2

I have understood that in general, given two normed spaces $(X,|\cdot |_X)$ and $(Y,|\cdot |_Y)$, an open set $U\subset X$ in $X$ and a function $f:U\to Y$, one has that:

$$f':U\to \mathcal{L}(X,Y)$$ $$f'':U\to\mathcal{L}(X,\mathcal{L}(X,Y))$$

that as an abstract concept is clear. Now, if we take in particular $X=Y=\mathbb{R}$ and $f(x)=x^3$, we have that $f':\mathbb{R}\to \mathcal{L}(\mathbb{R},\mathbb{R})$ is $f’(x;h_1)=(3x^2)h_1$. I can't do the next step for $f''(x)$. When in cassical real analysis we write that $f''(x)=6x$, who is the map $f'':\mathbb{R}\to \mathcal{L}(\mathbb{R},\mathcal{L}(\mathbb{R},\mathbb{R}))$?

Edit: I found that in this general sense, the second derivative of $x^3$ can't exists, because if it existed we would have:

$$f''(x_0)(h_1,h_2)=6x_0\cdot h_1$$

that obviously is not simmetric in $h_1,h_2$.

PS: $\mathcal{L}(X,Y)$ is the set of ontinuous linear mappings from $X$ into $Y$.

Arctic Char
  • 16,972

1 Answers1

3

Let me use the notation $Df$ and $D^2f$ to represent those maps and use $f'$ and $f''$ to mean the typical single variable notation for derivatives. Note that instead of thinking of $D^2f$ as a map $U \to \mathcal{L}(X,\mathcal{L}(X,Y))$, I think it is sometimes simpler to think of it as a map $U \to \mathcal{L}^2(X;Y)$ (the space of continuous bilinear maps $X\times X \to Y$; there is in fact a natural isometric (when equipped with operator norm) isomorphism between these spaces).

Then, of course, for all $x \in \Bbb{R}$, $f(x) = x^3$ implies $f'(x) = 3x^2, f''(x) = 6x, f'''(x), f^{(4)}(x) = 0$. In terms of the Frechet derivatives, these are: \begin{align} \begin{cases} Df:\Bbb{R} \to \mathcal{L}(\Bbb{R}; \Bbb{R}),\quad Df_x[h]&= 3x^2h\\ D^2f:\Bbb{R} \to \mathcal{L}^2(\Bbb{R}; \Bbb{R}),\quad D^2f_x[h_1, h_2]&= 6x h_1 h_2\\ D^3f:\Bbb{R} \to \mathcal{L}^3(\Bbb{R}; \Bbb{R}),\quad Df_x[h_1, h_2, h_3]&= 6h_1h_2h_3 \end{cases} \end{align} for $k \geq 4$, $D^4f = 0$ is identically $0$. From these expressions it is clear that for every $k \in \Bbb{N}$, and every $x \in \Bbb{R}$, $D^kf_x$ is a symmetric continuous multilinear mapping $\underbrace{\Bbb{R} \times \dots\times \Bbb{R}}_{\text{$k$ times}} \to \Bbb{R}$.


Notice that $f'(x) = Df_x[1]$ and $f''(x) = D^2f_x[1,1]$. In general for any sufficiently differentiable function $f: \Bbb{R} \to \Bbb{R}$, we will have $f^{(k)}(x) = D^kf_x[\underbrace{1, \dots, 1}_{\text{$k$ times}}]$, or equivalently, \begin{align} D^kf_x[h_1, \dots, h_k] &= f^{(k)}(x) \cdot h_1 \cdots h_k. \end{align} (Try a proof by induction if you want to be super rigorous).


Edit: Responding to OP's comment

Let's prove a slightly more general theorem. If we have a function $f:\Bbb{R} \to Y$, we shall use the notation $f'(x)$ to mean the limit $\lim_{h\to 0}\dfrac{f(x+h)-f(x)}{h}$. Since the domain is $\Bbb{R}$ and the target space is a normed vector space $Y$, it makes sense to talk about this limit. I shall take for granted that you know the following fact:

For $x \in \Bbb{R}$, $f'(x)$ exists if and only if $Df_x$ exists in which case, $Df_x[h] = f'(x) \cdot h$.

