Title: Quantum basin hopping with gradient methods
Dr David Bulger
Department of Statistics - Macquarie University
This talk will describe a global optimisation method involving two
well-known quantum algorithms: Grover's algorithm, and the quantum
Fourier transform. As a point of departure, consider multistart with
some gradient-based local optimisation (steepest descent, conjugate
gradient, et cetera). That is, we repeatedly choose a random domain
point x, and descend to the bottom of the "basin" containing x. When a
stopping criterion is reached, we report the best point seen, which is
the local minimum in the best basin we've found. Thus there are two
components: a searching component (searching for the best basin) and a
local optimisation component.
Grover's quantum searching algorithm can find a target in effort
proportional to 1/sqrt(p), where p is the target's size
(the corresponding effort on a Turing computer is proportional to 1/p).
Also, the quantum complexity of gradient estimation does not increase
with domain dimension (the corresponding Turing complexity increases
linearly with dimension).
The optimisation approach discussed in this talk combines
* the natural benefit of local optimisation achieved by classical
multistart,
* the quadratic acceleration of Grover's algorithm (loosely, a speed
increase by a factor of the square root of the number of basins), and
* the acceleration due to faster gradient estimation (a speed increase
by a factor of the domain dimension).
It is uncertain how far we can get in 50 minutes. We'll start with a
quick & simple overview of quantum computation. Next we'll walk through
the execution of Grover's algorithm in the simplest nontrivial case.
Next we'll look at the dynamic oracle construction trick which converts
Grover's algorithm to a minimisation method. Next we'll look at
gradient estimation. Putting it all together may have to be left as an
exercise!