Search by quantum walk and extended hitting time Andris Ambainis, Martins Kokainis University of Latvia
Exhaustive search Finite search space. Some elements might be marked. Find a marked element!
Search with structure Finite search space. Some elements might be marked. Find a marked element! After checking A, it may be easier to check B than C.
Example: search on grids N N grid. In one step, we can: – check if vertex marked; – move 1 step;
Search by random walk Random walk, following the locality constraints. Stop after finding a marked vertex.
Szegedy’2004 Random walk: T steps Quantum walk: O( T) steps
Szegedy’2004 (fine print) Random walk finds marked element: T steps Quantum walk detects if marked element exists: O( T) steps
Quantum walk detects if marked element exists: O( T) steps
Quantum walk detects if marked element exists: O( T) steps | start - starting state; No marked element - | start unchanged; Marked elements - | start diverges to an almost orthogonal state | . | not concentrated on marked element
Open question Random walk finds marked element: T steps Quantum walk finds marked element: O( T) steps ?
Krovi, Ozols, Magniez, Roland (previous talk) Quantum algorithm that finds marked element in O( HT + ) steps, HT + - extended hitting time. HT + = HT if there is 1 marked element; HT + can be larger than HT. How large can HT + be?
This talk Weak (upper) bound on HT+. Two big gaps between HT+ and HT.
DEFINITIONS
Markov chains /3 1/3 2 2/3
Classical hitting time
Matrix form Eigenvalues – real. 1 = 1, eigenvector – stationary distribution. 1 > 2 ... n. Spectral gap: probability of transition i j
UPPER BOUND ON HT +
Upper bound
1 - 2
Corollary
How strong is this result?
Unstructured search (Grover, 1996)
Grover’s algorithm Query Q: check if an element marked; Diffusion D: – | start | start ; – | -| , | | start . Repeat D, Q, D, Q,..., D, Q.
Diffusion Diffusion D: – | start | start ; – | -| , | | start . Markov chain: – |v 1 |v 1 ; – |v i i |v i , i 1- . Can implement diffusion with O(1/ ) steps of Markov chain.
Summary KMRO algorithm: – at least as good as Magniez-Nayak-Roland-Santha; – finds 1 marked element optimally. More general description when KMRO works well?
GAPS BETWEEN HT+ AND HT
Example 1 Stationary distribution: π x =1/3 for all x. M = {1, 2}. HT 10/9. HT+ 1/(4 ) 1- 0.1-
Example 1 If =0, two eigenvectors with i = 1. If 0, 2 1. Large contribution to HT +, causing HT + 1- 0.1-
Gap between HT and HT+, for a natural search space?
2D grid N N grid. Spectral gap: (1/N). Possible: HT + N HT.
2D grid: example 1 N N grid. HT = (1). HT + = (N). (1) fraction of vertices marked. Classical search easy – no need for quantum.
2D grid: example 2 Gap persists, unless the number of marked vertices small.
marked unmarked
2D grid: example 2 Outside: divide into k k squares, mark corners. Inside: divide into (2k) (2k) squares, mark corners. Regular pattern, with different densities outside and inside.
Classical hitting time Lower bound: hitting time with 1 marked vertex in each (2k) (2k) square. 1 of 4k 2 vertices marked.
Extended hitting time Calculation, using eigenvectors of the grid. HT + = (N), for any density of marked vertices.
KMRO algorithm Result: uniform superposition over marked vertices.
KMRO algorithm Pr=1/2 Pr=4/5
Marking more elements may increase HT + HT + = (1)HT + = (N)
What else can we try?
A, Bačkurs, et al., TQC’ D grid, 1 marked vertex. Standard quantum walk. After O( N log N) steps, state orthogonal to | start . Measurement: Pr[marked] = o(1); Pr[distance N from marked] const.
Idea 1 Does final state | final have large probability on vertices that are close to marked? If true - measure| final , obtain v, search the neighbourhood of v classically.
Idea 2 If HT = T, classical walk P hits a marked vertex in O(T) steps, with probability 1- . G’ – neighbourhood of the starting vertex where P stays during O(T) steps. Quantum walk on G’ instead of the full space?
Conclusions Upper bound for HT +, via spectral gap. KMRO algorithm at least as good as Magniez- Nayak-Roland-Santha. Two examples of gaps between HT + and HT. Optimal quantum algorithm should not be producing the uniform superposition of marked vertices!