2

Let $\mathbb{F}$ be a field, and consider $\mathbb{F}^\mathbb{F}$ as an algebra over $\mathbb{F}$ with the standard function multiplication. Let $D$ be a linear transformation on a subalgebra of $\mathbb{F}^\mathbb{F}$ closed under function composition that satisfies the chain rule. Does $D$ necessarily satisfy the product rule for arbitrary $\mathbb{F}$? (Inspired by a comment on this question.) What if the subalgebra must be unital?

mathlander
  • 3,997

2 Answers2

1

Let $\mathbb{F}=\mathbb{F}_2$ and consider $D:\mathbb{F}_2^{\mathbb{F}_2}\to\mathbb{F}_2^{\mathbb{F}_2}$ which sends the constant functions to $0$ and the nonconstant functions to $1$. It is easy to check that this satisfies the chain rule but it does not satisfy the product rule since $D(i\cdot i)=D(i)=1\neq 2iD(i)=0$ where $i$ denotes the identity function.

On the other hand, if $\mathbb{F}$ has characteristic different from $2$, then the chain rule actually does imply the product rule (assuming your subalgebra $A\subseteq\mathbb{F}^\mathbb{F}$ is unital and contains the identity function $i$). First, note that $$D(1)=D(1\circ 0)=(D(1)\circ 0)\cdot D(0)=0$$ since $D(0)=0$. Also, for any $f\in A$, $$D(f)=D(f\circ i)=D(f)D(i).$$ Taking $f=i$ gives $D(i)=D(i)^2$, so $D(i)$ can only take the values $0$ and $1$. Let $S\subseteq\mathbb{F}$ be the set of inputs on which $D(i)$ is $0$; then $D(f)$ vanishes on $S$ for all $f\in A$. For each $a\in\mathbb{F}$, let $s_a$ be the function $a-i$. Note that $s_a\circ s_a=i$ so $$D(i)=(D(s_a)\circ s_a)\cdot D(s_a).$$ Comparing the vanishing sets of each side of this equation, we see that $S\supseteq s_a^{-1}(S)=s_a(S)$. Since $a\in\mathbb{F}$ is arbitrary, this means $S$ must be either $\emptyset$ or all of $\mathbb{F}$. If $S=\mathbb{F}$ then $D(i)=0$ and thus $D=0$ and the conclusion is trivial. So let us assume $S=\emptyset$, which means $D(i)=1$.

Now let $f=D(i^2)$. For any $a,b\in\mathbb{F}$, we have $$D((ai+b)^2)=(f\circ (ai+b))D(ai+b)=af\circ(ai+b)$$ but also $$D((ai+b)^2)=D(a^2i^2+2abi+b^2)=a^2f+2ab.$$ That is, for each $x\in\mathbb{F}$, $$af(ax+b)=a^2f(x)+2ab,$$ or $$f(ax+b)=af(x)+2b$$ as long as $a\neq 0$. Plugging in $x=0$ gives $$f(b)=af(0)+2b.$$ Since $\mathbb{F}$ has more than $2$ elements (so there are multiple different choices for $a$) this implies $f(0)=0$ and thus $f(b)=2b$. That is, $f=2i$.

Now let $f,g,\in A$ be arbitrary and consider $D((f+g)^2)$. On one hand, $$D((f+g)^2)=(2i\circ (f+g))\cdot D(f+g)=2(f+g)D(f+g).$$ On the other hand, $$D((f+g)^2)=D(f^2+2fg+g^2)=2fD(f)+2D(fg)+2gD(g).$$ Since $2\neq 0$, comparing these two equations gives $D(fg)=fD(g)+gD(f)$, as desired.

(I don't know what can be said if $\mathbb{F}$ has characteristic $2$ but more than $2$ elements.)

Eric Wofsey
  • 330,363
-1

Not an answer, but a helpful heuristic.

