As seen in this question, the derivatives can be easily characterized if we know $C^\infty(\mathbb{R}).$ How can we simply characterize $C^\infty(\mathbb{R})$ if we can't use limits?
-
7You can use the Fourier transform. Smooth functions are functions whose Fourier transform decays rapidly. – Stefan Lafon Nov 30 '22 at 01:21
-
I edited the question to disallow limits because now it seems like there are many ways to solve this using limits. – mathlander Nov 30 '22 at 01:24
-
1I’d suspect that something like this might work: Consider function spaces $V$ closed under multiplication and composition that allow a unique operator $D$ that satisfies product and chain rule and maps the identity to one and has the constants as kernel. I think $C^\infty$ is maximal with this property. At least I suspect so far that a too large space either as a too big kernel or no or non-unique $D$ … – Hagen von Eitzen Nov 30 '22 at 21:11
-
1It might be the closure of some (much simpler, "algebraic") space under some norm or at least topology. – Torsten Schoeneberg Dec 01 '22 at 03:22
1 Answers
Consider data $(V,D)$ with the following properties:
$$\tag1 V\text{ is a subalgebra of }\Bbb R^{\Bbb R}$$ $$\tag2 (\mathbf 1\colon x\mapsto1)\in V$$ $$\tag3 (\operatorname{id}\colon x\mapsto x)\in V$$ As a consequence, the algebra of polynomial functions is a subalgebra of $V$. $$\tag 4D\text{ is a linear map }V\to V$$ $$\tag 5D(f\cdot g)=D(f)\cdot g+D(g)\cdot f\quad\text{for }f,g\in V$$ $$\tag 6 D(\operatorname{id})=\mathbf 1$$ This introduces the derivative, but in a purely algebraic fashion. Below, we will not need the chain rule for $D$ (and in fact do not even demand that $V$ be closed under composition). I need a few more conditions to hold that can be viewed as not using limits and even avoid to mention continuity directly, but they introduce enough topological structure to reach the desired goal:
If $a$ and $b$ are real numbers with $a<b$, then with $I:=[a,b]$ $$\tag7f|_I \text{ is bounded}\quad\text{for }f\in V$$ $$\tag8\text{If }D(f)|_I\ge0\text{, then }f|_I\text{is non-decreasing}.$$
It may be possible to weaken these conditions even further, but at least they are good enough to prove the following
Claim. If $(V,D)$ with properties $(1)$ to $(8)$, then $V\subseteq C^\infty(\Bbb R)$ and $D=\frac{\mathrm d}{\mathrm dx}$. In particular, if $(V,D)$ is maximal, then $V=C^\infty(\Bbb R$).
Proof. Fix $x_0\in\Bbb R$ and $r>0$, and let $I=[x_0-r,x_0+r]$. By $(7)$, for every $f\in V$, there exists $M$ with $|D(f)(x)|\le M$ for all $x\in I$. Then $D(M\operatorname{id} \pm f)(x)=M\pm D(f)(x)\ge 0$ for all $x\in I$, hence the two functions $M\operatorname{id} \pm f$ are non-decreasing on $I$. Thus for $0\le h\le r$, $$M(x_0+h)\pm f(x_0+h)\ge Mx_0\pm f(x_0)\ge M(x_0-h)\pm f(x_0-h), $$ i.e., $$\tag9 |f(x)-f(x_0)|\le M|x-x_0|$$ for all $x\in I$. It follows that $f$ is continuous at $x_0$.
Let $g=f-D(f)(x_0)\operatorname{id}$. This makes $D(g)=D(f)-D(f)(x_0)\mathbf 1$ and in particular $D(g)(x_0)=0$. Apply $(9)$ to $D(g)$ to obtain (with a different constant $M'$) $$ |D(g)(x)|\le M'|x-x_0|$$ for $|x-x_0|\le r$. Hence when applying $(9)$ to $g$ with $0<r'\le r$, we can use $M=M'r'$. As long as $0<|x-x_0|\le r$, we may set $r'=|x-x_0|$ and arrive at $$g(x)-g(x_0)\le M'|x-x_0|^2. $$ This implies $g'(x_0)=0$, so $f'(x_0)=D(f)(x_0)$. As $x_0$ was arbitrary, ultimately $f'=D(f)$, as desired. $\square$

- 2,184

- 374,180
-
-
2I don't think you use the product rule anywhere in this argument. It seems that the conditions $(6)$, "$D(f) |_I$ is bounded" (weaker than $(7)$), and $(8)$ are all that is needed to force the linear operator $D$ to be the derivative and hence give the result. – jawheele Jan 06 '23 at 17:33
-
3This is very nice in that it indicates how natural the derivative really is: it effectively says that the derivative is the unique linear operator which can measure rate of change, as $(8)$ detects local monotonicity and $(6)$ normalizes the rate measurement. – jawheele Jan 06 '23 at 17:44