Consider the piecewise-deterministic Markov process on $\mathbf{R}$ which

  • moves according to the vector field $\phi (x) = 1$,
  • experiences events at rate $\lambda(x) = 1$, and
  • at events, jumps according to the transition kernel

\begin{align} Q (x \to dy) = \frac{\exp(- y - \exp(-y))}{\exp( - \exp (-x))} \, \mathbf{I} [y \leqslant x] dy \end{align}

i.e. the location is resampled according to a Gumbel distribution, conditioned to be to the left of $x$.

It can be shown that this process admits the Gumbel distribution as its stationary measure, that is, the measure with density $G(x) = \exp(- x - \exp(-x))$. Moreover, its generator is given by

\begin{align} \mathcal{L} f (x) = f'(x) + \int_{-\infty}^x [f(y) - f(x)] Q (x \to dy). \end{align}

It is worth remarking that the process is non-reversible, and hence its generator is not self-adjoint.

For reference, the generator's adjoint in $L^2 (dG)$ is given by

\begin{align} \mathcal{L}^* f (x) = - f'(x) + \exp (-x) \int_x^{\infty} [f(y) - f(x)] Q^* (x \to dy) \end{align}


\begin{align} Q^* (x \to dy) = \exp(- (y - x)) \mathbf{I} [y \geqslant x] dy, \end{align}

which can be derived by computing the time-reversal of the original process.

Now, I'm interested in studying the eigenfunctions of the generator, i.e. solutions to

\begin{align} \mathcal{L} f (x) = \lambda f (x) \end{align}

After some calculations, I can simplify this equation to

\begin{align} f''(x) + (\exp ( -x) - 1 ) f'(x) = \lambda \left\{ \exp (-x) f(x) + f' (x) \right\} \end{align}

which can be put into `quasi-Sturm-Liouville' form as

\begin{align} \frac{d}{dx} \left( G(x) f'(x) \right) &= \lambda G(x) \left\{ f' (x) + \exp (-x) f(x)\right\}, \end{align}

and even into the form

\begin{align} \frac{d}{dx} \left( G(x) f'(x) \right) &= \lambda \exp (-x) \frac{d}{dx} \left( \exp ( - \exp (-x)) f(x) \right), \end{align}

which could conceivably be useful.

Anyways, although this all looks very reminiscent of Sturm-Liouville theory, it's not quite there, in as much as the term with the eigenvalue $\lambda$ don't just involve $f$, but also its derivatives.

My main questions are thus as follows:

  1. Is there a name for this type of system? i.e. S-L-type structure, but where the $\lambda$-term depends on the derivatives of $f$ as well. If so, I'd appreciate relevant references.

  2. Are there any other techniques which might be of use in trying to solve this eigensystem? Ideally there will be solutions of the form

\begin{align} f(x; \lambda_n) = \{ \text{polynomial of degree } n \} \times \text{fixed function}, \end{align}

but this might be a bit (/very) optimistic on my part.

A related consideration (which I'm in the process of doing the calculations for) is to carry out a spectral decomposition of $\mathcal{L} \mathcal{L}^*$ and $\mathcal{L}^* \mathcal{L}$. Given that these will be self-adjoint, some of the theory might make life a little easier; on the other hand, both operators involve double integrals, so it may be time-consuming. I'll update with details once I've had a go of that. Any advice on that would also certainly be welcome.


Your Answer

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

Browse other questions tagged or ask your own question.