Jerry's Digital Garden

BibliographyComments
/Part 5: Calculus/Chapter 27: Differential Calculus

27.2 Differentiation With Respect to Sets

Definition 27.2.1 (Small).label Let $E, F$ be TVSs over $K \in \RC$, $\sigma \subset \mathfrak{B}(E)$ be a covering ideal, $r: E \to F$, and $n \in \natz$, then the following are equivalent:

  1. (1)

    For each $A \in \sigma$, $r(th)/t^{n} \to 0$ uniformly on $A$.

  2. (2)

    If $r_{t}(x) = r(tx)/t^{n}$, then $r_{t} \to 0$ as $t \to 0$ with respect to the $\sigma$-uniform topology on $F^{E}$.

  3. (3)

    For each $A \in \sigma$, $\seq{a_k}\subset A$, and $\seq{t_k}\subset K \setminus \bracs{0}$ with $t_{k} \to 0$ as $n \to \infty$, $r(t_{k}a_{k})/t_{k}^{n} \to 0$ as $n \to \infty$.

If the above holds, then $r$ is $\sigma$-small of order $n$.

The set $\mathcal{R}_{\sigma}^{n}(E; F)$ is the $K$-vector space of all $\sigma$-small functions of order $n$ from $E$ to $F$. For simplicity, $\mathcal{R}_{\sigma}(E; F)$ denotes $\mathcal{R}_{\sigma}^{1}(E; F)$.

Proposition 27.2.2.label Let $E, F$ be TVSs over $K \in \RC$ with $F$ being separated, $\sigma \subset \mathfrak{B}(E)$ be a covering ideal, and $\mathcal{R}_{\sigma}(E; F)$ be the space of $\sigma$-small functions, and $\mathcal{H}\subset B_{\sigma}(E; F)$ be a subspace, $(\mathcal{H}, \mathcal{R}_{\sigma}(E; F))$ is a system of derivatives and remainders.

Proof. Let $T \in B_{\sigma}(E; F)$ and suppose that there exists $V \in \cn_{E}(0)$ circled and $r \in \mathcal{R}_{\sigma}(E; F)$ such that $T|_{V} = r|_{V}$. For any $x \in V$, $\bracs{x}\in \sigma$, so $T(x) = \lim_{t \downto 0}T(tx)/t = 0$ as $F$ is separated.$\square$

Definition 27.2.3 ($\sigma$-Derivative).label Let $E, F$ be TVSs over $K \in \RC$ with $F$ being separated, $\sigma \subset \mathfrak{B}(E)$ be a covering ideal, $U \subset E$ be open, $f: U \to F$, and $x_{0} \in U$, then $f$ is $\tilde \sigma$-differentiable at $x_{0}$ if there exists $V \in \cn_{E}(0)$, $T \in B_{\sigma}(E; F)$, and $r \in \mathcal{R}_{\sigma}(E; F)$ such that

\[f(x_{0} + h) = f(x_{0}) + Th + r(h)\]

for all $h \in V$.

The linear map $T \in B_{\sigma}(E; F)$ is the $\tilde \sigma$-derivative of $f$ at $x_{0}$, denoted $D_{\tilde \sigma}f(x_{0})$. If $T \in L(E; F)$, then $f$ is $\sigma$-differentiable at $x_{0}$, and $T$ is the $\sigma$-derivative of $f$ at $x_{0}$.

Definition 27.2.4 (Differentiable).label Let $E, F$ be TVSs over $K \in \RC$ with $F$ being separated, $\sigma \subset \mathfrak{B}(E)$ be a covering ideal, $U \subset E$ be open, and $f: U \to F$, then $f$ is $\sigma$/$\tilde \sigma$-differentiable on $U$ if it is $\sigma$/$\tilde \sigma$-differentiable at every point in $U$. In which case, the map $D_{\sigma} f: U \to B_{\sigma}(E; F)$ is the $\sigma$/$\tilde \sigma$-derivative of $f$.

