Adiabatic quantum computation – a tutorial for computer scientists Itay Hen Dept. of Physics, UCSC Advanced Machine Learning class UCSC June 6th 2012 Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Outline introduction I: what is a quantum computer? introduction II: motivation for adiabatic quantum computing adiabatic quantum computing simulating adiabatic quantum computers future of adiabatic computing [AQC and machine learning(?)] Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 What is a Quantum Computer? Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 input state computer computation 011 010011 What is a quantum computer? output state both classical and quantum computers may be viewed as machines that perform computations on given inputs and produce outputs. Both inputs and outputs are strings of bits (0’s and 1’s). classical computers are based on manipulations of bits. at any given time the system is in a unique “classical” configuration (i.e., in a state that is a sequence of 0’s and 1’s). Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 input state input state computer computation output state computation 10010 11101 01010 a classical algorithm (circuit) looks like this: 01011 011 010011 What is a quantum computer? output state at any given time the system is in a unique “classical” configuration (i.e., in a state that is a sequence of 0’s and 1’s). Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 input state computer computation 011 010011 What is a quantum computer? output state quantum computers on the other hand manipulate objects called quantum bits or qubits for short. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 input state computer computation 011 010011 What is a quantum computer? output state quantum computers on the other hand manipulate objects called quantum bits or qubits for short. what are qubits? Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 What are qubits? a qubit is a generalization on the concept of bit. motivation / idea behind it: small particles, say electrons, obey different laws than ‘big’ objects (such as, say, billiard balls). small particles obey the laws of Quantum Physics. big objects are classical entities – they obey the laws of Classical Physics. (however, big objects are collections of small particles!) most notably: a quantum particle can be in a superposition of several classical configurations. classical world I’m here Itay Hen or I’m there quantum world I’m here and Advanced Machine Learning class– UCSC I’m there June 6th, 2012 What are qubits? while a bit can be either in the a |0 state or a |1 state (this is Dirac’s |𝑘𝑒𝑡 notation). a qubit can be in a superposition of these two states, e.g., |𝜑 = 1 2 |0 + |1 imagine |0 and a |1 as being two orthogonal basis vectors. while a classical bit is either of the two, a qubit can be an arbitrary (yet normalized) linear combination of the two. in order to access the information stored in a qubit one must perform a measurement that “collapses” the qubit. extremely nonintuitive. important: only 1 bit of information is stored in a qubit. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 What are qubits? take for example a system of 3 bits. Classically, we have 23 = 8 possible “classical configurations”: |000 , |001 , |010 , |011 , |100 , |101 , |110 , |111 , the corresponding quantum state of a 3-qubit system would be: 7 |𝜑 = 𝑐𝑖 |𝑖 𝑖=0 this is a (normalized) 8-component vector over the complex numbers (8 complex coefficients). |𝑖 enumerates the 8 classical configurations listed above. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Encoding optimization problems suppose that we are given a function 𝑓: {0,1}3 → ℝ and are asked to find its minimum and the corresponding minimizing configuration. this is the same as having the diagonal matrix: 𝐹= and asking which is the smallest eigenvalue (minimum of f ) and corresponding eigenvector (minimizing configuration). here, the basis vectors 𝑒𝑖 correspond to |𝑖 and the elements on the diagonal are the evaluation of f on them. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Encoding optimization problems in matrix-form we can generalize classical optimization problems to quantum optimization problems by adding off-diagonal elements to the matrix. we still look for the smallest eigen-pair, i.e., 𝑡 min 𝜑|𝐹|𝜑 = min 𝜑 𝐹𝜑 |𝜑 𝜑 but problem now is much harder. (this is Dirac’s 𝑏𝑟𝑎| notation). in quantum mechanics, the minimal (eigen)vector is called the “ground state” of the system and the corresponding minimal (eigen)value is called the “ground state energy”. also, the matrix F is called the Hamiltonian (usually denoted by H). quantum mechanics is just linear algebra in disguise. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 computer input state output state input state Itay Hen computation Advanced Machine Learning class– UCSC 10010 11010 + 11001 + 11101 01010 + 01001 quantum computers manipulate quantum bits or qubits for short. a quantum algorithm (circuit) may look like this: 01011 computation 011 010011 What is a quantum computer? output state June 6th, 2012 input state computer computation space of possible quantum states is huge. range of operations on qubits is huge. 011 010011 What is a quantum computer? output state capabilities are greater (name dropping: superposition, entanglement, tunneling, interference…). can solve problems faster. only problem with quantum computers: they do not exist! theory is very advanced but many technological unresolved challenges. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Motivation for adiabatic quantum computing Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Motivation clearly, a quantum computer is a generalization of a classical computer in what ways quantum computers are more efficient than classical computers? what problems could be solved more efficiently on a quantum computer? these are two of the most basic and important questions in the field of quantum information/computation. