What to do with a small quantum computer? Aram Harrow (MIT Physics)

Transcription

1 What to do with a small quantum computer? Aram Harrow (MIT Physics) IPAM Dec 6, 2016

2 simulating quantum mechanics is hard! DOE (INCITE) supercomputer time by category 15% 14% 14% 14% 12% 10% 10% 9% 2% Implications: 1. Quantum mechanics is hard to simulate classically. 2. Quantum computers can do this simulation exponentially more efficiently. 3. Maybe they can do other things more efficiently too?

3 Motivation Practical Social/economic value of computing End of Moore s Law Foundational Nature of information and computing

4 classical computers n bits = 2 n states gates x y x NAND y

5 quantum computers qubit = two-level system nuclear spin photon polarization ion e - states 1 qubit = 2 dimensions i 2C 2 superconducting charge or flux... n qubits = 2 n dimensions Schrödinger s eq: d i dt = ih(t) i i 2C 2n quantum gates I I U I I one- or two-qubit

6 randomized computation p0000 Bp0001 C B C 2n B.. C 2 A p1111 ZApplication: ZZ Z high-dimensional integrals dx1... dxn f (x1,..., xn ) FERMIAC

7 unstructured search x f(x) N 0 Given the ability to compute f, find x such that f(x)=1. classical: O(N) time needed Grover s algorithm: (1996) O( p N) on quantum computer similar speedups for maximizing, approximate counting, collisions, triangle finding, game-tree evaluation, backtracking algorithms

8 factoring = n digits?? Best classical algorithm: time O(exp(n 1/3 )) Shor s algorithm: (1994) time O(n 2 ) on a quantum computer (similar for discrete log, elliptic curves, )

9 linear systems of equations dimension N condition number κ Classical: O(Nsκ) x A B B C A = = b 0 1 B A up to s nonzero entries per row/column Quantum: O(log(N)sκ) A computable on the fly input bi and output xi Applications: ODEs PDEs regression machine learning

10 experimental QC 1-2 clean qubits 1000 noisy qubits IBM, Google, Microsoft, Intel, Alibaba, BBN, Rigetti, psitech, ionq, Quantum Circuits Inc, proposals for 50-qubit universal QCs in 2-10 years using superconductors and ion traps D-Wave (1qbit, QxBranch, Lockheed Martin)

11 combinatorial optimization Problem: minimize f(x 1,...,x n ) = (terms involving a few variables) Applications: minimum energy of classical systems best way for a protein to fold best model in a family explaining observed data math proof with fewest errors Difficulties: exponentially many local minima NP-complete even for good approximations

12 rest of the talk classical optimization algorithms adiabatic optimization variational algorithms and quantum supremacy

13 classical local search Local search Start at point x Choose a neighbor y Move to y if f(x) > f(y) Repeat n=3 problem: local minima

14 classical simulated annealing goal: sample from Gibbs distribution Pr[x] = e f(x) T annealing Z Metropolis algorithm x R x!y R y!x y R x!y R y!x = e f(x) f(y) T gradually lower T T= : unbiased random walk T=0: local search

15 Walks on the hypercube {0,1} n = n-dimensional hypercube adjacency matrix A A xy = 1 if (x,y) connected, 0 if not B C A n=3 Graph Laplacian: L = n I A continuous-time random walk: dp/dt = -Lp è uniform distribution Simulated annealing: dp dt = diag(e f 2T ) L diag(e f 2T )p

16 overview classical optimization algorithms adiabatic optimization variational algorithms and quantum supremacy

17 quantum mechanics to the rescue Diffusion equation: d p(~r, t) = dt Lp(~r, t) ~2 L=r Schrödinger equation: d (~r, t) = il (~r, t) dt

18 math-physics dictionary eigenvectors of H quantum states eigenvalues of H eigenvector with lowest eigenvalue energies ground state Schrödinger s eq: d i dt = ih(t) i energy = frequency

