10
$\begingroup$

I have a question about obtaining matrix elements of time evolution operator. I have the following Hamiltonian for a particle in a ring with magnetic field

$$H=\dfrac {\hbar ^{2}} {2mR^{2}}\left[ -i\dfrac {\partial } {\partial \theta}+\dfrac {\phi e} {h}\right] ^{2}$$

and since $\left[ H,P\right] =0$, for $P=-i\hbar\frac{\partial}{\partial\theta}$, I took momentum eigenstates as eigenstates of the Hamiltonian.

Then I expressed the time evolution operator $U\left( t_{2},t_{1}\right) =e^{\frac {-i} {\hbar }H\left( t_{2}-t_{1}\right) }$ in the eigenbasis of the Hamiltonian in the following way

$$-\frac {i} {\hbar }H\left( t_{2}-t_{1}\right) \lvert n\rangle =-\frac {i} {\hbar }\left( t_{2}-t_{1}\right) E_{n}\lvert n\rangle$$

$$\frac {-i} {\hbar }H\left( t_{2}-t_{1}\right) \sum_{n=-\infty}^\infty \lvert n\rangle \langle n\rvert =\frac {-i} {\hbar }\left( t_{2}-t_{1}\right) \sum_{n=-\infty}^\infty E_{n}\lvert n\rangle\langle n\rvert $$

And with the completeness property $\sum_{n=-\infty}^\infty \lvert n\rangle\langle n\rvert=1$ we get: $$U\left( t_{2},t_{1}\right) =e^{\frac {-i} {\hbar }H\left( t_{2}-t_{1}\right) }=\sum_{n=-\infty}^\infty e^{\frac {-i} {\hbar }\left( t_{2}-t_{1}\right) E_{n}}\lvert n\rangle\langle n\rvert $$

Now, I know that I can sandwich the operator between eigenstates to find matrix elements. But I want to find the matrix element in the $\theta$ basis. More specifically I want to find $\left\langle \theta_{2}\left| U\left( t_{2},t_{1}\right) \right| \theta_{1}\right\rangle$. The eigenstates of the Hamiltonian in the $\theta$ basis,$\left\langle \theta|n\right\rangle =\frac{1}{\sqrt{2\pi}}e^{in\theta}$

$\endgroup$
22
+50
$\begingroup$

You wouldn't think it, from how easy it is to pose this question, but it is ridiculously nontrivial. As it happens, it is entirely impossible to find the position-basis matrix elements of this propagator.

So far you've done good, and the identification $$ U\left( t_{2},t_{1}\right) =e^{\frac {-i} {\hbar }H\left( t_{2}-t_{1}\right) } =\sum_{n=-\infty}^\infty e^{-i\left( t_{2}-t_{1}\right) E_{n}/\hbar}\lvert n\rangle\langle n\rvert $$ is correct. Naively, the path is clear, because you know the eigenfunctions, $$ ⟨\theta|n⟩=\frac{1}{\sqrt{2\pi}}e^{in\theta}, $$ so you can in principle just sandwich $U(t_2,t_1)$ between two position kets and you'd be done. This gives you \begin{align} \langle \theta_2 | U\left( t_{2},t_{1}\right) | \theta_1\rangle & = \sum_{n=-\infty}^\infty e^{-i\left( t_{2}-t_{1}\right) E_{n}/\hbar}\langle \theta_2 | n\rangle\langle n|\theta_1\rangle \\ & = \frac{1}{2\pi} \sum_{n=-\infty}^\infty e^{-i\left( t_{2}-t_{1}\right) E_{n}/\hbar}e^{in(\theta_1-\theta_2)}, \end{align} and everything looks fine, right? You've just got one series to calculate and you'll be on your way. Moreover, you can simplify things here a fair bit: you can set $\theta=\theta_1-\theta_2$ and $\tau=t_2-t_1$ to simplify notation (i.e. the propagator only depends on those differences, which is nice), and you can also calculate the energies, $$ E_n=\frac{\hbar^2}{2mR^2}\left(n+\frac{e}{h}\phi\right)^2, $$ so that $ \frac{(t_2-t_1)E_n}{\hbar} = \frac{\hbar}{2mR^2}\left(n+\frac{e}{h}\phi\right)^2\tau =\kappa n^2\tau + \lambda n\phi\tau + \mu\phi^2\tau, $ and all of this makes for a nicely simplified series: \begin{align} \langle \theta_2 | U\left( t_{2},t_{1}\right) | \theta_1\rangle & = \frac{e^{-i\mu\phi^2\tau}}{2\pi} \sum_{n=-\infty}^\infty e^{-in^2\kappa \tau }e^{in(\theta-\lambda\phi\tau)}, \end{align} and this is clean and simple as can be. What could possibly go wrong?


