This is from Jaynes, Probability Theory: The Logic of Science, pp 27-28.

We have a function $F$ which is $\mathbb{R}^2 \rightarrow \mathbb{R}$, and we set $v = F(y,z)$. We discover that

$$ F_1(y,z) = { \partial F \over \partial y } = { H(v) \over H(y) } \\ F_2(y,z) = { \partial F \over \partial z } = r{ H(v) \over H(z) } $$

where $H$ is arbitrary, but can't change sign in the region of interest. (Specifically, $H$ is such that $F_2(y,z) / F_1(y,z)$ takes the form $r H(y) / H(z) $.)

We later discover $r = 1$, so I'm going to ignore that for clarity. Then, since $\mathrm d v = F_1 \mathrm d y + F_2 \mathrm d z $, we get

$$ { \mathrm d v \over H(v) } = { \mathrm d y \over H(y) } + { \mathrm d z \over H(z) } $$

So far, so good. Now we define

$$ w(x) = \exp\left( \int^x { \mathrm d x \over H(x) } \right) $$

and it follows that $w(v) = w(y) w(z)$.

I can kinda sorta see how this happens, but not really. Apparently,

$$ \int { \mathrm d v \over H(v) } = \int \left( { \mathrm d y \over H(y) } + { \mathrm d z \over H(z) } \right) = \int { \mathrm d y \over H(y) } + \int{ \mathrm d z \over H(z) } $$

And then we just apply $e^{a+b} = e^a e^b$. This makes some sense notationally, but I'm not familiar with the rigor behind it. (I've removed the ${}^x$ from the integrals because I don't really know what I'd do with it. Jaynes credits this proof to Cox (1961), which I looked up - it didn't include any intermediate steps to help me, but it also omitted the ${}^x$, so I feel somewhat comfortable doing so myself.)

I think there are a few specific things confusing me, where the notation just doesn't mean what I expect it to:

  • I expect expressions like $\int^x f(s) \mathrm d s$ to be a function of $x$, where $s$ is a bound variable, and we can rewrite $s$ as $p$ or $\alpha$ or anything, and rewriting $s$ as $x$ is just about the most confusing choice we can possibly make. In this case, it seems that the symbol used inside the integration is relevant?

  • $w$ looks like a function $\mathbb R \rightarrow \mathbb R $, but $w(v)$ depends upon $\mathrm d v$ as well as $v$?

Can anyone clear this up for me?

  • $\begingroup$ What is $H$? Some arbitrary function? A particular function? $\endgroup$ – Muphrid Mar 21 '14 at 14:37
  • $\begingroup$ Oh, sorry. It's arbitrary, but can't change sign in the region of interest. $\endgroup$ – philh Mar 21 '14 at 14:40

So you have a scalar field $F$ such that

$$\nabla F = [e_1/H(y) + e_2/H(z)] [H \circ F](y,z)$$

You can, of course, rearrange this to read

$$\frac{\nabla F|_{y,z}}{(H \circ F)(y,z)} = \frac{e_1}{H(y)} + \frac{e_2}{H(z)}$$

Now integrate on some curve $\ell(t) = \ell^y(t) e_1 + \ell^z(t) e_2$:

$$\int \left( \frac{e_1}{H \circ \ell^y} + \frac{e_2}{H \circ \ell^z} \right) \cdot \ell' \, dt = \int \frac{\nabla F|_\ell}{H \circ F \circ \ell} \cdot \ell' dt$$

Now, define

$$w(x;\lambda)=\exp \int^x \frac{\lambda'(t)}{(H \circ \lambda)(t)} \, dt$$

where $\lambda$ is some function. It should be clear that exponentiating the LHS just gives

$$\exp \left[\int^x \frac{(\ell^y)'}{H \circ \ell^y} \, dt + \int^x \frac{(\ell^z)'}{H \circ \ell^z} \, dt \right] = w(x; \ell^y)w(x; \ell^z)$$

The RHS is a little trickier. You need to realize that

$$\frac{d}{dt} (F \circ \ell)(t) = \ell' \cdot \nabla F|_{\ell(t)}$$

So the RHS can be written in terms of the function $G(t) = (F \circ \ell)(t)$ as

$$\int^x \frac{\ell'(t) \cdot \nabla F|_{\ell(t)}}{H \circ F \circ \ell} \, dt= \int^x \frac{G'(t)}{(H \circ G)(t)} \, dt$$

So exponentiating this, we get

$$w(x; F \circ \ell) = w(x; \ell^y) w(x; \ell^z)$$

Now, most people would call $y = \ell^y$ and $z = \ell^z$; this is a trivial choice of parameterization. Most people wouldn't bother writing $F \circ \ell$ and would just write $v$ as the authors did. So the letter of the proof holds: that $w(x;v) = w(x;y) w(x;z)$. I think writing it with one argument instead of two (as I did) is considerably less clear. Using the same variable for a bound of integration as the dummy variable is a big no-no, but some people do that when renaming would be less clear.

I think at the very least, $w$ ought to be a function of two arguments: a number $x$ and a function $\lambda: \mathbb R \mapsto \mathbb R$.

  • $\begingroup$ Thank you! I just worked through this and it makes sense. I proposed an edit fixing what I think were typos, please let me know if I'm wrong about them. Additional question, is it possible to turn w back into a function R -> R? Fixing the limits of integration would make w:(R->R)->R, but for example is there some function W:R->R such that W(F(y,z)) = w(x; F o l) for some fixed x and fixed curve l? Apologies if my question is unclear, or or the answer is obvious. $\endgroup$ – philh Mar 22 '14 at 22:02
  • $\begingroup$ I'm gonna have to think a bit on your question about $w$'s definition. I would say probably you can define $w$ the way the authors did and just not worry about how the path is parameterized. This would be a change of variables from $t$ to, for example, $\ell^y(t)$ and analagously elsewhere. What worries me most is what this does to the bounds. I'll have to think on that one. $\endgroup$ – Muphrid Mar 23 '14 at 1:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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