3
$\begingroup$

I'm interested in numerical algorithms for 1-dimensional Hamiltonians of the form

$$ H = -\frac{d^2}{dx^2} + V(x) \quad \quad (1) $$

defined on the line ($x\in\mathbb{R}$) or on the circle. The potential $V$ is such that the lowest part of the spectrum is made of eigenvalues.

My definition of best (and of 'integrating') is that the algorithm should return more accurately the lowest eigenvalues and corresponding eigenfunctions (say the lowest $N$ with $N$ of order 10-20).

I'm aware of Numerov's method but I think there should be better algorithms given my desiderata.

In particular I wonder wether Krylov methods exist for my problem. The potential $V$ is given explicitly so computing $H \psi$ for a nice smooth $\psi$ is easy and in general obtaining $\langle \phi, H^n \psi \rangle$ seems possible to a certain degree of accuracy.

Bonus points if the algorithm is implemented in some (available) software package.

$\endgroup$
  • $\begingroup$ See this physics.stackexchange.com question. The answer contain cut-and-pasteable code. $\endgroup$ – Igor Rivin Dec 19 '17 at 3:07
  • $\begingroup$ @IgorRivin Thank you but that post seems to deal with Numerov's method. I was hoping for something a bit more advanced (without reinventing the wheel) $\endgroup$ – lcv Dec 19 '17 at 3:36
1
$\begingroup$

To obtain the lowest eigenvalues by means of a Krylov subspace method you could use the Lanczos algorithm, as explained for example in:

Solving the discretized time‐independent Schrödinger equation with the Lanczos procedure

A new method is presented to find bound state solutions of the one‐, two‐, or three‐dimensional Schrödinger equation. The equation is turned into a sparse matrix eigenvalue problem by representing the potential energy surface and the wave function on a grid. The Laplacian is represented by a high (10th) order finite difference formula. Eigenvalues are found by the Lanczos procedure. Examples are given for the 1D Morse oscillator and the 2D Hénon‐Heiles potential. Numerical convergence can be checked easily and highly accurate results can be obtained. The algorithm is fast, easy to implement, and vectorizable.

$\endgroup$
  • $\begingroup$ Thank you this may work. Let me see if something else comes up. If not I will accept this. I was hoping to avoid to discretize space. $\endgroup$ – lcv Dec 19 '17 at 10:05

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.