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Motivation in what ways quantum computers are more efficient than classical computers? what problems could be solved more efficiently on a quantum computer? best-known examples are: Shor’s algorithm for integer factorization. Solves the problem in polynomial time (exponential speedup). current quantum computers can factor all integers up to 21. Grover's algorithm is a quantum algorithm for searching an unsorted database with 𝑁 entries in 𝑂 𝑁1/2 time (quadratic speedup). importance is huge (cryptanalysis, etc.). Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Motivation what other problems can quantum computers solve? people are considering “hard” satisfiability (SAT) and other optimization problems which are at least NP-complete. these are hard to solve classically; time needed is exponential in the input (exponential complexity). could a quantum computer solve these problems in an efficient manner? perhaps even in polynomial time? Adiabatic Quantum Computing is a general approach to solve a broad range of hard optimization problems using a quantum computer [Farhi et al.,2001] Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Adiabatic quantum computation Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 The nature of physical systems adiabatic quantum computation is different that circuit-based computation (there are still other models of computation). adiabatic quantum computation is analog in nature (as opposed to 0/1 “digital” circuits). it is based on the fact that the state of a system will tend to reach a minimum configuration. this is the principle of least action. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Analog classical optimization we are given a landscape f(x) for which we need to find the minimum configuration, i.e., the stable state of the ball. if energy landscape is “simple”: the ball will eventually end up at the bottom of the hill. f→ x→ Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Analog classical optimization we are given a landscape f(x) for which we need to find the minimum configuration, i.e., the stable state of the ball. if energy landscape is “jagged”: ball will eventually end up in a minimum but is likely to end up in a local minimum. this is no good. f→ x→ Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Analog classical optimization we are given a landscape f(x) for which we need to find the minimum configuration, i.e., the stable state of the ball. if energy landscape is “jagged”: ball will eventually end up in a minimum but is likely to end up in a local minimum. this is no good. f→ probability of jumping over the barrier is exponentially small in the height of the barrier. x→ ball is unlikely to end up in the true minimum. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Analog quantum optimization we are given a landscape f(x) for which we need to find the minimum configuration, i.e., the stable state of the ball. if energy landscape is “jagged”: ball will eventually end up in a minimum but is likely to end up in a local minimum. this is no good. f→ quantum mechanically, the ball can “tunnel through” thin but high barriers! x→ sometimes more likely to end up in the true minimum. tunneling is a key feature of quantum mechanics. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 The adiabatic theorem of QM the adiabatic theorem of QM tells us that a physical system remains in its instantaneous eigenstate if a given perturbation is acting on it slowly enough and if there is a gap between the eigenvalue and the rest of the Hamiltonian's spectrum. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 The adiabatic theorem of QM the adiabatic theorem of QM tells us that a physical system remains in its instantaneous eigenstate if a given perturbation is acting on it slowly enough and if there is a gap between the eigenvalue and the rest of the Hamiltonian's spectrum. ?? Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 The adiabatic theorem of QM the adiabatic theorem of QM tells us that a physical system remains in its instantaneous eigenstate if a given perturbation is acting on it slowly enough and if there is a gap between the eigenvalue and the rest of the Hamiltonian's spectrum. rephrasing: if the ball is at the minimum (ground state) of some function (Hamiltonian), and the function is changes slowly enough, ball will stay close to the instantaneous global minimum throughout the evolution. here, initial function is easy to find the minimum of. final function is much harder. take advantage of adiabatic evolution to solve the problem! Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 The adiabatic theorem of QM the adiabatic theorem of QM tells us that a physical system remains in its instantaneous eigenstate if a given perturbation is acting on it slowly enough and if there is a gap between the eigenvalue and the rest of the Hamiltonian's spectrum. real QM example: change the strength of a harmonic potential of a system in the ground state: harmonic potential 𝜓(𝑥, 𝑡) Itay Hen Advanced Machine Learning class– UCSC 2 June 6th, 2012 The adiabatic theorem of QM the adiabatic theorem of QM tells us that a physical system remains in its instantaneous eigenstate if a given perturbation is acting on it slowly enough and if there is a gap between the eigenvalue and the rest of the Hamiltonian's spectrum. real QM example: change the strength of a harmonic potential of a system in the ground state: an abrupt change (a diabatic process): harmonic potential 𝜓(𝑥, 𝑡) Itay Hen Advanced Machine Learning class– UCSC 2 June 6th, 2012 The adiabatic theorem of QM the adiabatic theorem of QM tells us that a physical system remains in its instantaneous eigenstate if a given perturbation is acting on it slowly enough and if there is a gap between the eigenvalue and the rest of the Hamiltonian's spectrum. real QM example: change the strength of a harmonic potential of a system in the ground state: a gradual slow change (an adiabatic process): wave function can “keep up” with the change. harmonic potential 𝜓(𝑥, 𝑡) Itay Hen Advanced Machine Learning class– UCSC 2 June 6th, 2012 The quantum adiabatic algorithm (QAA) the mechanism proposed by Farhi et al., the QAA: 1. take a difficult (classical) optimization problem. 