The chain rule and the product rule cannot be compared. A product of functions, which you called standard, requires two functions with the same domain and range: $f,g\, : \,D\rightarrow R$ since we plug in the same value and multiply in a set where multiplication has to be defined: $$(f\cdot g)(x)=f(x)\cdot g(x).$$ On the other hand, a composition of functions requires that the range from one function is within the domain of another: $f\, : \,R\rightarrow S$ and $g\, : \,D \rightarrow R$ in order to make $$ (f\circ g)(x)=f(g(x)) $$
possible. If you simplify all these requirements by setting e.g. $D=R=S=\mathbb{R}$ then you disguise those subtleties. They are still there. Hence product and composition are two very different operations. It is even more obvious if we consider the derivatives: \begin{align*} D_p(f\circ g)&=D_{g(p)}(f)\cdot D_p(g)\\ D_p(f\cdot g)&=D_p(f)\cdot g+ f\cdot D_p(g) \end{align*} The fact that the evaluation points differ significantly makes them incomparable.

mathlander
  • 3,997
Marius S.L.
  • 2,245
  • 2
    We're specifically considering $\Bbb F^{\Bbb F}$, or a subalgebra thereof that is closed under composition, because this is the context where it makes sense to ask the question. It's not clear what exactly you mean by "incomparable". – Karl Dec 20 '22 at 19:35
  • I mean that the derivative of a composition takes place in different tangent spaces, whereas the derivative of a product takes place in a single tangent space. That is qualitatively different, whether all those spaces are identical or not. – Marius S.L. Dec 20 '22 at 19:41
  • @mathlander How and why? – Marius S.L. Dec 20 '22 at 19:42
  • 1
    Okay, but what does "incomparable" mean in this context? – mathlander Dec 20 '22 at 19:51
  • That's how I read the question. The derivative of a composition is a product of derivatives, whereas the derivative of a product is a derivation. Is that abstract enough? – Marius S.L. Dec 20 '22 at 19:51
  • 1
    Yes, but I don't think that this actually answers my question. – mathlander Dec 20 '22 at 19:53
  • I thought I already answered it: "I mean that the derivative of a composition takes place in different tangent spaces, whereas the derivative of a product takes place in a single tangent space. That is qualitatively different, whether all those spaces are identical or not." – Marius S.L. Dec 20 '22 at 19:54
  • 1
    Now I understand why this is a (partial) answer. – mathlander Dec 20 '22 at 19:55
  • 2
    However, a "true" answer would either prove that the chain rule implies the product rule or provide a counterexample. – mathlander Dec 20 '22 at 19:56
  • Your question is whether one implies the other. I'd say no and I gave the reason for this assessment. How could I prove "it does not imply"? Even if everything is fine, how could $f\circ g$ be compared to $f\cdot g$? $D_p(f\cdot 1)=D_p(f)\cdot 1+f\cdot D_p1=D_p(f)$ and $D_p(f\circ 1)=D_{1(p)}1\cdot D_p(f)=D_p(1)\cdot D_p(f)=1\cdot D_p(f)$ since the derivative of the identity is the identity again. We get identical results in this particular case. However, this doesn't say anything about the general case. – Marius S.L. Dec 20 '22 at 20:05
  • I think I'm still missing something. If we think of $D_p$ (for a fixed $p$) as mapping a function to a scalar, then your expression $D_p(f\circ g)=D_{g(p)}(f)\cdot D_p(g)$ for the chain rule makes sense, but for the product rule, shouldn't $D_p(f\cdot g)=D_p(f)\cdot g+ f\cdot D_p(g)$ actually be $D_p(f\cdot g)=D_p(f)\cdot g(p)+ f(p)\cdot D_p(g)$? Then we're studying an operator $D$, from functions to functions, that takes $f$ to $(p\mapsto D_p(f))$, and the "types" seem to work out well enough to ask the original question. – Karl Dec 20 '22 at 21:10
  • (Of course this relies on the assumption that the space of function values is the same as the space of derivatives-at-a-point.) – Karl Dec 20 '22 at 21:20
  • 1
    My argument is that there is a qualitative difference since the spaces "can" be different. The fact that they are the same does not change this. $f\circ g$ involves three different spaces, $f\cdot g$ only two. If all of them are simply $\mathbb{F},$ then calculations are easier. Geometrically, we still have three spaces for composition, two for multiplication. E.g. you can compare the unit circle in the complex plane for calculus with rotations of the real plane for physics. They are the same, yet qualitatively different in their roles. – Marius S.L. Dec 20 '22 at 21:36