Definition 27.2.5.label Let $E, F$ be TVSs over $K \in \RC$ with $F$ being separated, $\sigma^{E}_{\text{Fin}}, \sigma^{E}_{c}, \sigma^{E}_{b} \subset 2^{E}$ be the collection of all finite, precompact, and bounded subsets, respectively, then differentiability with respect to $\sigma^{E}_{\text{Fin}}, \sigma^{E}_{c}, \sigma^{E}_{b}$ correspond to Gateaux, Hadamard, and Fréchet differentiability.

Proposition 27.2.6 (Chain Rule).label Let $E$, $F$, $G$, be TVSs over $K \in \RC$ with $F, G$ being separated, $\sigma \subset \mathfrak{B}(E)$ and $\tau \subset \mathfrak{B}(F)$ be covering ideals. If:

  1. (a)

    For any $r \in \mathcal{R}_{\sigma}(E; F)$ and $T \in L(F; G)$, $T \circ r \in \mathcal{R}_{\sigma}(E; G)$.

  2. (b)

    For any $r \in \mathcal{R}_{\sigma}(E; F)$, $T \in L(E; F)$, and $s \in \mathcal{R}_{\tau}(F; G)$, $s \circ (T + r) \in \mathcal{R}_{\sigma}(E; G)$.

then for any $U \subset E$ and $V \subset F$ open, $f: U \to V$ $\sigma$-differentiable at $x_{0} \in U$, $g: V \to F$ $\tau$-differentiable at $f(x_{0}) \in V$, $g \circ f: U \to F$ is $\sigma$-differentiable at $x_{0}$ with

\[D_{\sigma}(g \circ f)(x_{0}) = D_{\tau} g(f(x_{0})) \circ D_{\sigma} f(x_{0})\]

Proof [Proposition 4.5.2, BS17]. Since $g$ is $\tau$-differentiable at $f(x_{0})$, there exists $s \in \mathcal{R}_{\tau}(F; G)$ such that

\[g(f(x_{0}) + h) = g \circ f (x_{0}) + D_{\tau} g(f(x_{0}))h + s(h)\]

for all $h \in F$ such that $f(x_{0}) + h \in V$. By differentiability of $f$, there exists $r \in \mathcal{R}_{\sigma}(E; F)$ such that

\[f(x_{0} + h) = f(x_{0}) + D_{\sigma} f(x_{0})h + r(h)\]

for all $h \in E$ such that $x_{0} + h \in U$. Therefore for all $h \in E$ with $x_{0} + h \in U$,

\begin{align*}g \circ f(x_{0} + h)&= g \circ f (x_{0}) + D_{\tau} g(f(x_{0})) \circ D_{\sigma} f(x_{0})h \\&+ D_{\tau} g(f(x_{0})) \circ r(h) + s(D_{\sigma} f(x_{0})h + r(h))\end{align*}

where $D_{\tau} g(f(x_{0})) \circ r \in \mathcal{R}_{\sigma}(E; G)$ by assumption (a), and $d \circ (D_{\sigma} f(x_{0}) + r) \in \mathcal{R}_{\sigma}(E; G)$ by assumption (b).$\square$

Proposition 27.2.7 ([Corollary 4.5.4, BS17]).label Let $E$, $F$, $G$, be TVSs over $K \in \RC$ with $F, G$ being separated. If $\sigma \subset \mathfrak{B}(E)$ and $\tau \subset \mathfrak{B}(F)$ correspond to the following families of sets on $E$ and $F$:

  1. (1)

    Precompact sets.

  2. (2)

    Bounded sets.

then

  1. (1)

    For any $r \in \mathcal{R}_{\sigma}(E; F)$ and $T \in L(F; G)$, $T \circ r \in \mathcal{R}_{\sigma}(E; G)$.

