0
$\begingroup$

This is a question I've met recently while doing calculation. To make the point clear, let's just consider that a scalar field, say $\phi(x)$, as well as a functional, say $$B(\phi(x))=f(x)\phi(x).$$

Now we consider the following derivative: $$ \frac{\delta\phi(y)}{\delta B(\phi(x))} =\frac{\delta \phi(y)}{\delta (f(x)\phi(x))} =\frac{1}{\frac{\delta (f(x)\phi(x))}{\delta \phi(y)}} =\frac{1}{\frac{\delta f(x)}{\delta \phi(y)}\phi(x)+f(x)\delta^3(x-y)}.$$

Then, in this case, how could we understand this delta function in denominator? Or, eventually, if we put simply $$\frac{\delta \phi(x)}{\delta \phi(y)}=\frac{1}{\frac{\delta \phi(y)}{\delta\phi(x)}}=\frac{1}{\delta^3(x-y)},$$ where is the mistake in this issue? Since we know that $$\frac{\delta \phi(x)}{\delta \phi(y)}=\delta^3(x-y)$$ and delta function cannot be its inverse.

$\endgroup$
2

3 Answers 3

5
$\begingroup$

To expand on @mikestone's answer, the required result is not$$\frac{\delta\phi(x)}{\delta\phi(y)}\frac{\delta\phi(y)}{\delta\phi(x)}=1,$$but$$\int_{\Bbb R^3}\frac{\delta\phi(x)}{\delta\phi(y)}\frac{\delta\phi(y)}{\delta\phi(z)}d^3y=\frac{\delta\phi(x)}{\delta\phi(z)}.$$Indeed, this is just$$\int_{\Bbb R^3}\delta^{(3)}(x-y)\delta^{(3)}(y-z)d^3y=\delta^{(3)}(x-z).$$This is the functional equivalent of the multivariable chain rule$$\sum_j\frac{\partial x^i}{\partial y^j}\frac{\partial y^j}{\partial x^k}=\frac{\partial x^i}{\partial x^k}=\delta^i_k$$relating rival coordinate systems $x^i,\,y^j$; in particular, the equivalent of the summation operator $\sum_j$ is the integration operator $\int_{\Bbb R^3}d^3y$.

$\endgroup$
2
  • $\begingroup$ Thanks for your kind illustragtion! If I'm true, in this case I think then we have $\int \frac{\delta \phi(x)}{\delta \phi(y)}\frac{\delta \phi(y)}{\delta \phi(x)}d^3y=\delta^3(x-x)=\delta^3(0)$, which means, as you mentioned, we don't have $\frac{\delta \phi(x)}{\delta \phi(y)}=1/[\frac{\delta \phi(y)}{\delta \phi(x)}]$. However, may I ask that, consider there is a $\psi=\psi(\phi)$, then how would you calculate $\frac{\delta \phi(x)}{\delta \psi(y)}$? $\endgroup$
    – Chunhui
    Commented Jul 28, 2021 at 1:36
  • $\begingroup$ Sorry, I mean in the case that one cannot solve $\psi=\psi(\phi)$ and get $\phi=\phi(\psi)$ $\endgroup$
    – Chunhui
    Commented Jul 28, 2021 at 1:51
3
$\begingroup$

Functional derivatives can be understood in the sense of Gateaux as follows. Let $\phi\in C^\infty(\mathbb R^n)$ be a smooth function and $S:C^\infty(\mathbb R^n) \rightarrow \mathbb R$ be a functional. For a given $\eta\in C^\infty(\mathbb R^n)$, we define $$\frac{d}{d\epsilon} S[\phi+\epsilon \cdot \eta] \bigg|_{\epsilon=0}$$ to be the directional derivative of $S$ along $\eta$ at $\phi$. If this can be written in the form

$$\int \mathrm d^nx E\big[x,\phi(x), \partial \phi(x),\ldots\big] \eta(x)$$

for some function $E$, then we write $E\big[\phi(x),\partial \phi(x),\ldots\big]=: \frac{\delta S}{\delta \phi(x)}$ and call it the functional derivative of $S$ with respect to $\phi(x)$ (or somewhat more accurately, the functional derivative of $S$ evaluated at $\phi(x)$).


Example

Having made that preliminary defininition, consider the functional $S$ which eats a smooth function and integrates its square over $\mathbb R^n$. Then

$$\frac{d}{d\epsilon} S[\phi + \epsilon \cdot \eta] \bigg|_{\epsilon=0}= \frac{d}{d\epsilon} \int \mathrm d^n x \big(\phi(x)+\epsilon \cdot \eta(x)\big)^2\bigg|_{\epsilon=0} = \int \mathrm d^n x \ 2\phi(x) \eta(x)$$

$$\implies \frac{\delta S}{\delta \phi(x)} = 2\phi(x)$$


Next, let $T:C^\infty(\mathbb R^n)\rightarrow C^\infty(\mathbb R^n)$ be an operator. Once again we consider a Gauteaux-derivative which will now be function-valued rather than $\mathbb R$-valued:

$$\frac{d}{d\epsilon} T[\phi+ \epsilon \eta;x]\bigg|_{\epsilon=0}$$ where the notation $T[\ldots;x]$ is used to remind us that after plugging in a function $\phi$ to the first slot of $T$, the result is a (possibly generalized) function of the dummy variable $x$. If this can be expressed as

$$\frac{d}{d\epsilon} T[\phi+ \epsilon \eta;x]\bigg|_{\epsilon=0} = \int \mathrm d^n y F[x,y,\phi(y),\ldots] \eta(y)$$

for some $F$, then as before we call $F[\ldots]=: \frac{\delta T[\phi;x]}{\delta \phi(y)}$ the derivative of $T$ with respect to (or evaluated at) $\phi(y)$.


