13
$\begingroup$

I am trying to speedup the calculation of eigenvalues, given that I have good guesses for the eigenvectors. From what I know of Arnoldi/Lanczos, my good guesses should be helpful. Unfortunately, I am unable to see any speedup when I give these guesses to Mathematica.

Here is a simple and very artificial example.

I first create a random vector:

testvec = SparseArray[Table[RandomInteger[2000] -> Random[], {500}],{2000}]

I then create a matrix for which this is an eigenvector:

testmat = KroneckerProduct[testvec, testvec]

By construction the eigenvalue is

ev=(testvec.testmat.testvec)/(testvec.testvec)

On my computer (a macbook air running Mathematica 10), the following takes about 0.24 seconds:

AbsoluteTiming[Eigenvalues[testmat, 1, Method -> "Arnoldi"]]

Indeed it finds the right eigenvalue. I now feed it information about the eigenvector:

AbsoluteTiming[Eigenvalues[testmat, 1,Method -> {"Arnoldi", "StartingVector" -> testvec}]]

This once again takes 0.24 seconds.

  1. Why does this not give me a speedup?
  2. What can I do differently to get a speedup?

Note: I realize that I can get about a factor of 4 speedup by using the "Shift" option -- but it seems I should be able to get a couple orders of magnitude better by using a good starting vector.

$\endgroup$
  • $\begingroup$ In playing around with your problem, I was unable to select a badstartvec that would make it worse either. Various randomizations, or all zeros except for a random "1"...result seems to be independent of the "StartingVector". Got the same eigenvalue and same timing (percentage points) no matter what. $\endgroup$ – MikeY Apr 6 '17 at 16:01
5
+25
$\begingroup$

Since the test matrix you use has rank 1, many iterative schemes will converge very rapidly. Starting with a more general real symmetric, positive-definite matrix, we compare 3 ways of computing the largest eigenvalue (and corresponding eigenvector)

a = RandomReal[NormalDistribution[0, 1], {2000, 2000}];
testmat = Transpose[a].a;

A generic approach

AbsoluteTiming[{val, vec} = Eigensystem[testmat, 1];]
(* {0.893087, Null} *)

The uninitialised Arnoldi method is 5 x faster

AbsoluteTiming[{val1, vec1} = 
   Eigensystem[testmat, 1, Method -> "Arnoldi"];]
(* {0.162465, Null} *)

The initialised Arnoldi method is almost 4 x faster again

AbsoluteTiming[{val2, vec2} = 
   Eigensystem[testmat, 1, 
    Method -> {"Arnoldi", "StartingVector" -> First[vec]}];]
(* {0.046821, Null} *)

All give the same eigenvalue

Flatten[{val, val1, val2}]
(* {7935.59, 7935.59, 7935.59} *)

EDIT

I also note that your examples evaluate much faster if converted to full, rather than sparse arrays.

$\endgroup$
  • $\begingroup$ In an actual application that will necessitate the use of Arnoldi, converting to a dense array will likely be prohibitive in memory. $\endgroup$ – J. M. will be back soon Apr 10 '17 at 23:23

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.