anthony leverrier (inria paris) 29 march 2018 orap 41st...
TRANSCRIPT
Quantum algorithms
Anthony Leverrier (Inria Paris)
29 March 2018
ORAP 41st Forum: Quantum Computing
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 1 / 13
Outline
I 1982 - 1993: the genesis of quantum computing, a scientific curiosity
I 1994 - 2009: Shor, Grover and generalizations
I 2010 - : solving linear equations, quantum machine learning
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 2 / 13
Genesis of quantum computingHow can we simulate the behaviour/evolution of quantum systems, since we need tokeep track of an exponential number of parameters?
Feynman 1981
“Can quantum systems be probabilistically simulated by aclassical computer? [. . . ]The answer is almost certainly, No!”=⇒ use quantum systems to simulate quantum systems!
Deutsch 1985
I quantum Turing machine
I existence of a universal machine
Bernstein, Vazirani 1993
I efficient quantum Turing machine (complexity class BQP)
I Bernstein-Vazirani problem: f : {0, 1}n → {0, 1} such that f (x) = a · xFind a. =⇒ ok with 1 quantum query vs n classically
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 3 / 13
Genesis of quantum computingHow can we simulate the behaviour/evolution of quantum systems, since we need tokeep track of an exponential number of parameters?
Feynman 1981
“Can quantum systems be probabilistically simulated by aclassical computer? [. . . ]The answer is almost certainly, No!”=⇒ use quantum systems to simulate quantum systems!
Deutsch 1985
I quantum Turing machine
I existence of a universal machine
Bernstein, Vazirani 1993
I efficient quantum Turing machine (complexity class BQP)
I Bernstein-Vazirani problem: f : {0, 1}n → {0, 1} such that f (x) = a · xFind a. =⇒ ok with 1 quantum query vs n classically
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 3 / 13
Genesis of quantum computingHow can we simulate the behaviour/evolution of quantum systems, since we need tokeep track of an exponential number of parameters?
Feynman 1981
“Can quantum systems be probabilistically simulated by aclassical computer? [. . . ]The answer is almost certainly, No!”=⇒ use quantum systems to simulate quantum systems!
Deutsch 1985
I quantum Turing machine
I existence of a universal machine
Bernstein, Vazirani 1993
I efficient quantum Turing machine (complexity class BQP)
I Bernstein-Vazirani problem: f : {0, 1}n → {0, 1} such that f (x) = a · xFind a. =⇒ ok with 1 quantum query vs n classically
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 3 / 13
The first algorithms
Simon, Shor 1994exponential speedups for
I period finding
I factoring
I discrete logarithm
=⇒ exploits Quantum Fourier Transform=⇒ consequences for public-key cryptography: a main argument to obtain
funding in quantum computation, seen as a real threat today by crypto agencies(NIST competition for post-quantum cryptography)
Grover 1996I search an n-item list with O(
√n) queries
I Lots of applications (find collisions, approximate counting,shortest path)
but only quadratic improvement
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 4 / 13
Hidden subgroup problem (≈ exponential speedup)
The problem for group G
I input: function f “periodic” on subgroup H ≤ G :
f (x) = f (y) ⇐⇒ y ∈ xH
I output: generator set for H
Applications
I Simon problem: G = (Z2)n, H = {0, s}I factoring: G = Z, H = rZ
I discrete logarithm: G = Z2, H = {(rx , x) : x ∈ Z}
I Pell’s equation: G = R
I graph isomorphism: G = Sn
Polynomial-time algorithms (in log |G |)ok for Abelian groups G (Shor 1995), for normal subgroups H (2000)
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 5 / 13
Polynomial speedup
Generalizations of Grover’s search: “blackbox tools”
I amplitude amplification: unstructured problems
I quantum walks (Szegedy 2004)
I span programs (Reichardt, Spalek 2008): AND-OR tree evaluations
I learning graphs (Belovs 2012): triangle finding, etc
Applications to symmetric cryptography
I find a key: quadratic speedup =⇒ double key size
I break hash functions: find collisions x , y such that h(x) = h(y)
Lower bounds
If one models complexity as the number of queries to the function/oracle, thenquantum speedup for many problems is at most polynomial.
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 6 / 13
This was the situation until not so long ago.
Main algorithms: Shor and Grover (+ generalizations)
(also quantum chemistry . . . )
then came HHL and the prospect of quantum machine learning
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 7 / 13
This was the situation until not so long ago.
Main algorithms: Shor and Grover (+ generalizations)
(also quantum chemistry . . . )
then came HHL and the prospect of quantum machine learning
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 7 / 13
The HHL algorithm, “exponential speedup”
the problem
Given an n × n matrix A and a vector b, solve the linear system:
Ax = b
I classically, complexity at least n2
I quantumly, O(log n)!!
How is that even possible??by cheating a little bit... Read the fine prints!
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 8 / 13
The HHL algorithm, “exponential speedup”
the problem
Given an n × n matrix A and a vector b, solve the linear system:
Ax = b
I classically, complexity at least n2
I quantumly, O(log n)!!
How is that even possible??by cheating a little bit... Read the fine prints!
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 8 / 13
The HHL algorithm, “exponential speedup”
the problem
Given an n × n matrix A and a vector b, solve the linear system:
Ax = b
I classically, complexity at least n2
I quantumly, O(log n)!!
How is that even possible??
by cheating a little bit... Read the fine prints!
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 8 / 13
The HHL algorithm, “exponential speedup”
the problem
Given an n × n matrix A and a vector b, solve the linear system:
Ax = b
I classically, complexity at least n2
I quantumly, O(log n)!!
How is that even possible??by cheating a little bit... Read the fine prints!