19 t adiabatic algorithm Schrödinger s eq: d i dt = ih(t) i H initial H(t) H final ground state (t=0) ground state (t=t final ) easy to prepare desired output Adiabatic theorem: Slowly changing H(t) à stay in ground state H initial H final

20 adiabatic optimization H initial = L H final = diag(f) ground state (t=0) ground state (t=t final ) easy to prepare 1 p 2 n H(t) = C. A 1 1 t T final L + Schrödinger s eq: d i dt t T final diag(f) desired output B A 0 = ih(t) i arg min x f(x)

21 0 n/4 n/2 n z tunneling and optimization Analogue for adiabatic optimization f( z ) z = # 1 s in string z

22 adiabatic tunneling beats simulated annealing Adiabatic optimization fast if barrier area n 1/2 Simulated Annealing exponential slowdown

23 quantum advantage in tunneling? classical algorithms can only walk over this landscape. Quantum computers can tunnel through the landscape making it faster to find the lowest point. The D-Wave processor considers all the possibilities simultaneously to determine the lowest energy from

24 ground states of stoquastic Hamiltonians = H(t) = 0 apple0 apple0. 1 t T final 1 apple0 apple0... apple0 apple0 C A... L + t T final diag(f) sign-problem free stoquastic ground state ψ has only nonnegative entries non-universal form of quantum computing Quantum Enhanced Optimization

25 quantum Monte Carlo (QMC) (actually a classical algorithm) stoquastic classical n n H = f + Γ L m Vertical bonds: ferromagnetic energy log(m/γ) i.e.disagree prob O(Γ/m) Horizontal bonds: = f / m

26 n QMC vs barriers Thm: For barrier sizes/shapes where the adiabatic algorithm tunnels through rapidly then quantum Monte Carlo does too. Elizabeth Crosson Proof idea: Only a few rows on the barrier at once. m

27 simulated annealing vs quantum adiabatic optimization vs quantum Monte Carlo

28 implications Evidence against tunneling as a route to quantum speedup. Routes to quantum advantage? Non-stoquastic adiabatic optimization Tasks where quantum Monte Carlo fails

29 overview classical optimization algorithms adiabatic optimization variational algorithms and quantum supremacy

30 variational quantum algorithms t 1 t 2 t 3 t 4... H = L diag(f) L diag(f) p(z) Theorem: Quantum Supremacy Outputs of quantum algorithm are hard to simulate classically.

31 things I don t know Why are there quantum speedups? Exponentially large dimension not enough. Interference needed too. What good is a 100-qubit quantum computer? Stoquastic adiabatic evolution: quantum supremacy or efficient classical simulation?

32 deleted scenes

33 standard part of QMC 1. Use local moves (Glauber or Metropolis) to generate samples from π(z 1,..., z L ). Run-time/accuracy tradeoff unknown in general. 2. Use sampling-to-counting equivalence to estimate Z or O =tr[o e -βh ]/Z. Problem reduces to bounding mixing time (equiv. gap) of a classical Markov chain.

34 n The Markov chain m jump z 2 z weight(z) = exp(-β(f(z 1 ) f(z L )) / m) (βγ/m)# vertical jumps rapidly mixing?

35 the path measure random walk z 1,...,z m on hypercube {0,1} n conditioned to return (z m+1 = z 1 ) alternatively can use open boundary conditions. typically βγn total jumps [see also JSIBMTN ] Suppose that f(z) depends only on Hamming weight z. look only at Hamming weight: {0,1} n -> {0,1,,n}. take n-> and {0,1,,n} à [0,1]. Brownian motion, or with closed B.C., Brownian bridge with local Z fields -> Brownian motion with drift Ornstein-Uhlenbeck bridge dx(t) = θ(μ-x(t)) dt + σ db(t) θ = drift, μ = mean, σ = diffusion imaginary time Hamming weight

36 local times of Brownian motion Local time: L x (t) = amount of time Brownian motion B(t) spends at point x. Lévy s theorem: {L 0 (t): t 0} and {S(t): t 0} have the same distribution, where S(t) = sup 0 s t B(s). In fact, (S-B, S) = d ( B, L 0 ) Additionally, S = d B.