Now, as it happens, this series is exactly summable, and it represents something called a Jacobi $\vartheta_3$ function, and you can read all about it in the DLMF . More precisely, in this case, the relationship comes down to \begin{align} \langle \theta_2 | U\left( t_{2},t_{1}\right) | \theta_1\rangle & = \frac{1}{\sqrt{4\pi i \kappa\tau}} e^{-i\mu\phi^2\tau} e^{\frac{i (\lambda\phi\tau-\theta)^2}{4 \kappa\tau}} \vartheta_3\mathopen{}\left( \frac{\pi }{2 \kappa\tau}(\lambda\phi\tau-\theta), e^{i\frac{ \pi^2}{\kappa\tau}} \right)\mathclose{} , \tag{$\ast$} \end{align} and it kinda looks like you're done: you've reduced your series to some special function, and you obviously cannot whittle that down any further, so there's nothing else to do.

Unfortunately, however, you're not actually done yet. If you look more closely at the theta function, it is normally defined as $$ \vartheta_3(z,q)=1+2\sum_{n=0}^\infty q^{n^2}\cos(2nz), $$ where $q$ and $z$ are complex parameters, and we will want to set $q=e^{i \pi^2/\kappa\tau}$. It is obvious from the definition that the series will converge if $|q|<1$, and that it will diverge if $|q|>1$. The quantum mechanics result we just got is balanced on the knife edge, at $|q|=1$: there are no convergence guarantees, but the series is oscillatory, so maybe we can hope for some conditional convergence brought about by the oscillations?

Alas, there's no joy there either. The full answer is that $\vartheta_3$ has something called a natural boundary, which means that for any point $q_0$ on the unit circle (so $|q_0|=1$) and any $z\in\mathbb C$, the limit $\lim_{q\to q_0^-} \vartheta_3(z,q)$ diverges to infinity. (I'm using the notation $q\to q_0^-$ to imply that $q$ stays inside the disk during the limit, just to keep things easy.) If you want the big ugly mathematician details, this is because the series is something called a lacunary series, which I asked about here on MathOverflow and here on Maths SE; in particular, here the Fabry gap theorem implies that the theta function has a ring of fire around it, which sort of looks like this:

Mathematica code through Import["https://raw.githubusercontent.com/halirutan/Mathematica-SE-Tools/master/SETools/SEImageExpressionDecode.m"]["http://i.stack.imgur.com/wM9dN.png"]

The problem here is that our $\vartheta_3$ is sitting right on top of the natural boundary: it is an analytical function for $|q|<1$, but there is provably no way to provide an analytical continuation past its circle of convergence (and that is kind of a bummer), and there is provably no way to make any sense of it for $|q|=1$, which is exactly where we are.


So... where does that leave us? Our nice propagator $(\ast)$ looked neat enough, but a closer inspection shows that it is actually mathematical nonsense. What does this mean, and how do we fix it?

Well, there's a couple of ways, but they both essentially mean that you are only allowed to interpret $\langle \theta_2 | U\left( t_{2},t_{1}\right) | \theta_2+\theta \rangle$ as something called a generalized function, also known as a distribution: an object, kind of like the Dirac delta function, that doesn't make sense on its own, but only when integrated in products with other functions.

There's two easy ways to do this:

  • One is to take the complex-$q$ thing seriously. This is easily achieved by giving the time $\tau=\tau_0-i\delta$ a nonzero imaginary component, which brings $|q|$ down to under $1$, and lets $⟨ \theta_2 | U\left( t_{2},t_{1}\right) | \theta_1 ⟩$ make sense.

    If you do this, the key thing to note is that normally the only physically meaningful quantities where you will use the propagator are matrix elements of the form $⟨f|U(t_2, t_1)|g⟩$ for physical states $|f⟩$ and $|g⟩$. Those ultimately boil down to integrals of the form \begin{align} ⟨f|U(t_2, t_1)|g⟩ & = \int_0^{2\pi} ⟨0|U(t_2, t_1)|\theta⟩ F(\theta)\mathrm d\theta \\&= \int_0^{2\pi} \frac{ e^{-i\mu\phi^2\tau} e^{\frac{i (\lambda\phi\tau-\theta)^2}{4 \kappa\tau}} }{\sqrt{4\pi i \kappa\tau}} \vartheta_3\mathopen{}\left( \frac{\pi }{2 \kappa\tau}(\lambda\phi\tau-\theta), e^{i\frac{ \pi^2}{\kappa\tau}} \right)\mathclose{} F(\theta)\mathrm d\theta , \end{align} and here the recipe says that you should read those integrals as the limit of integrals with a nonzero $\delta$ as that $\delta$ goes to zero: \begin{align} ⟨f|U(t_2, t_1)|g⟩ & = \lim_{\delta\to 0} \int_0^{2\pi} \frac{ e^{-i\mu\phi^2\tau} e^{\frac{i (\lambda\phi\tau-\theta)^2}{4 \kappa\tau}} }{\sqrt{4\pi i \kappa\tau}} \vartheta_3\mathopen{}\left( \frac{\pi }{2 \kappa\tau}(\lambda\phi\tau-\theta), e^{i\pi^2/\kappa(\tau-i\delta)} \right)\mathclose{} F(\theta)\mathrm d\theta . \end{align} If you do that, what happens is that for each $\delta$ the integrand is well defined, analytical and perfectly well behaved, and it gives you a perfectly normal result for the integral. More importantly, those results will approach a limit as $\delta\to0$ as long as $F(\theta)$ is not absolutely crazy.

