14
$\begingroup$

Shor's algorithm is an algorithm which factors integers in polynomial time on a quantum computer. If one tries to run it on a classical computer, one runs into the problem that the state vector that is being operated on is of exponential size, so it cannot be run efficiently.

However, let us make the following observation with respect to MCMC (Markov Chain Monte Carlo) algorithms; see example 1 in Persi Diaconis' paper "The MCMC revolution" available at http://statweb.stanford.edu/~cgates/PERSI/papers/MCMCRev.pdf.

Notice that the state space in example 1 is of exponential size just like the state space of Shor's algorithm, yet the MCMC converges rapidly to the desired state. Also, the matrix involved in the MCMC is of exponential size just like the unitary matrices in Shor's algorithm. From this observation, I was thinking that we could do something similar with Shor's algorithm to get it to run in polynomial time on a classical computer, via Monte Carlo simulation.

Of course, one obvious problem with such an approach is that the unitary matrices in Shor's algorithm involve complex numbers, so they cannot be stochastic matrices. Also, the $l2$ norm, not the $l1$ norm, of the columns of the unitary matrices in Shor's algorithm equal one. (Furthermore, the unitary matrices in Shor's algorithm are all different, so it is not even close to a Markov Chain!)

But why should this stop us? Here is how I was thinking we could potentially fix this problem: First, convert the unitary matrices used in Shor's algorithm to orthogonal matrices so that each entry $a+bi$ gets converted to the matrix

$\left( \begin{array}{cc} a & b \\ -b & a \end{array} \right)$.

But these new matrices cannot be stochastic, since they involve negative numbers. To fix this problem, I thought about converting each entry $a$ in the orthogonal matrix to the matrix of only nonnegative entries

$\left( \begin{array}{cc} a & 0 \\ 0 & a \end{array} \right)$, if $a>0$ and

$\left( \begin{array}{cc} 0 & -a \\ -a & 0 \end{array} \right)$, if $a<0$.

(The idea here is that the matrix acts only on vectors of form $(x,0)^t$ when $x>0$ and $(0,-x)^t$ when $x<0$.)

So we have quadrupled the column and row dimensions of the unitary matrices in Shor's algorithm. Next, we can normalize the columns of the new matrices, so that they all add to one.

(Added: As one of the commenters, zeb, pointed out after my original post, normalizing the columns of the new matrices could potentially mess up the algorithm, but I was thinking that since in Shor's original paper, http://arxiv.org/abs/quant-ph/9508027, the quantum gates used to make up the Fourier transform are so incredibly simple, this might not be such a big problem.)

My question is: is there anything fundamentally stopping us from using these new matrices to efficiently factor integers via Monte Carlo simulation in much the same way as is described in Diaconis' paper (except of course, all of the matrices here are different)? Did I miss something?

$\endgroup$
  • $\begingroup$ Any boolean circuit can be turned into a Bayesian network, so if MCMC always works then P = NP. $\endgroup$ – Geoffrey Irving Apr 26 '15 at 6:42
  • $\begingroup$ It can be easier to typeset matrices using the command that exists to that end $\begin{pmatrix} 1 & 2 \\ 3 & 4 \end{pmatrix}$ $\begin{pmatrix} 1 & 2 \\ 3 & 4 \end{pmatrix}$ $\endgroup$ – user9072 Apr 26 '15 at 9:24
  • 1
    $\begingroup$ Wait - "normalize the columns of the new matrix"? Doesn't this step completely mess up the algorithm? $\endgroup$ – zeb Apr 26 '15 at 10:09
  • $\begingroup$ @Zeb, in general yes. But when looking at Shor's original paper arxiv.org/abs/quant-ph/9508027, the quantum gates that he uses for the Fourier transform are simple enough that perhaps normalizing the columns in the new matrices will not mess things up too badly. $\endgroup$ – Craig Feinstein Apr 26 '15 at 15:18
  • $\begingroup$ the problem you are running into with the negative weight in your Monte Carlo algorithm is the infamous "negative sign problem"; it is believed to be NP-hard --- arxiv.org/abs/cond-mat/0408370 $\endgroup$ – Carlo Beenakker Apr 26 '15 at 15:38
15
$\begingroup$

If this kind of simple trick involving just the Fourier transform, and not taking advantage of special properties of multiplication modulo $n$, worked, then it would provide a fast classical algorithm not just for factoring but for the more general abelian hidden subgroup problem - e.g. identifying the period of an arbitrary periodic sequence.

But there is no fast classical algorithm for the abelian hidden subgroup problem, since if you measure the value of a function $\mathbb Z \to \pm 1$ at $n$ integers, then you have not ruled out the possibility of the function being periodic with period some prime $p$ that does not divide the difference of any two of your numbers. Many such primes of size around $n^2$ necessarily exist, so you have no hope of figuring out the period in time less than the square root of the size of the period.

So MCMC can't work for this problem.

$\endgroup$
  • 1
    $\begingroup$ I don't follow your reasoning. Shor's algorithm doesn't just take the Fourier Transform. It computes $x^a (\bmod n)$ too. But the unitary matrices associated with these computations do not appear to be problematic, as they are just permutation matrices, right? $\endgroup$ – Craig Feinstein Apr 26 '15 at 18:04
  • 1
    $\begingroup$ @CraigFeinstein Yes, the classical steps of the computation are permutation matrices, which are not problematic. What I mean by "just the Fourier transform" is that the only thing we do to the sequence $x^a$ mod $n$ is compute it and then apply the Fourier transform. We don't investigate its properties using deep number theory technology. $\endgroup$ – Will Sawin Apr 26 '15 at 18:40
  • $\begingroup$ I think you are correct, @willsawin. $\endgroup$ – Craig Feinstein Apr 27 '15 at 0:49
  • 1
    $\begingroup$ @DominiqueUnruh It is not proven for real-world problems but it is known for oracle problemss. The abelian hidden subgroup problem is an oracle problem, which unless I am very confused has no fast classical algorithm for the reason I described. $\endgroup$ – Will Sawin Dec 20 '16 at 12:10
  • 1
    $\begingroup$ @DominiqueUnruh Well, one should read what I wrote in context. The reason the oracle problem is relevant here is that, if Craig Feinstein's strategy works for factoring, it necessarily works for the oracle problem as well. $\endgroup$ – Will Sawin Dec 21 '16 at 15:17

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

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