37 local times of Brownian motion Local time: L x (t) = amount of time Brownian motion B([0,t]) spends at point x. Lévy s theorem: (S-B, S) = d ( B, L 0 ) Proof: consider discrete r walk: W(k) = X(1) X(k) with X(t) = ±1. Let M(k) = max(w(0),..., W(k)) W (t) M(t) Y(t) = M(t)-W(t) d W(t) t Y (t) =M(t) M = # of times M-W remains at 0 d L 0 W (t) t figure adapted from Brownian Motion by Mörters and Peres.

38 Hamming weight + spike f( z ) FGG 02 R 04 KC 15 BvD 16 JSIBMTN 16 width n a 1 exp(-n a+(b-1)/2 ) height n b height b 0.5 n 1/2-a-b 0 n n n 4 2 z 0 Ω(1) 0 width a 0.5

39 QMC and tunnelling spike (width n a, height nb ) imaginary time ST = normalized spike time d N(0, n a-1/2 ) proof using either Lévy s thm or quantum-classical correspondence Hamming weight Feynman-Kac thm: Pr[path spike] = exp(-βst n b ) Pr[path no spike] à if a+b<1/2 then typical paths don t notice the spike.

40 instantons on the cheap spike (width n a, height nb ) a<1/2 2a+b<1 cf JSIBMTN 16 1 imaginary time Hamming weight steps to traverse spike n 2a min ST = n 2a / βγn Feynman-Kac à prob reduced by exp(-n 2a+b-1 ) 2a+b 1 is the theshold to cross the spike once.

41 canonical paths Given Markov chain P(x,y) with stationary distribution π(x) and Q(x,y) = P(x,y) π(y) = Q(y,x). TFAE: P has a 1/poly(n) gap between the top two eigenvalues The conductance Φ is 1/poly(n). Φ = min S Q(S, S c ) / π(s) π(s c ) conductance For any x,y there exists a path γ xy from x -> y routing π(x) π(y) units of flow such that each edge e has load poly(n) Q(e). ( canonical paths/flows ) Heuristics analyze some plausible cut. Proofs analyze all cuts or construct paths. canonical paths

42 1-d canonical path x 1,1 x 2.1 x 3,1 x 4,1 x 1,2 x 2,2 x 3,2 x 4,2 x 1,3 x 2,3 x 3,3 x 4,3 x 1,4 x 2,4 x 3,4 x 4,4 x 1,5 x 2,5 x 3,5 x 4,5... y 1,1 y 2.1 x 3,1 x 4,1 y 1,2 y 2,2 x 3,2 x 4,2 y 1,3 y 2,3 x 3,3 x 4,3 y 1,4 y 2,4 x 3,4 x 4,4 y 1,5 x 2,5 x 3,5 x 4,5 energy penalty: 2 new jumps 1 term from H D (L bonds each with weight 1/L.) x 1,6 x 2,6 x 3,6 x 4,6 y 1,6 x 2,6 x 3,6 x 4,6 y 1,1 x 2.1 x 3,1 x 4,1 x 1,2 x 2,2 x 3,2 x 4,2 x 1,3 x 2,3 x 3,3 x 4,3 x 1,4 x 2,4 x 3,4 x 4,4 x 1,5 x 2,5 x 3,5 x 4,5 x 1,6 x 2,6 x 3,6 x 4,6... y 1,1 y 2.1 y 3,1 y 4,1 y 1,2 y 2,2 y 3,2 y 4,2 y 1,3 y 2,3 y 3,3 y 4,3 y 1,4 y 2,4 y 3,4 y 4,4 y 1,5 y 2,5 y 3,5 y 4,5 y 1,6 y 2,6 y 3,6 y 4,6