  • The other approach is similar, but it's series-based, i.e. it requires you to take our original series before the evaluation to $\vartheta_3$, \begin{align} \langle \theta_2 | U\left( t_{2},t_{1}\right) | \theta_1\rangle & = \frac{e^{-i\mu\phi^2\tau}}{2\pi} \sum_{n=-\infty}^\infty e^{-in^2\kappa \tau }e^{in(\theta-\lambda\phi\tau)}, \end{align} and see it as indicating that you need to take the $|n|\to\infty$ limit after you've done any relevant inner products you need to take. For the integral above, for example, this means the understanding that \begin{align} \int_0^{2\pi} ⟨0|U(t_2, t_1)|\theta⟩ F(\theta)\mathrm d\theta &= \frac{e^{-i\mu\phi^2\tau}}{2\pi} \sum_{n=-\infty}^\infty e^{-in^2\kappa \tau } \int_0^{2\pi} e^{in(\theta-\lambda\phi\tau)} F(\theta)\mathrm d\theta , \end{align} with the series after the integral. If $F(\theta)$ is well-behaved, then its Fourier series coefficients will go down with $|n|$ reasonably fast, and the series will (whew!) converge again - as it needs to do for a physically relevant matrix element.

  • Alternatively, you can take the series expression we had earlier, \begin{align} \langle \theta_2 | U\left( t_{1}+\tau,t_{1}\right) | \theta_2+\theta\rangle & = \frac{e^{-i\mu\phi^2\tau}}{2\pi} \sum_{n=-\infty}^\infty e^{-in^2\kappa \tau }e^{in(\theta-\lambda\phi\tau)}, \end{align} and simply see it as an identity between distributions.

    This paradigm requires you to see each mapping $\theta\mapsto e^{-in^2\kappa \tau }e^{in(\theta-\lambda\phi\tau)}$ not as a function but as a distribution, which you are then adding up together as a series. This is a pretty strong thing to do, since it requires you to stop seeing $⟨\theta_2|n⟩⟨n|\theta_1⟩$ as something which has a value if you give it a specific $\theta_1$, $\theta_2$ and $n$, but if you do this you get some nice mathematical advantages.

    More specifically, each of the distributions $\theta\stackrel{\mathrm{DS}}{\mapsto} e^{-in^2\kappa \tau }e^{in(\theta-\lambda\phi\tau)}$ is a tempered distribution (i.e. the kind of distribution that plays well with Fourier transforms and quantum mechanics; see these notes, these ones or these ones for more details), and the series is bounded above termwise by the Dirac comb, which is also a tempered distribution, and this is (probably?) plenty to show that the series sums to a tempered distribution.

    Thus, in this paradigm, you get that the propagator exists and it is as nice as you'd like it to be (within the bounds that it is a distribution) and that it is indeed the limit of the series on the right hand side, so long as you take it in the distributional sense.


That's sort of it, really, and I think this is as clear as I can make it. I'll close with the two references where I learned this stuff,

L. Schulman. A path integral for spin. Phys. Rev. 176 no. 5, p. 1558. (1968).

and

L. S. Schulman. Techniques and applications of path integration (Dover, New York, 2005), pp. 190-196.

and finally with a quote from Schulman (actually present in both references), which describes the situation here rather accurately:

The degree of pathology exhibited by this Green's function is entertaining, especially in view of the elementary nature of the example.

$\endgroup$
  • $\begingroup$ Have you since stumbled upon some physical insight as to why the math turns out to be so pathological? Is the example somehow too artificial? In simple enough cases, I feel that sometimes divergences come out of idealizations (such as ignoring backreaction on the magnetic field?) but here it seems that assigning dynamics to the magnetic field would summon the fifty shades of infinity from the QED :/ :( $\endgroup$ – Feynmans Out for Grumpy Cat May 11 at 10:54
  • $\begingroup$ No, I haven't, I'm sorry to say. I've shown this a couple of times in a lecture context, with some smart people attending, and no deeper insight bubbled up. Maybe it's worth having a separate question asking why this behaviour happens - it might prompt the Qmechanics and V Morettis of the world to examine it more closely and potentially come across a good explanation. $\endgroup$ – Emilio Pisanty May 11 at 12:22

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.