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 8 / 13
The HHL algorithm, “exponential speedup”
the problem
Given an n × n matrix A and a vector b, solve the linear system:
Ax = b
the algorithm (Harrow, Hassidim, Lloyd 2009)
Given access to a matrix A and a vector b such that
1. b can be efficiently loaded in a quantum memory
2. A has sparsity s per row and condition number κ
there exists a Q algo that
I outputs |x〉 =∑n
i=1 xi |i〉 where Ax = b
I runs in time polynomial in (s, κ) and logarithmic in the dimension
Requires A to be sparse and well-conditioned
Is this algorithm useful?? (does not return the complete answer)
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 9 / 13
The HHL algorithm, “exponential speedup”
the problem
Given an n × n matrix A and a vector b, solve the linear system:
Ax = b
the algorithm (Harrow, Hassidim, Lloyd 2009)
Given access to a matrix A and a vector b such that
1. b can be efficiently loaded in a quantum memory
2. A has sparsity s per row and condition number κ
there exists a Q algo that
I outputs |x〉 =∑n
i=1 xi |i〉 where Ax = b
I runs in time polynomial in (s, κ) and logarithmic in the dimension
Requires A to be sparse and well-conditioned
Is this algorithm useful?? (does not return the complete answer)
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 9 / 13
Quantum machine learningHHL initiated the field of quantum machine learning (subroutine for many algos).
Least Square Fitting [Wiebe, Braun, Lloyd 12]
I input: N labelled points (xi , yi )
I output: A fit function f (x , λ) =∑
j fj(x)λj that minimizes
err =N∑i=1
|f (xi , λ)− yi |2
Using HHL, the algorithm returns |λ〉.
Support Vector Machine [Lloyd, Mohseni, Rebentrost 13]I input: M labelled N-dimensional points
(xi, yi ), xi ∈ RN , yi ∈ {−1, 1}I output: A maximum margin hyperplane that separates the
classes
SVM can be recast as a system of linear equations=⇒ the quantum algorithm returns |w〉, the normal vector to the hyperplane
Application: classify data by estimating the inner product with w (l2-norm)
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 10 / 13
Quantum machine learningHHL initiated the field of quantum machine learning (subroutine for many algos).
Least Square Fitting [Wiebe, Braun, Lloyd 12]
I input: N labelled points (xi , yi )
I output: A fit function f (x , λ) =∑
j fj(x)λj that minimizes
err =N∑i=1
|f (xi , λ)− yi |2
Using HHL, the algorithm returns |λ〉.
Support Vector Machine [Lloyd, Mohseni, Rebentrost 13]I input: M labelled N-dimensional points
(xi, yi ), xi ∈ RN , yi ∈ {−1, 1}I output: A maximum margin hyperplane that separates the
classes
SVM can be recast as a system of linear equations=⇒ the quantum algorithm returns |w〉, the normal vector to the hyperplane
Application: classify data by estimating the inner product with w (l2-norm)
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 10 / 13
Quantum machine learning
Singular Value Estimation
I Sampling eigenvalues/vectors [Lloyd, Mohseni, Rebentrost 13]For a psd matrix with trace 1, quantum algorithm that efficiently samples aneigenvector with corresponding eigenvalue
I Singular Value Estimation [LMR 13, Prakash 15]Given a matrix and a singular vector, outputs an estimate of the singular value
Not clear whether these algorithms offer a true speedup
I to obtain a good complexity, the data should be “nice”, b should be efficientlyaccessible. Maybe there are efficient classical algorithms in such cases?
=⇒ HHL should be seen as a template for a generic algorithm: one needs to providea setup where
I b can be loaded efficiently in the quantum memory
I the data is sufficiently nice
I we’re not interested in x but in quantities efficiently computable from |x〉
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 11 / 13
Quantum machine learning
Singular Value Estimation
I Sampling eigenvalues/vectors [Lloyd, Mohseni, Rebentrost 13]For a psd matrix with trace 1, quantum algorithm that efficiently samples aneigenvector with corresponding eigenvalue
I Singular Value Estimation [LMR 13, Prakash 15]Given a matrix and a singular vector, outputs an estimate of the singular value
Not clear whether these algorithms offer a true speedup
I to obtain a good complexity, the data should be “nice”, b should be efficientlyaccessible. Maybe there are efficient classical algorithms in such cases?
=⇒ HHL should be seen as a template for a generic algorithm: one needs to providea setup where
I b can be loaded efficiently in the quantum memory
I the data is sufficiently nice
I we’re not interested in x but in quantities efficiently computable from |x〉
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 11 / 13
Quantum machine learning: speedup for a real-life problem
Quantum recommendation systems [Kerenidis, Prakash 16]
I input: a hidden preference matrix T with Ti,j ∈ {0, 1}, depending on whetherproduct j is “good” for user i
I output: a high value of row i (i.e. a recommendation for user i)
I complexity: polylog(mn) (= exponential speedup)
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 12 / 13
Conclusion
Summary
I there are quantum algorithms beyond Shor and Grover
I e.g., generalization of these techniques
I lower bounds in the query complexity model
“Hot topics”
I applications of HHL =⇒ quantum machine learningreal applications?
I cryptanalysis of quantum-resistant (“postquantum”) cryptography
I not mentioned but probably relevant in the next few years:algorithms for quantum chemistry
thanks!
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 13 / 13
Conclusion
Summary
I there are quantum algorithms beyond Shor and Grover
I e.g., generalization of these techniques
I lower bounds in the query complexity model
“Hot topics”
I applications of HHL =⇒ quantum machine learningreal applications?
I cryptanalysis of quantum-resistant (“postquantum”) cryptography
I not mentioned but probably relevant in the next few years:algorithms for quantum chemistry
thanks!
Anthony Leverrier (INRIA) ORAP Forum, 29 March 2018 13 / 13