This shouldn't be too hard to prove; if I remember correctly, its only a 3-4 line proof. Now, we can state the conclusion of the theorem slightly differently. Define the map $T: Y \to \mathcal{L}(\Bbb{R}, Y)$ as \begin{align} T(\alpha) &:= (h\mapsto h\cdot \alpha) \end{align} This can easily be checked to be a linear isometric isomorphism from $Y$ onto $\mathcal{L}(\Bbb{R},Y)$ (equipped with operator norm). With this notation, we can state the conclusion of the theorem as:

$f'(x)$ exists if and only if $Df_x$ exists in which case, $Df_x = T(f'(x))$, or equivalently, $Df = T \circ f'$.

Now, how do we calculate $D^2f_x$? We simply use the chain rule and the fact that derivatives of linear transformations are themselves: \begin{align} D^2f_x &= DT_{f'(x)} \circ D(f')_x \\ &= T \circ D(f')_x \tag{$T$ is linear} \\ &= T \circ [T(f''(x))], \end{align} where in the last line I applied the highlighted result to $f'$. In this notation, $D^2f_x$ is an element of $\mathcal{L}(\Bbb{R}, \mathcal{L}(\Bbb{R},Y))$. Now evaluate first on $h_1$ then evaluate on $h_2$, and then you'll see that by how $T$ is defined, \begin{align} (D^2f_x[h_1])[h_2] &= f''(x) \cdot h_1 h_2. \end{align} Or if we abuse notation slightly and refer to $D^2f_x$ as the associated bilinear continuous map, then \begin{align} D^2f_x[h_1, h_2] &= f''(x) \cdot h_1 h_2. \end{align}

I leave it to you to prove inductively the following theorem:

For a function $f: U \subset \Bbb{R} \to Y$ ($U$ an open set), and for any integer $k\geq 0$, and any $x \in U$, the "usual derivative" $f^{(k)}(x)$ exists if and only if $D^kf_x$ exists (doesn't matter if you think of this as multilinear or not, because these two interpretations are related simply by application of a linear isomorphism... which doesn't affect differentiability). In this case, \begin{align} D^kf_x[h_1, \dots, h_k] &= f^{(k)}(x) \cdot h_1 \cdots h_k. \end{align}

peek-a-boo
  • 65,833
  • Deriving recursively w.r.t. the generic vectors $h_1$ and $h_2$, I obtain (in your notation) $D^2f_x(h_1,h_2)=6xh_1$ and not $6xh_1h_2$. I have done calculation more times, and I always obtain this result. Could I ask you, please, to let me see how did you obtain your expression for $D^2f_x(h_1,h_2)$? – Nameless Jul 07 '20 at 19:47
  • @Nameless I have provided a slightly more abstract proof which is easily generalizable (the details using induction left to you), hopefully that helps. But notice also that $Df_x[h] = 3x^2 h$ means that $Df_x = 3x^2 \cdot \text{id}_{\Bbb{R}}$, from here it is also possible to give a direct calculation from the definition (using various error estimates etc... but this is I think very tedious) that $(D^2f_x[h_1])(h_2) = 6x h_1 h_2$. – peek-a-boo Jul 07 '20 at 20:15
  • Thanks to your reply, I also found my error. So now, using the definition:

    $$D^2f_x(h_1,h_2)=D_{h_1}D_{h_2}f(x)=\lim_{t_1\to 0} \frac{D_{h_2}f(x+t_1 h_1)-D_{h_2}f(x)}{t_1}=$$ $$=\lim_{t_1\to 0} \frac{\lim_{t_2\to 0}\frac{f(x+t_2h_2+t_1 h_1)-f(x+t_1 h_1)}{t_2}-\lim_{t_2\to 0}\frac{f(x+t_2h_2)-f(x)}{t_2}}{t_1}=$$ $$=\lim_{t_1\to 0} \frac{3(x+t_1h_1)^2h_2-3x^2h_2}{t_1}=6xh_1h_2$$

    Thank you very much!

    – Nameless Jul 08 '20 at 05:07
  • PS: I can't give you the bounty until the next 14 hours, you unfortunately have to wait a bit, sorry. – Nameless Jul 08 '20 at 05:13