Example

Let $T$ be the operator which eats a function and spits that function right back out, i.e. $T[\phi;x] = \phi(x)$. Then

$$\frac{d}{d\epsilon}T[\phi+\epsilon\cdot \eta;x]\bigg|_{\epsilon=0} = \frac{d}{d\epsilon}\big(\phi(x)+\epsilon\cdot \eta(x)\big)\bigg|_{\epsilon=0} = \eta(x) = \int \mathrm d^4n y \delta(x-y) \eta(y)$$ $$\implies \frac{\delta T[\phi;x]}{\delta \phi(y)} = \delta(x-y)$$

As a mild abuse of notation, this is often written $\frac{\delta \phi(x)}{\delta \phi(y)}= \delta(x-y)$. Note that in elementary calculus we usually write the derivative of the function $\mathrm{sq}:x\mapsto x^2$ as $\frac{d(x^2)}{dx} = 2x$ rather than $\frac{d(\mathrm{sq})}{dx} = 2x$; this is precisely the same abuse of notation as employed here.


Finally, consider some operator $T:C^\infty(\mathbb R^n)\rightarrow \mathbb C^\infty(\mathbb R^n)$ and some function $\phi$ such that in a neighborhood of $\phi$, $T$ can be inverted, i.e. $T^{-1}\big[T[\phi;\cdot];x\big] = \phi(x)$. After a bit of work, one can show that if

$$\frac{d}{d\epsilon} T[\phi+ \epsilon \eta;x]\bigg|_{\epsilon=0} = \int \mathrm d^n y F\big[x,y,\phi(y),\ldots\big] \eta(y)$$ and (letting $\rho(x) = T[\phi;x]$ for brevity) $$\frac{d}{d\epsilon} T^{-1}\big[\rho+ \epsilon \eta;x\big]\bigg|_{\epsilon=0} = \int \mathrm d^n y G\big[x,y,\rho(y),\ldots\big] \eta(y)$$

then we must have that

$$\int \mathrm d^n z G\big(x,z,T[\phi;z],\ldots\big) F\big(z,y,\phi(y),\ldots\big) = \delta(x-y)$$

or, somewhat less unpleasantly, $$\int \mathrm d^n z \frac{\delta T^{-1}\big[T[\phi;\cdot],x\big]}{\delta T[\phi;z]} \frac{\partial T[\phi;z]}{\delta \phi(y)} = \delta(x-y)$$

This is essentially the functional(/operator) derivative generalization of the inverse function theorem, and it is in this sense that these derivatives are inverses of one another - not simple multiplicative inverses, but rather integral inverses in the sense above.

It's illuminating to note that if we consider functionals and operators $f:\mathbb R^n\rightarrow \mathbb R^n$ rather than on $C^\infty(\mathbb R^n)$, then the functional derivatives reduce to the corresponding matrices $\frac{\partial f^i}{\partial x^j}$ and the inverse function theorem reduces to a statement about matrix multiplication; in that sense, the derivatives of $f$ and $f^{-1}$ are matrix inverses of one another, not simply multiplicative inverses. Of course, if we continue to simplify and specialize to $n=1$, then we arrive at the inverse function theorem from Calculus 101 in which simple multiplicative inverses finally emerge.


Example

As a trivial example, let $T[\phi;x]=\phi(x)$. Then $T^{-1}[\rho;x] = \rho(x)$ (i.e. $T$ is its own inverse). Both of them have the same derivative as given in the previous example, and indeed

$$\int \mathrm d^n z \delta(x-z) \delta(z-y) = \delta(x-y)$$

Example 2

As a less trivial example, let $T[\phi;x] = f(x)\phi(x)$ as you propose. Then $T^{-1}[\rho;x] = \frac{1}{f(x)}\rho(x)$ (assuming that $f(x)\neq 0$). The derivatives of these operators are simply

$$\frac{\delta T[\phi;x]}{\delta \phi(y)} = f(x)\delta(x-y)\qquad \frac{\delta T^{-1}[\rho;x]}{\delta \rho(y)} = \frac{1}{f(x)} \delta(x-y)$$

and once again,

$$\int \mathrm d^n z \frac{1}{f(x)} \delta(x-z) f(y) \delta(z-y) = \delta(x-y)$$

Example 3

Just for fun, let $T[\phi;x] = \frac{1}{(2\pi)^{n/2}}\int \mathrm d^n k \ e^{-ikx} \phi(k)$ and $T^{-1}[\rho;x] = \frac{1}{(2\pi)^{n/2}}\int \mathrm d^n k\ e^{ikx} \rho(k)$. Then

$$\frac{\delta T[\phi;x]}{\delta \phi(y)} = e^{-ixy}/(2\pi)^{n/2} \qquad \frac{\delta T^{-1}[\rho;x]}{\delta \rho(y)} = e^{ixy}/(2\pi)^{n/2}$$ $$\implies \int \mathrm d^n z \frac{e^{-i(x-y)z}}{(2\pi)^n} = \delta(x-y)$$

$\endgroup$
2
$\begingroup$

You have to invert the matrix with of partial derivatives with entries labled by $x$, $y$. You do not just invert the entries: if matrix $1+M$ has entries $\delta_{ij}+m_{ij}$ then $(1+M)^{-1}$ does not have entries $1/(\delta_{ij}+m_{ij})$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.