2. encode its solution in the ground state of a quantum “problem” Hamiltonian 𝐻𝑝 . 3. prepare the system in the ground state of another easily solvable “driver” Hamiltonian” 𝐻𝑑 . 𝐻𝑝 , 𝐻𝑑 ≠ 0. 4. vary the Hamiltonian slowly and smoothly from 𝐻𝑑 to 𝐻𝑝 until ground state of 𝐻𝑝 is reached. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 The quantum adiabatic algorithm (QAA) the interpolating Hamiltonian is this: 𝐻(𝑡) = 𝑠(𝑡)𝐻𝑝 + 1 − 𝑠(𝑡) 𝐻𝑑 𝐻𝑝 is the problem Hamiltonian whose ground state encodes the solution of the optimization problem 𝐻𝑑 is an easily solvable driver Hamiltonian, which does not commute with 𝐻𝑝 the parameter 𝑠 obeys 0 ≤ 𝑠 𝑡 ≤ 1, with 𝑠 0 = 0 and 𝑠 𝒯 = 1. also: 𝐻 0 = 𝐻𝑑 and 𝐻 𝒯 = 𝐻𝑝 . here, 𝑡 stands for time and 𝒯 is the running time, or complexity, of the algorithm. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 The quantum adiabatic algorithm (QAA) the interpolating Hamiltonian is this: 𝐻(𝑡) = 𝑠(𝑡)𝐻𝑝 + 1 − 𝑠(𝑡) 𝐻𝑑 the adiabatic theorem ensures that if the change in 𝑠 𝑡 is made slowly enough, the system will stay close to the ground state of the instantaneous Hamiltonian throughout the evolution. one finally obtains a state close to the ground state of 𝐻𝑝 . measuring the state will give the solution of the original problem with high probability. how fast can the process be? Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Quantum phase transition bottleneck is likely to be something called a Quantum Phase Transition happens when the energy differenece between the ground state and the first excited state (i.e., the gap) is small. there, the probability to “get off track” is maximal. a schematic picture of the gap to the first excited state as a function of the adiabatic parameter 𝑠. gap ∆𝐸1 Itay Hen 0 𝐻 = 𝐻𝑑 QPT 1 𝑠 𝐻 = 𝐻𝑝 Advanced Machine Learning class– UCSC June 6th, 2012 Quantum phase transition Landau-Zener theory tells us that to stay in the ground state the running time needed is: 𝒯∝1 2 Δ𝐸𝑚𝑖𝑛 exponentially closing gap (as a function of problem size N) exponentially long running time exponential complexity. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Quantum phase transition Landau-Zener theory tells us that to stay in the ground state the running time needed is: 𝒯∝1 2 Δ𝐸𝑚𝑖𝑛 exponentially closing gap (as a function of problem size N) exponentially long running time exponential complexity. a two-level avoided crossing system remains in its instantaneous ground state Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 The quantum adiabatic algorithm most interesting unknown about QAA to date: could the QAA solve in polynomial time “hard” (NP-complete) problems? or for which hard problems is 𝒯 polynomial in 𝑁? Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Simulating Adiabatic Quantum Computers Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Exact diagonalization quantum computers are currently unavailable. how does one study quantum computers? quantum systems are huge matrices. sizes of matrices for n qubits: 𝑁 = 2𝑛 . one method: diagonalization of the matrices. however, takes a lot of time and resources to do so. can’t go beyond ~25 qubits. another method: “Quantum Monte Carlo”. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Quantum Monte Carlo methods quantum Monte Carlo (QMC) is a generic name for classical algorithms designed to study quantum systems (in equilibrium) on a classical computer by simulating them. in quantum Monte Carlo, one only samples the exponential number of configurations, where configurations with less energy are given more weight and are sampled more often. this is usually a stochastic Markov process (a Markovian chain). there are statistical errors. not all quantum physical systems can be simulated (sign problem). quantum systems have an additional “extra” dimension (called “imaginary time” and is periodic). for example a 3D classical system is similar to 2D quantum systems (with notable exceptions). Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Method main goal: determine the complexity of the QAA for the various optimization problems study the dependence of the typical minimum gap Δ𝐸min = min Δ𝐸 𝑠∈ 0,1 on the size (number of bits) of the problem. this is because: 𝒯∝1 2 Δ𝐸𝑚𝑖𝑛 polynomial dependence polynomial complexity! Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Future of adiabatic Quantum Computing Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Future of adiabatic QC so far, there is no clear-cut example for a problem that is solved efficiently using adiabatic quantum computation (still looking though). the first quantum adiabatic computer (quantum annealer) has been built. ~128 qubits. built by D-Wave (Vancouver). one piece has been sold to USC / Lockheed-Martin (~1M$). a strong candidate for future quantum computers. there are however a lot of technological challenges. people are looking for other possible uses for it. Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Adiabatic QC and Machine Learning one of many possible avenues of research is Machine Learning. people are starting to look into it. supervised and unsupervised machine learning. could work if problem can be cast in the form of an optimization problem. Hartmut Neven’s blog: http://googleresearch.blogspot.com/2009/12/machinelearning-with-quantum.html#!/2009/12/machine-learning-withquantum.html Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012 Adiabatic quantum computation – a tutorial for computer scientists Itay Hen Dept. of Physics, UCSC Advanced Machine Learning class UCSC June 6th 2012 Itay Hen Advanced Machine Learning class– UCSC June 6th, 2012