  2. (2)

    For any $r \in \mathcal{R}_{\sigma}(E; F)$, $T \in L(E; F)$, and $s \in \mathcal{R}_{\tau}(F; G)$, $s \circ (T + r) \in \mathcal{R}_{\sigma}(E; G)$.

and by Proposition 27.2.6, $\sigma$-derivatives and $\tau$-derivatives satisfy the Chain rule.

Proof. (1): Let $A \in \sigma$ and $U \in \cn_{G}(0)$. Since $T$ is continuous, there exists $V \in \cn_{F}(0)$ such that $T(V) \subset U$. Since $r \in \mathcal{R}_{\sigma}(E; G)$, there exists $t > 0$ such that $r(sA)/s \in V$ for all $s \in (0, t)$. In which case, $T \circ r(sA)/s \in U$ for all $s \in (0, t)$.

(2): To show that $s \circ (T + r) \in \mathcal{R}_{\sigma}(E; G)$, it is sufficient to show that for every $A \in \sigma$, $\seq{t_n}\subset \real_{> 0}$ with $t_{n} \downto 0$ as $n \to \infty$, and $\seq{a_n}\subset A$,

\[\limv{n}\frac{1}{t_{n}}s \circ (T + r)(t_{n}a_{n}) = \limv{n}\frac{1}{t_{n}}s\braks{t_n\paren{Ta_n + \frac{r(t_{n}a_{n})}{t_{n}}}}= 0\]

Since $\bracs{t_n^{-1}r(t_na_n)|n \in \natp}$ is a convergent sequence, it is contained in a compact set. Thus

\[B = \bracs{Ta_n + \frac{r(t_{n}a_{n})}{t_{n}}\bigg | n \in \natp}\subset T(A) + \bracs{\frac{r(t_{n}a_{n})}{t_{n}}|n \in \natp}\]

is contained in a compact set if $A$ is compact, and bounded if $A$ is bounded. Given that $s \in \mathcal{R}_{\sigma}(E; F)$, $t^{-1}s(tx) \to 0$ as $t \downto 0$ uniformly on $B$. Therefore

\[\limv{n}\frac{1}{t_{n}}s\braks{t_n\paren{Ta_n + \frac{r(t_{n}a_{n})}{t_{n}}}}= 0\]

$\square$

Remark 27.2.1.label In Definition 27.2.1, the system $\sigma$ can be chosen based on the bornology of $E$, and the definition of small-ness depends exclusively on $\sigma$. As such, there is an apparent disconnect between differentiation and the topology of the domain.

Consider for example a Hilbert space equipped with its norm and weak topology. The norm itself is differentiable with respect to both topologies, because the bounded sets coincide. Moreover, the data for differentiability needs to only come from a neighbourhood of $0$ in the norm topology. As such, a function may be differentiable even if its domain is too small to have an interior.

A method of extending this sense of differentiability is to require that every extension of the function to some open set, or to the entire space is differentiable. Given that this paves way to confusion for related definitions of differentiability, this definition is not formally included here.

Theorem 27.2.8 (Interchange of Limits and Derivatives).label Let $E$ be a TVS over $K \in \RC$, $F$ be a separated locally convex space over $K$, $\sigma \subset \mathfrak{B}(E)$ be a covering ideal, $U \subset E$ be open, and $n \in \natp$. Let $\fF \subset 2^{\tilde D_\sigma^n(U; F)}$ be a filter such that:

  1. (a)

    There exists $f: U \to F$ such that $\fF \to f$ pointwise.

  2. (b)

    For each $1 \le k \le n$, there exists $f^{(k)}: U \to B^{k}_{\sigma}(E; F)$ such that for all $x \in U$, and $A \in \sigma$ with $x + [0, 1]A \subset U$, $D_{\sigma}^{k}(\fF) \to f^{(k)}$ uniformly on $x + [0, 1]A$.

then $f \in \tilde D_{\sigma}^{n}(U; F)$ and $D^{k}_{\sigma} f = f^{(k)}$ for all $1 \le k \le n$. In particular, if $\sigma$ is saturated, then $(b)$ may be replaced by

  1. (b)

    For each $1 \le k \le n$, there exists $f^{(k)}: U \to B^{k}_{\sigma}(E; F)$ such that $D_{\sigma}^{k}(\fF) \to f^{(k)}$ uniformly on every $A \in \sigma$.

