1)]$. Then, $\varepsilon$ itself is a bounded linear map, so for any $T$, we have \begin{align} d \varepsilon_T = \varepsilon \tag{$*$} \end{align}

Using this, we can say \begin{align} \phi'(t) &= \varepsilon (df_{(x+t,y+t)}) \\ &= (\varepsilon \circ df)(x+t,y+t) \end{align} Hence, by another application of the chain rule, we find that \begin{align} \phi''(t) &= \left [d \varepsilon_{df_{(x+t,y+t)}} \circ d(df)_{(x+t,y+t)} \right] (1,1) \\ &= [\varepsilon \circ d^2f_{(x+t,y+t)}](1,1) \tag{by $(*)$}\\ &:= \left[d^2f_{(x+t,y+t)}(1,1) \right](1,1) \end{align} Now, if we make use of the isomorphism, to think of second differetials as a bilinear maps, we get that \begin{align} \phi''(t) = d^2f_{(x+t,y+t)} \left[ (1,1), (1,1)\right] \end{align}

I hope you can now use this logic to compute $F''(t)$.


Additional Remarks about Notation:

By the way, just to clarify the notation, if $f:V \to W$ if differentiable at $a$, then I use $df_a$ to mean the Frechet derivative. But in the special case where the domain $V$ is just $\Bbb{R}$, then $f$ is a mapping $\Bbb{R} \to W$, so the first differential is an element $df_a \in L(\Bbb{R},W)$. But note that $L(\Bbb{R},W)$ is naturally isomorphic to $W$ via the isomorphism $T \mapsto T(1)$. So, rather than writing $df_a(1)$, I rather use the notation $f'(a)$. This is no misuse of notation, because one can show that \begin{align} df_a(1) = \lim_{h \to 0} \dfrac{f(a+h) - f(a)}{h} \end{align} Hence the use of $f'(a)$ conforms with the single variable usage.

peek-a-boo
  • 55,725
  • 2
  • 45
  • 89
0

Fix me if I'm wrong, please.

$$F'(t) = \frac{\partial f}{\partial u}(x+t, y+t) + \frac{\partial f}{\partial v}(x+t, y+t) - \frac{\partial f}{\partial u}(x-t, y-t) - \frac{\partial f}{\partial v}(x-t, y-t) + \frac{\partial f}{\partial u}(x+t, y-t) - \frac{\partial f}{\partial v}(x+t, y-t) - \frac{\partial f}{\partial u}(x-t, y+t) + \frac{\partial f}{\partial v}(x-t, y+t)$$

$$F''(t) = \frac{\partial^2 f}{\partial u^2}(x+t, y+t) + 2\frac{\partial^2 f}{\partial u \partial v} (x+t, y+t) + \frac{\partial^2 f}{\partial v^2}(x+t, y+t) + \frac{\partial^2 f}{\partial u^2}(x-t, y-t) + 2\frac{\partial^2 f}{\partial u \partial v} (x-t, y-t) + \frac{\partial^2 f}{\partial v^2}(x-t, y-t) + \frac{\partial^2 f}{\partial u^2}(x+t, y-t) + \frac{\partial^2 f}{\partial v^2}(x+t, y-t) + \frac{\partial^2 f}{\partial u^2}(x-t, y+t) + \frac{\partial^2 f}{\partial v^2}(x-t, y+t)$$

  • I just checked in the answer in the back of my textbook and there says that $F''(t) = f''(x,y) ((4,0),(4,0)).$ – Victor Jul 05 '19 at 20:46
  • @VictorRafael, I suppose it's misunderstanding in designations. Can you plz give the name of book. – Eugene Sirkiza Jul 05 '19 at 20:50
  • The book i'm using is avaible only in Portuguese and is called Topologia e Análise no Espaço $\mathbb{R}^n$ - Ronaldo Freire de Lima, but this more general notion of derivative can be found on classics such as Bartle or Rudin. – Victor Jul 05 '19 at 20:52