Proof. Assume without loss of generality that $n = 1$. For any $\varphi \in \tilde D^{1}_{\sigma}(U; F)$, $x \in U$, and $h \in E$ such that $x + h \in U$,

\begin{align*}f(x + h) - f(x) - f^{(1)}(x)h&= \underbrace{\varphi(x + h) - \varphi(x) - D_\sigma\varphi(x)h}_{\in \mathcal{R}_\sigma(E; F)}\\&+ (f - \varphi)(x + h) - (f - \varphi)(x) \\&+ (D_{\sigma}\varphi - f^{(1)})(x)h\end{align*}

Since $\fF \to f$ pointwise, for any $S \in \fF$,

\[(f - \varphi)(x + h) - (f - \varphi)(x) \in \overline{\bracs{(g - \varphi)(x + h) - (g - \varphi)(x)|g \in S}}\]

By the Mean Value Theorem, for any $g \in \tilde D^{1}_{\sigma}(U; F)$,

\[(g - \varphi)(x + h) - (g - \varphi)(x) \in \ol{\conv}\bracs{D_\sigma(g - \varphi)(x + th)h|t \in [0, 1]}\]

Hence

\[(f - \varphi)(x + h) - (f - \varphi)(x) \in \ol{\conv}\bracs{D_\sigma(g - \varphi)(x + th)h|g \in S, t \in [0, 1]}\]

so for any $t \in (0, 1)$ and $A \in \sigma$,

\begin{align*}&(f - \varphi)(x + tA) - (f - \varphi)(x) \\&\subset \ol{\conv}\bracs{D_\sigma(g - \varphi)(x + sth)th|g \in S, s \in [0, 1], h \in A}\\&= t\ol{\conv}\bracs{D_\sigma(g - \varphi)(x + sth)h|g \in S, s \in [0, 1], h \in A}\end{align*}

and

\begin{align*}&t^{-1}[(f - \varphi)(x + tA) - (f - \varphi)(x)] \\&\subset \ol{\conv}\bracs{D_\sigma(g - \varphi)(x + sth)h|g \in S, s \in [0, 1], h \in A}\\&\subset \ol{\conv}\bracs{D_\sigma(g - \varphi)(x + sh)h|g \in S, s \in [0, 1], h \in A}\end{align*}

In addition, since $D_{\sigma}(\fF) \to f^{(1)}$ pointwise,

\[t^{-1}(f^{(1)}- D_{\sigma}\varphi)(x)(tA) \subset \ol{\conv}\bracs{D_\sigma(g - \varphi)(x + sh)h|g \in S, s \in [0, 1], h \in A}\]

as well.

Now, let $V \in \cn_{F}(0)$ be convex and circled. Using assumption (b), let $S \in \fF$ such that for any $\varphi \in S$,

\[\ol{\conv}\bracs{D_\sigma(g - \varphi)(x + sh)h|g \in S, s \in [0, 1], h \in A}\subset V\]

Fix $\varphi \in S$, then as $\varphi$ is differentiable at $x$, there exists $\delta \in (0, 1)$ such that

\[t^{-1}[\varphi(x + tA) - \varphi(x) - D_{\sigma}\varphi(x)(tA)] \subset V\]

for all $t \in (0, \delta)$.

So

\[t^{-1}[f(x + tA) - f(x) - f^{(1)}(x)(tA)] \subset 3V\]

for all $t \in (0, \delta)$. Therefore $f$ is $\tilde \sigma$-differentiable at $x$ with $D_{\sigma} f(x) = f^{(1)}(x)$.$\square$

Post a Comment

Name:Email:
Please enter the tag of the current page (GW) to post the comment.
Tag:

Direct References

Jerry's Digital Garden

BibliographyComments

Direct References